America | Tech | Opinion | Culture | Charts

SaaSpocalypse Interrupted

The actual story of the SaaSpocalypse continues to be more nuanced than the narrative.

It’s definitely true that software companies generally are experiencing some headwinds, but whether that’s strictly speaking “AI-driven” and/or something else, is a bit of an open question.

Plus, there’s plenty of reason to think that AI could be an accelerant to incumbent software businesses–indeed, before the big saas selloff, that was the narrative: incumbents incorporating AI stood to gain massively.

The most likely outcome for “AI v. SaaS” is, of course, “it depends.” Some companies are intrinsically more vulnerable, for sure, but a lot will simply depend on execution. Can companies effectively incorporate AI into their stacks, products, and pricing models (or not)?

For now, some data suggests that some software companies are doing just-fine (while others, less so). Based on a panel of “stable software spenders,” the picture is mixed: there are big-winners amongst incumbent saascos, especially amongst the biggest buyers of software, but AI companies are definitely making gains:

• Hubspot stands out for seeing large recent increases amongst the biggest spenders, and the largest yoy median increase for the panel as a whole.

• Figma, Box, Cloudflare, Datadog, Semrush, and Monday, also grew 25%+ amongst the biggest spenders (but only Figma experienced any gains at the median)

That’s good news for Hubspot and Figma, two companies that many believe are directly in AI crosshairs. For its part, Hubspot has been telling investors for a while now that it was investing heavily in AI, so perhaps this is part of the payoff.

In general, though, the same data shows some tough(er) sledding ahead for non-AI software.

Subscribe now

Amongst those companies increasing their tech spending 10%+, AI continues to see an increasing share of overall spend:

Over 60% of the panel now allocate 5%+ of their spending to AI, in contrast to the ~12% of the panel, from just one year ago.

What that means is that AI is capturing more and more of the incremental tech spending–i.e. when companies spend more on tech, they are spending it on AI.

Again though, that doesn’t mean that all categories of software companies are equally exposed. From a recent CIO survey, which asked “which area(s) of IT budget do you expect the greatest reduction in budget growth?” it’s clear that some categories are clearly more exposed than others:

71% of CIO respondents identified as Systems Integrators as the area expected to bear the cost of AI spending reallocation.

In contrast, front-office (i.e. crms), vertical saas, and infra were somewhere in the middle, while back-office, cloud, and cyber(!) appear to be relatively safe.

Make of that what you will, but the point is that the battle for enterprise tech spending is very much on, but we’re still in the early-innings of how all of this will eventually shakeout.

Consumer AI: Room to Run

Over on the consumer side of AI things, there too, we’re still very much in the early innings.

While AI adoption is quite high, paying AI adoption (amongst consumers) is just getting going. According to BofA’s internal data, only 3% of households are paying subscribers to an AI service.

3% of households is a relatively tiny number, although it’s been growing fairly quickly, over the past year:

The number of households “making AI payments” is up nearly 40%, relative to February 2024.

That growth is impressive, but it’s nothing like the growth we’ve seen on the enterprise side. At only 3% penetration, there is an awful lot of room-to-run in consumer-AI.

The other thing to keep in mind is that among households who are spending on AI, the bigger spenders are increasingly a larger share of the pie:

While the majority of paying households are still paying $20 or less monthly, over 40% are now paying more than that, and every “big spending” group increased its relative share.

Consumers who are spending on AI, are spending more and more.

In terms of who specifically is driving those gains? Well, the data is pretty clear on that too:

Gen-Z and Younger-Millennial households have increased their median AI spending by ~54% over the past year, while older households have been relatively flat.

Younger people are accelerating their AI spending much more quickly, over the past year. They also represent the households with the largest share of AI spenders–but even amongst Gen Z, only 5% of households are paying for AI services.

Again, however quickly growth has been, we’re still very early in this game.

Ads Come for Everything

Streaming prices are going up.

Netflix just bumped its ad-free Standard plan to $19.99. Disney+ premium is $18.99. A three-service, no-ads bundle now costs more than $55/month – roughly what a basic cable package used to cost.

We have been liberated from cable only to recreate . . . cable.

But the more interesting trend is what’s happening alongside the price hikes: the ad tier is taking over.

Netflix subscribers on ad-supported tiers went from 32% to 40% last year. Disney+ from 37% to 45%. And Prime Video now sits at 86% ad-supported – though that’s partly by design since Amazon defaults all subscribers to the ad tier and charges extra to opt out.

The streaming platforms are converging on the same playbook: raise the price of the ad-free tier until only the most price-insensitive subscribers stick around. Everyone else drops to the ad tier, where the platform collects a cheaper subscription plus ad revenue on top. The streamers are extracting maximum revenue from the people who hate ads, and maximum attention from the people who tolerate them.

If the trend holds, ad-supported will be the majority of accounts on every major platform not named Apple TV within a year or two.

Trucks Are Busy

There are some early signs that manufacturing activity may be picking up. Most of the evidence so far is survey-based, but if you want a more tangible signal, well, we’ve got that too.

Consider the freight market. Freight may be a leading indicator of industrial and manufacturing activity, which makes sense, given that more manufacturing means more stuff to ship.

And right now, some data suggests that demand for freight is surging.

The national outbound tender rejection index has more than doubled from roughly 5-8% last year to over 15% today.

When a shipper has freight to move, they “tender” it to a carrier under contract. The carrier can accept or reject. When rejection rates are high, it means that carriers are turning down shipments, usually because demand is outstripping supply.

Flatbed is even more striking, with nearly half of all contracts rejected:

A 49% rejection rate for flatbed trucking is higher than it was during the early-2022 pandemania surge.

Flatbed is the segment most tied to industrial and construction activity (steel, lumber, heavy equipment), so when carriers are rejecting nearly half of all tenders, it tends to mean the physical economy is busier than the headlines suggest.

Intermodal rates are telling a similar story – after trading flat for months, rates spiked roughly 10% in March:

Of course, rejections can spike for two reasons: more freight or fewer trucks. And after three years of a freight downturn (that drove trucking companies out of business), there are genuinely fewer trucks out there. In other words, surely some of what we’re seeing is a supply-side story, as well.

That said, 49% flatbed rejections are hard to explain on supply alone. Domestic demand for big-stuff shipping is ripping. Whether it keeps on ripping remains to be seen, but the freight data says the physical economy is tighter than it’s been in years.

Earnings Go Up, Prices Go Down

Here’s something that you don’t usually see.

Normally, when earnings go up, share prices go up too. And yet, forward earnings for the S&P 500 just hit an all-time high, but the index price is rolling over:

When earnings increase, but share prices decline, it necessarily means one thing: multiple compression.

Forward P/E for the index has retreated to ~21.5x, which is still quite high, but not quite as high as its top centile range of ~23x.

Why are earnings multiples compressing? There’s never just one reason, but some possibilities include simple mean reversion, plus a healthy dose of geopolitical risk, volatile commodity prices, and general uncertainty.

Another likely factor is earnings quality. Profits are still going up, but free cash flow is going down, down, down:

For the hyperscalers, revenue and operating profits continue to rise, but free cash flow has plummeted, as they plow more and more money into the AI buildout.

Gotta spend money to make money, as the saying goes. Doesn’t mean the market is going to like it, however.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

This week, someone hijacked one of the most popular packages on the internet and used it to install a backdoor on every machine that ran npm install.

The package was Axios, which is the HTTP library that powers network requests in millions of applications. It gets downloaded over 100 million times a week. If you’ve built anything in JavaScript in the last decade, you’ve almost certainly used it, or something you depend on has used it for you.

The attacker didn’t change a single line of Axios’s source code. They compromised a maintainer’s account, added one new dependency to the package manifest, and published an update. That dependency, a brand-new package called plain-crypto-js that had been registered hours earlier, ran a script on install that detected your operating system, downloaded a remote access trojan tailored to your machine, executed it, and then deleted itself. By the time you looked at your node_modules folder, it was gone. The malware had already phoned home.

This is the new shape of software supply chain attacks: fast, quiet, and invisible to traditional security tools. And it’s accelerating.

Everyone knows supply chain attacks are a problem. That’s not the point.

If you work in security, you’ve heard this story before. SolarWinds. Log4Shell. The XZ Utils backdoor. Supply chain attacks make headlines, the industry wrings its hands, and then everyone goes back to business.

So what’s different now?

Two things changed in the last twelve months, and they changed at the same time. First, frontier coding models got good enough to write and ship code autonomously, and that made both sides faster. AI agents are pulling in dependencies at machine speed with minimal human review. Second, attackers figured out how to weaponize the dependency graph at scale - not just targeting one package, but cascading across entire ecosystems in days.

The result is a mismatch that should terrify anyone building software: the attack surface is expanding faster than any human can monitor, and the entities making dependency decisions are increasingly not human.

Dependencies are the new perimeter

It’s always been true that the code you write is a small fraction of the code you ship.The average application contains over 1,100 open source components. A bare-bones Next.js project installs 282 packages before you write a single line. The median JavaScript project on GitHub has 755 transitive dependencies - packages that your packages depend on, chosen by nobody on your team.¹

When an attacker compromises one node in that graph, they don’t need to target you specifically. Every npm install during the exposure window is a potential victim. The Axios attacker didn’t need to know who you are or what you built. They compromised the foundation of the ecosystem and waited for it to collapse.

Subscribe now

Modern-day security tooling looks for the wrong things. Most software composition analysis tools work by checking your dependencies against a database of known vulnerabilities - CVEs. But a deliberately planted backdoor doesn’t have a CVE. There’s no database entry for a brand-new malicious package. Running npm audit on the compromised Axios version would have returned a clean bill of health, because the malware had already self-destructed.

One stolen token, five ecosystems, eight days

The Axios hack was headline-grabbing, but it was arguably not even the most consequential supply chain attack this quarter. That distinction belongs to a campaign called TeamPCP, which demonstrated what cascading supply chain failure actually looks like.²

It started with Trivy, a widely-used open source vulnerability scanner. An attacker exploited a misconfiguration in its GitHub Actions workflow to steal an access token. From there, they force-pushed malicious code to nearly all of Trivy’s version tags, meaning anyone pulling the scanner was now running the attacker’s code. The payload harvested credentials from the CI/CD environment: SSH keys, cloud tokens, and npm tokens.

Those stolen npm tokens fueled the next stage. A self-propagating worm called CanisterWorm spread across 66+ npm packages, using blockchain-based command infrastructure that couldn’t be conventionally taken down. It harvested more tokens from infected developer environments and automatically published malicious updates to every package those tokens had access to.³

Within eight days, the same campaign had cascaded from GitHub Actions to Docker Hub, npm, PyPI, and the VS Code extension marketplace. With just one token across five ecosystems, thousands of organizations were potentially impacted.

Supply chain attacks used to be surgical, patient, and targeted;the most common exploits were things like a nation-state actor spending two years earning maintainer trust on XZ Utils. Now they’re automated, self-propagating, and ecosystem-spanning. The timeline compressed from years to days.

AI coding, AI attack surface

The same AI tools that are making developers 2–4x more productive are also making the supply chain dramatically more vulnerable. And this cuts both ways: AI isn’t just writing vulnerable code; attackers are using the same tools to generate malicious packages, automate propagation, and find weaknesses at machine speed.

A study of over 117,000 dependency changes across thousands of GitHub repositories found that AI agents select known-vulnerable dependency versions 50% more often than humans. Worse, the vulnerable versions they pick are harder to fix, requiring major-version upgrades far more frequently.

Hallucinated packages are the sleeper threat. LLMs regularly invent package names that don’t exist. One study found that nearly 20% of AI-recommended packages were fabrications, and 43% of those hallucinated names appeared consistently across queries. Attackers have figured this out. The technique is called “slopsquatting“: register the fake package name the AI keeps recommending, fill it with malicious code, and wait. One researcher demonstrated this by uploading a dummy package with a commonly hallucinated name and watching it accumulate 30,000 downloads - largely from AI-driven workflows - in weeks.

The autonomous coding agents now entering production can install dependencies, execute builds, and open pull requests without a human ever touching the keyboard. They optimize for “does this work?” not “is this safe?” And they do it at speeds that compress the window for security review to essentially zero.

There’s also the risk that coding agents are themselves becoming formidable instruments of attack, even when they aren’t explicitly asked. Truffle Security recently found that Opus 4.6, along with other models, would discover and exploit SQL injection vulnerabilities when asked to fetch basic information from websites, in cases where the legitimate path to that information was blocked. So imagine what these coding agents are capable of when they’re wielded explicitly to mount sophisticated supply chain attacks.

We are building a world where machines write the code, machines choose the dependencies, and machines ship the updates. The AI agents are building the software. If we don’t secure the supply chain they rely on, the AI agents are cooked.

Your Scanner Wouldn’t Have Caught It

The industry average time to detect a supply chain breach is 267 days. SolarWinds went undetected for 14 months. XZ Utils took two years to surface.

Socket, an a16z portfolio company, detected the malicious dependency in the Axios attack within 6 minutes of its publication. That’s roughly 63,000 times faster than the industry average. They flagged it 16 minutes before the first compromised Axios version was even published, because they caught the suspicious dependency package itself.

How? Socket doesn’t check packages against a vulnerability database. It analyzes what the code actually does: Does it access the network? Spawn shell processes? Obfuscate its payload? Execute postinstall scripts? Read environment variables? This behavioral approach catches novel malicious packages that have no CVE and no prior history - exactly the kind of attack that traditional tools miss entirely.

The Axios attack was live for about three hours before npm pulled the compromised versions. In that window, tens of thousands of installations likely occurred. On 135 endpoints monitored by one security vendor alone, the malware executed and phoned home to the attacker’s server within 89 seconds of install. The gap between what’s possible (6-minute detection) and what’s typical (267-day detection) is where the damage happens.

Save the AI agents. Save the Robots

The software supply chain has become the most critical and least-defended attack surface in modern software development. AI is accelerating both sides of the equation: attackers using it to generate malicious packages at scale, defenders racing to build behavioral detection systems that can keep up. And the shift to autonomous AI agents writing and deploying code without human review has compressed the timeline from months to minutes.

The teams that are winning this fight share a common trait: they stopped waiting to be told they’d been compromised. They moved security controls to the point closest to the threat - the moment a dependency enters the build. They stopped relying solely on CVE databases and started analyzing actual package behavior. And they started treating their dependency graph not as a list of libraries, but as a living attack surface that needs continuous monitoring.

Software is being built at the speed of machines now. If defenders don’t match that speed, the machines shipping the code will be shipping the malware too.

Save the AI agents. Secure the supply chain.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Drew Baglino is founder and CEO of Heron Power, an industrial power electronics company focused on affordably scaling electrification. He is formerly a longtime engineer turned executive as SVP of Powertrain and Energy at Tesla.

The US electricity grid is one of the greatest infrastructure achievements in history. Built over a few remarkable pioneering decades, it now spans an entire continent and has served incredible growth throughout our cities and society. But this infrastructure has been quietly aging and much of the grid is now approaching end of life at the exact moment we desperately need it to support tremendous growth.

Electricity demand is accelerating faster than at any point in a generation — driven by datacenters, electric vehicles, heat pumps, re-industrialization, and AI. Every indicator points to that growth continuing: more of our energy economy runs on electricity each year, and more of that electricity comes from renewables. Getting to a fully electrified economy means generation needs to grow more than four times from today’s levels — and that estimate doesn’t yet account for the next wave of AI infrastructure now coming online.

I founded Heron Power in 2024 because I believe the grid itself is the next frontier for power electronics. The engineering capability exists. The talent exists. What’s needed now is the same concentration of focus, urgency, and scale that transformed electric transportation — directed at the infrastructure that powers everything else.

We are asking a nineteenth-century grid to power a twenty-first-century economy.

Electricity demand growth is running headlong into physical limits. In the US, seventy percent of transmission lines and large power transformers are over 25 years old, and more than half of all distribution transformers are nearing the end of their useful lives. In 2025, the American Society of Civil Engineers gave U.S. energy infrastructure a D+.

We have to renew existing infrastructure while planning for the future, and we have to do it affordably. The minute power bills jump, public support weakens, regulators get cautious and necessary projects stall. Every rate case is a balancing act between investing in new infrastructure and keeping power affordable. If utilities can’t serve future demand without higher rates, approvals slow down. And if bills rise anyway, the backlash ripples as delay and disruption to entire growth sectors.

That is why the next wave of grid investment must be different. We cannot simply replace aging equipment with more of the same. We need a new generation of grid-scale solutions that unlock capacity, increase utilization, improve reliability, and reduce loss — without pushing costs onto customers. A sustainable economy can only grow as fast as affordable electricity can propel it.

The superpowers of power electronics

The hero of this story is power electronics. Advances in computing, communications, heat pumps, renewable energy, efficient lighting, battery storage. Electric cars, AI datacenters, next-generation manufacturing. None of these scale without efficient, reliable, high-performance power conversion. Five decades of continuous advancement in power electronics has quietly enabled nearly every major technology of our era.

Subscribe now

I saw this firsthand over nearly two decades at Tesla, where I led the development and scaling of electric vehicle drivetrains, battery storage, and fast-charging infrastructure — none of which would have been possible without power electronics. By combining the best available device technology with rigorous engineering and manufacturing at scale, we transformed industries and unlocked cost curves that once seemed unreachable. But I also kept running into the same wall: progress at the grid’s edge was consistently outpacing the infrastructure on the other side of the wire. The build out of energy installations and factories stalled — not for lack of capital, technology, or permits, but because the grid equipment they needed to connect to was undersized, supply-constrained, or prohibitively expensive.

What’s limiting the grid

Since 2019, average residential electricity rates have climbed roughly 25-30%. But the cause isn’t what most people assume. Generation costs have actually declined — natural gas expanded, wind and solar fell dramatically, battery storage reached grid scale. The problem is delivery. Transmission and distribution now accounts for nearly half of what customers pay in many regions. Generation is getting cheaper, but delivery is more expensive. The net result is more expensive power.

Two structural problems explain why.

First, incentives are misaligned. Regulated utilities earn returns on capital deployed — the more they spend building, the more they earn. Decades of flat demand left equipment suppliers with little pressure to innovate. When the only tools available are large, slow, hardware-heavy upgrades, investment scales accordingly. Much of this spending is necessary — replacing aging assets, hardening against wildfires, expanding to serve new loads. But the incentive structure means there’s no natural brake on inefficient deployment.

Second, the system is engineered conservatively by necessity. Keeping the lights on is a literal matter of life and death, requiring simultaneous management of supply-demand balance, frequency, voltage, line congestion, fault-isolation and power quality in real time. Operators value reliability over cost minimization. Batteries and VPPs are making headway — aggregated home batteries and utility-controlled VPPs can defer pole and wire upgrades — but capacity is only part of the equation. Planners still need to solve for everything else, and today’s equipment gives them few tools to do so.The basic hardware hasn’t fundamentally changed in decades. Large mechanical switchyards direct power flow, breakers and reclosers protect circuits slowly with wide safety margins. The equipment is durable, but blunt. It does not sense or control dynamically. It cannot actively optimize. Transformers — on telephone poles, in green boxes in your neighborhood — convert voltage at a fixed ratio. No sensing. No telemetry. No control. Without visibility at each node, planners assume worst-case scenarios and overbuild accordingly.

An aging supply chain compounds the problem. Transformer demand has more than doubled since 2019, costs are up 80%, and the U.S. faces a 30% supply deficit. A single grossly under-capacitized domestic producer makes the steel required to build them. Eighty percent of supply is imported — for the most critical component of our infrastructure, at the exact moment we need to scale it.

The deeper challenge is distributed electrification: millions of EV chargers, heat pumps, rooftop solar, and commercial fleets connecting to circuits designed for a different world. These loads are dispersed, variable, and increasingly bidirectional. These are tractable problems for operators with dispatchable, continuous control of power flow. But with today’s passive, mechanical, analog equipment, planners default to what they know — overbuilding.

The problem is clear as day. If we want Americans to have affordable access to the abundant power we can now freely generate and propels prosperity, growth, and sustainability, things must change. We need better tools: tools to extract more capacity, more reliability, and more flexibility from the infrastructure already in the ground.

Applying Moore’s Law to power

Fortunately better technology exists today, born of decades of innovation in power semiconductors. Before they arrived, electrical engineers could only use mechanical devices to control current — think the light switch in your home, or the electro-mechanical relay that locks your car door. These switches are slow to change state, degrade as arcs pit their contacts over time, and top out at 5,000–10,000 lifetime cycles. At home wiring voltages they’re manageable, but at higher voltages they grow to the size of refrigerators — painfully expensive, and slow to actuate. In all cases, electrical switches that are fundamentally mechanical in nature have limited lifetime.

In the late 1970s and early 80s, pioneering scientists unlocked an entirely different class of device and entirely new means of controlling electricity. Thyristors, power diodes, and IGBTs could switch on and off in microseconds, block high voltages, and cycle trillions of times without fail. Power electronics enable electricity to control itself: in the right circuit topologies, engineers apply them to convert AC to DC, DC to AC, control power flow, or step between voltage levels with precision. Much like transistors in compute, power semiconductors have followed their own Moore’s law — smaller, faster, cheaper, and dramatically more capable on a relentless curve of improvement.

Since power semiconductors were first industrialized in the early 1980s, device switching speeds have improved by four orders of magnitude and current density by more than three. Fifty years ago, silicon MOSFETs switched in hundreds of microseconds and blocked 60–100 volts. Today, silicon carbide MOSFETs switch in nanoseconds and block thousands of volts in a millimeter-scale package. The power semiconductor business now generates roughly $60 billion in annual sales, enabling terawatts of new conversion capability every year.

And every step along that curve has unlocked new industries. Variable frequency drives transformed industrial efficiency in the 80s and 90s by precisely matching motor speed to load. In the 2000s, continued advances supported the build out of solar, wind and industrial automation. In the 2010s, another leap in performance helped EV powertrains and utility-scale batteries move into the mainstream. Today, a 140-watt USB-C charger uses GaN semiconductors switching a million times per second, packs 20X more power density than a comparable 1990s power supply, uses adaptive controls to share power across ports, and runs cool enough to require no fan at all. Each generation: lower cost, higher efficiency, higher power density.

The major inflection for higher-voltage applications came with silicon carbide (SiC), a semiconductor perfect for power. SiC devices operate at higher voltages and switching frequencies with less loss, run cooler, and enable smaller, more efficient converters than traditional silicon. It’s also an American success story — the technology emerged from North Carolina State and was commercialized by Cree, now Wolfspeed, in the 90s in Durham. Global competition is accelerating, and we should consider the national security priority of power chips alongside next-generation GPUs.

I lived this transition firsthand at Tesla. The 2006 Roadster and 2012 Model S used silicon IGBTs in their traction inverters. With the 2017 Model 3, we bet on SiC MOSFETs — and it paid off in efficiency, range, thermal performance, and package size. That same decision shaped the success of Megapack. Today most major EV and storage companies have followed, with more than two terawatts of SiC-enabled devices deployed in EVs alone every year.

Five decades in, nearly all power semiconductor progress has accrued to low-voltage applications: telecoms, datacenters, vehicles, chargers, solar, batteries. The distribution grid, operating at medium voltage between 7.2kV and 34.5kV, remains dominated by passive, mechanical hardware.

Medium voltage is the next frontier.

Solid-state transformers: one solution for many problems

Today’s best silicon carbide devices block up to 3.3 kilovolts and switch at hundreds of kilohertz. That combination makes something genuinely new possible: modular, medium-voltage, high-frequency power electronics stages that replace passive 60-hertz oil-filled transformers with software-controlled silicon. The latest SiC devices are more than a better switch. They are the foundation for solid-state transformers, or SSTs. Since the 2000s, SSTs have been at the forefront of research into medium-voltage applications and grid integration. Leading academic contributors such as Alex Q. Huang (NC State, FREEDM Systems Center), Marco Liserre (Kiel University), and Frede Blaabjerg (Aalborg University), along with institutions like UT Austin, RWTH Aachen, and Tsinghua University, have shaped the field.

Where a conventional transformer is a passive iron core on a pad, an SST is a software-defined power-conversion platform. It consolidates voltage conversion, fault isolation, power factor correction, frequency regulation, harmonics control, and phase balancing into a single package, managed in real time by software, far smaller and more power-dense than anything it replaces. Reliability comes not from mass and simplicity, but from modularity where no single failure is fatal, real-time monitoring that catches problems early, and protective response measured in microseconds. SiC devices already survive decades of thermal and electrical stress in EVs and grid storage. Bringing that durability to medium voltage is an engineering challenge, not a physics barrier.

The economics are compelling across every customer type. For solar and battery developers, NREL research found inverter failures accounted for 36% of lost energy in utility-scale systems over a 27-month study period, with inverters failing 300 to 500 times more frequently than PV modules. An SST with built-in redundancy and real-time monitoring directly attacks that problem. For utilities, an SST replaces not just a transformer but the switchgear, tap changers, capacitor banks, and phase balancers that utilities can rarely afford to deploy at scale. Every passive node replaced becomes a programmable one.

The cost trajectory only moves one direction. Power semiconductors get cheaper and more capable with every generation and scale economy. Steel and copper follow commodity cycles and are never structurally cheaper. And because SSTs are software-defined, they improve after installation. A conventional transformer that’s wrong for the job gets ripped out and replaced. An SST gets a firmware update.

Combined with distributed resources, the compounding effect on grid planning is significant. Batteries provide capacity. SSTs offer a new paradigm of visibility and control at the circuit level to maximize how much power existing infrastructure can serve. Move more kilowatt-hours across the same poles and wires with less capital and lower maintenance costs, and rates stabilize or fall. That can be deployed now, on circuits already scheduled for replacement.

Building a grid for the future

The grid will be rebuilt. Aging equipment will be replaced and demand will continue to grow. The only real question is whether we replace yesterday’s hardware with more of the same, or with systems designed for a more dynamic, electrified economy.

Grid hardware deployed today will remain in service for decades. It must support higher load density, distributed generation, bidirectional power flow, and tighter reliability requirements. It must do more than convert voltage. It must sense, respond, and adapt.

We have seen this transition before. Telecommunications moved from fixed-function electro-mechanical systems to programmable, software-defined networks. Costs fell, capacity expanded, and entirely new mobile and cloud industries emerged on top. Power is at a similar inflection point. There is real work ahead: solid-state transformers must integrate with existing protection systems, satisfy cybersecurity requirements, and prove long-term reliability in the field. But these are engineering challenges, not scientific ones. The physics is proven. The semiconductor supply chain exists. The capital will be deployed regardless.

The only choice is what we build. A grid rebuilt from passive equipment locks in today’s constraints for another generation. A grid rebuilt from active, programmable, multi-function SSTs gets smarter, more efficient, and more capable every year. It carries more power across the same infrastructure, stabilizes rates, and creates the foundation for everything the next economy will demand.

America has scaled transformative technologies before, from aerospace to semiconductors to the internet, when engineering and economics align behind a clear mission. The grid is that mission now. The tools exist. The need is urgent. And the opportunity, for the first time in a century, is to deploy dynamic, controllable hardware solutions that maximize our existing infrastructure and unlock future growth.

The future needs more electricity. Let’s build a grid worthy of delivering it.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

In the tech industry, all the attention goes to the frontier. And why not? Companies like OpenAI, Anthropic, and Google are pushing the boundaries of what’s possible, and billions of dollars chase each and every new breakthrough. But the largest market within technology is far more prosaic than any of that; it’s IT and security services.

Every single organization of any size leverages technology, and this infrastructure needs constant maintenance and support. Not every organization has the ability or the desire to handle that internally. So there’s a massive global market for managed service providers, or MSPs: the companies that provide IT, security, and compliance services to the vast majority of the organizations you interact with every day, from school districts to SMBs to the dentist’s office down the street. Add it all up, and we’re talking about a massive industry. IT spend is projected to hit $6 trillion in 2026, and there are over 40,000 MSPs in the United States alone.

And yet, despite being the largest market in technology, it is the last to meaningfully innovate.

There’s a reason for this. This enormous market is built on a legacy operating model that hasn’t meaningfully changed in two decades. Most MSPs are slow-moving lifestyle businesses; the rest are being rolled up by private equity. In either case, very little structural investment goes into modernizing how IT services are actually delivered. Humans are still the connective tissue across fragmented systems and service layers, even as automation and intelligence have reshaped every other corner of the tech industry.

This has real consequences for the millions of organizations that depend on these services. Only 23 percent of SMBs report being satisfied with their IT providers, largely because of slow response times and a lack of innovation.

IT spend also scales way past where it needs to. Complexity accumulates: you hire one MSP for general IT, another for security, a third for compliance, maybe a fourth for procurement, and before long you’re managing a constellation of providers, each speaking their own language, none fully coordinating with the others.

This is the problem that Treeline exists to solve. We are building what we call the Modern IT Operating System: a fundamental reimagining of how IT services work, one that uses software and AI to bridge the gap between technology’s frontier and its biggest market.

To understand why this matters, and why nobody has done it before, you have to understand the false choice that most organizations face when it comes to IT infrastructure.

The false choice

You can build it yourself. Hire an IT manager, then a security specialist, then a compliance officer. Each role costs six figures plus benefits. Each hire takes months. The team that worked when you had 100 employees is overwhelmed when you have 300. You’re constantly recruiting potential hires, onboarding new employees, and managing turnover. And none of this is your company’s core function. For a $50 million company, the historical rule of thumb is to spend 3 to 5 percent of revenue on IT, security, and compliance: that’s about $1.5 to $2.5 million a year, much of it headcount. And that’s the optimistic lower bound.

Subscribe now

Or you can outsource to a traditional MSP. This seems simpler: you pay for expertise and avoid the hiring headaches. But traditional MSPs grew out of local computer repair shops. They are services businesses with stagnant operating models, built on rigid workflows and reactive support models. The way most of them work today is the way they worked in 2005. And because fixing problems permanently means fewer billable hours, there’s no structural incentive to actually make anything better.

The fragmentation makes it worse. Most of those 40,000-plus U.S. MSPs have fewer than 10 employees. And so quality varies wildly. You might get a great technician who knows your particular systems inside and out, but when you need help with cloud architecture or a modern compliance framework, that same MSP is out of its depth. You’re back to adding vendors and stitching things together yourself.

What does this actually look like day-to-day? It’s not pretty. Slow, reactive fixes when critical systems fail. Error-prone offboarding that exposes the company to insider risk. Revenue stalled by a failed SOC 2 audit because nobody was tracking compliance in real time. Weak credentials. Unmanaged access. Employees waiting hours for basic support tickets. Overspending on SaaS licenses that nobody uses but nobody cancels. And this all adds up. Every dollar and minute that goes to managing infrastructure could have gone to product or customers.

There’s an irony worth noting here: the technology industry, which has transformed every other sector through software and automation, has not managed to transform its own plumbing. MSPs operate in largely the same way they did 20 years ago. While the rest of the tech industry moved to cloud-native architectures and (later) AI-powered workflows, the delivery of IT services stayed manual and labor-intensive.

What’s changed

For a long time, this was just how things were. Automation tools were brittle and inflexible. AI couldn’t handle the judgment calls that IT work actually requires: the ambiguous triage decisions, the integration headaches, and the context-dependent security responses. The expertise needed to secure a network and manage complex integrations was genuinely scarce and couldn’t be codified into software.

That has changed quickly in the last few years. End-to-end employee lifecycle management across systems, intelligent ticket triage with automated resolution, continuous security enforcement, real-time compliance tracking, proactive issue prevention through root-cause analysis – three years ago, you needed a person for every one of these. Today, they can be automated reliably.

But the mission-critical nature of IT means that pure automation isn’t sufficient, either. When systems go down, when a security incident happens, when somebody has to make a judgment call about architecture or compliance strategy, you need a human who knows what they’re doing. The right model is software handling everything that can be automated, while human expertise gets concentrated on the work that actually benefits from human judgment: strategy, architecture, novel problems. Five years ago the technology wasn’t good enough to pull this off. Now it is.

And that is what we’re building at Treeline.

The Modern IT Operating System

We recognized that the connective tissue across vendors, tools, and processes no longer needed to be human. So we productized the IT department’s core activities. Instead of relying on people to stitch together disparate and sometimes archaic systems, we built a centralized software layer that standardizes workflows, integrates tools into a single operational system, and automates the coordination that used to eat up most of an IT team’s time. IT, security, and compliance stop being separate silos managed by separate vendors and start functioning as one connected system.

While “tool consolidation” has been promised before, it usually just means replacing five dashboards with one dashboard. We aren’t repeating that mistake. We are building a new operating model: tickets get triaged automatically, and many issues get resolved before a human ever sees them. The system gets more intelligent with time, learning from patterns and fixing root causes. Security controls are enforced continuously. Compliance documentation is generated in real time. And human expertise gets applied deliberately, instead of being spread thin as default manual glue.

The hidden cost of traditional IT isn’t in the salary or in the vendor invoices. It’s in coordination. As an organization grows, more tools and more handoffs create more failure points, which creates pressure to add headcount or stack on another vendor just to keep pace. Simply put, IT costs scale faster than the organization they support. Treeline exists to break that pattern. Companies scale through systems instead of through people, and the link between organizational growth and ballooning IT spend gets severed.

This is urgent because the demands on IT infrastructure keep increasing. Security threats are escalating across the board. Compliance requirements multiply as regulations catch up to the age of AI-assisted hacking. Meanwhile, the 3 to 5 percent of revenue that goes to IT, security, and compliance becomes genuinely painful for any scaling business.

The current setup forces organizations to choose between expensive and inadequate. Build it yourself and watch costs spiral; or, outsource and accept inconsistent service. Businesses shouldn’t have to make that choice. We believe that IT infrastructure should be like electricity: reliable, scalable, something you rarely have to think about. When it actually works that way, companies can redirect the resources they used to burn on managing infrastructure toward all of the things that differentiate their business.

Treeline exists to make that real.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

IT, as defined by pop culture, is a dark room and someone hacking their way through a crisis. In reality, modern IT means managing devices, access, security, compliance, and support – continuously, invisibly, and at scale. It’s one of the most operationally demanding functions in any business, as any founder or operations lead will tell you. Because of the constant demands of the role, more than 80% of U.S. businesses outsource this mission-critical function to Managed Service Providers (MSPs) - specialized services firms that function as an external IT and security department.

The MSP market is the backbone of the US tech economy. MSPs collectively generate over $200B in annual revenue across roughly 40,000 independent businesses. As technology becomes more central to how every business runs, the need for someone to manage it has only deepened.

The vast majority of MSPs are small, lifestyle companies that deliver service the same way they always have – reacting to tickets and providing labor augmentation. But something important has changed. Advances in software and AI are making it possible to reimagine this model that historically required greater headcount for greater impact.

Enter Treeline – a team of engineers and operators building the AI-enabled IT services provider. To reinvent this legacy model and better service customers at scale, Treeline built a software layer that sits at the center of IT operations, standardizing workflows, automating routine work, and turning every resolved issue into an insight that makes the system smarter. The result is IT that gets more efficient over time – noticing patterns, addressing root causes, and bringing humans in where expertise matters.

Treeline has integrated leading MSPs into its offering in order to offer the market both world-class software and deep industry expertise. What’s different about the Treeline model, however, is that they aren’t just an “AI rollup.” Treeline’s platform completely reimagines how partner MSPs operate. Automated workflows dramatically reduce ticket response times, technicians offer white-glove support, and their broad product offering allows their customers to be forward thinking in ways traditionally reserved for the F100. All of this represents something the IT industry hasn’t seen before: leading human expertise and modern software, finally built into a single operating model.

Working with Treeline has afforded me the great pleasure of working with Peter Doyle and Hussain Kader. Peter and I met a decade ago and when he introduced me to Hussain, his close friend from undergrad at Stanford, it was clear they were the dream pairing to attack this enormous problem.

Investing in Treeline is an incredible full circle moment, and I couldn’t be more thrilled to be their partner. Onward!

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Walk into any building and you’re surrounded by systems you never think about. The pipes routing water through every floor. The ducts pushing conditioned air into every room. The wires connecting every outlet, circuit, and fire alarm. These systems didn’t appear by magic. An architect designed the shell. A structural engineer made it stand. A team of MEP consultants (mechanical, electrical, plumbing) designed the systems inside it. A general contractor coordinated the build. Specialty trades installed everything on site.

Behind every building is a chain of firms you’ve never heard of, doing work you’ve never thought about, in a $13 trillion industry that technology has largely bypassed.

They’re all, still, running on software from 1997.

The AEC Ecosystem

The architecture, engineering, and construction industry — AEC — is one of the largest and least digitized industries in the world. Global construction spend is $13 trillion annually and growing, driven by data center buildout, housing shortages, electrification, and infrastructure investment. The design and engineering services that are mission critical to construction projects account for hundreds of billions of this spend.

The industry operates by the coordination of many fragmented stakeholders. A developer or owner funds the project and sets the brief. An architect designs the building shell. MEP consultants design the mechanical, electrical, and plumbing systems that make the building function. A general contractor manages the overall build, bidding out installation work to specialty subcontractors — the electricians, plumbers, and HVAC technicians who physically put everything in place.

Inside each firm, the tools are a mess. Most of the software these firms run on was built in the late 1990s — desktop applications, installed locally, not connected or hardly connected to the cloud. They require enormous amounts of manual labor just to operate. An engineer doesn’t spend most of their day doing engineering. According to research by Autodesk and FMI, construction professionals spend 35% of their time, or over 14 hours a week, on non-productive activities like looking for project information, resolving conflicts, and dealing with mistakes and rework. Much of that is just keeping disconnected tools in sync with the actual state of the project.

Then there’s the problem between firms. A hospital project might involve dozens of companies — architect, structural engineer, MEP consultants, GC, a dozen specialty subs — each running their own version of their own building design file and tools, none of which are the same, almost none of which are cloud-based, all of which are sharing information through PDFs, emails, and weekly uploads to centralized drives. Changes cascade unpredictably. A structural beam moves on Tuesday. The MEP consultant doesn’t find out until Friday. By then the ductwork is routed through it. This cycle repeats constantly throughout a project. Someone has to fix it. In construction, fixing things late is orders of magnitude more expensive than getting them right early.

The consequences show up in the numbers. 85% of construction projects exceed their budgets and three quarters finish late. When things go wrong, they go expensively wrong: the average construction dispute in North America is worth $60.1 million and takes nearly 12.5 months to resolve. In the US alone, rework, miscommunication, and time spent hunting for project data costs the industry $177 billion every year — and more than 70% of that rework traces back to design errors, not site conditions or bad weather. The drawings were wrong before anyone broke ground. Of all the software in this chain, one tool above all others defines how the industry works — and has for the better part of three decades.

The Monopoly Nobody Loves

In 1997, a team of engineers set out to build a better design tool for architects. Their idea was simple: model the building in 3D from the start, with every element connected to every other. Change a wall and the floor plan updates. Move a window and the ceiling elevation adjusts to scale. They called the company Revit — a contraction of “revise instantly.”

Autodesk acquired Revit in 2002 for $133 million. At the time it was considered an expensive bet. It turned out to be one of the greatest acquisitions in enterprise software history. Revit is now the standard BIM (building information modeling) authoring platform at virtually every major architecture, engineering, and construction firm in the world. It does roughly $3 billion in annual recurring revenue and has over 95% market share. It is taught in every architecture and engineering school globally and it is the system of record for the industry.

It is to AEC what Excel is to finance: the connective tissue between every firm, every workflow, and every deliverable in the industry.

The “taught in schools” detail matters more than it sounds. Almost every engineer and architect who enters the workforce already knows Revit. Every firm defaults to it because their people know it, their clients expect it, and their project libraries — custom components, material specifications, firm-specific templates built up over years — live inside it. Those libraries, or Revit “families”, are stored in Autodesk’s proprietary file formats, which don’t export cleanly to anything else. Decades of project history, component libraries, and firm-specific standards are locked inside the Revit ecosystem.

Autodesk bolted MEP functionality onto Revit in 2007 and has done very little with it since. The interface looks almost identical to what it did twenty years ago. Collaboration happens almost entirely outside of the Revit system — everyone saves their model locally, syncs it to a master file, and spends time reconciling the conflicts that accumulated. It’s the construction equivalent of finance professionals emailing Excel files back and forth.

Subscribe now

The platform also doesn’t connect the 3D model to the physics-based calculations engineers need to run alongside it. Those happen in separate tools — notably Excel and other third-party software — all manually kept in sync by a variety of services providers. Instead of solving these problems and charging for new features, Autodesk relies on their monopoly status and drives revenue growth through price increases.

This may sound like a setup for a simple disruption story. It isn’t. Revit’s dominance has persisted not because Autodesk is protecting it aggressively, but because the problem it solves is genuinely hard. AEC design requires reasoning simultaneously about 3D geometry, physical constraints, building code, room occupancy, and equipment specs — while coordinating with teams making their own changes in parallel. Every serious challenger has stalled at the same wall: how do you teach software to actually *understand* a building, not just display it? How can you enable real-time collaboration? And if you can solve that, how do you get teams to trust your software when the stakes are this high?

Why Now

For a long time, the answer was: you can’t.

Two big things have changed this calculus:

The first is obviously the advent of LLMs and vision models enabling new tools to attack problem spaces previously left unaddressed by Revit. A BIM model contains enormous amounts of metadata — room types, occupancy classifications, equipment specs, spatial relationships — but historically that data was entered inconsistently and in formats that varied by firm and project. Previous systems couldn’t make sense of it. They couldn’t look at a room and understand that it’s a server room, which means a certain cooling load, which maps to specific code requirements, which constrains where you can route the ductwork. LLMs can do that. They parse the unstructured metadata, classify it semantically, and hand structured inputs to downstream engineering algorithms.

The second is the urgency for better tooling due to enormous infrastructure buildouts. As the infrastructure for AI gets built out, many have started to note the bottleneck is in data center construction and MEP infrastructure. The talent pipeline isn’t keeping pace, and there is an existential constraint on how these firms grow and demand is met in a timely fashion.

Three Ways to Attack a $13 Trillion Market

Given a dominant incumbent that hasn’t innovated in decades, a fragmented supply chain running on manual labor, and three technical unlocks that have only recently aligned — it’s worth being precise about how AI actually attacks this market. We’re seeing three distinct approaches in the early days of this market.

1. Attack Revit directly. Build a better BIM platform from scratch that’s cloud-native, AI-enabled, and collaborative by default, and win at the authoring layer. This is the most audacious approach and, historically, the graveyard. Every serious challenger for two decades has stalled trying to reach feature parity with a platform that firms have spent years customizing and their people have spent careers learning. GTM presents a few key challenges: you’re asking firms to retrain their workforce, rebuild their component libraries, and trust a new system of record with their most complex, highest-stakes work.
   
   Motif, founded by Autodesk’s former co-CEO, is the most credible current attempt. We believe now may be the time for this to finally happen as AI makes feature parity achievable on a timeline that wasn’t possible before — and that AI demand plus the industry’s frustration with Autodesk’s stagnation has created genuine appetite for a rip-and-replace.
   

2. Build new software surface area around Revit. Don’t fight it on its own turf. Own the workflows that Revit handles poorly or not at all (of which there are many) and use that foothold to expand. Most of the actual work in AEC happens around Revit, not inside it. Revit is the system of record, but engineers spend most of their time in Excel, Word, and disconnected third-party tools. The document review process that happens before anything goes to bid — checking design documents for errors, coordination failures, and omissions across disciplines — happens entirely outside Revit. Today it’s done by humans, takes three to six weeks, costs $50,000 to $100,000 per project, and still only catches about 30% of the issues that eventually become field change orders.
   
   LightTable is attacking exactly this workflow: AI that reads construction documents comprehensively, flags issues across disciplines, and prioritizes them by cost impact. The wedge doesn’t require anyone to change how they use Revit. It slots into the gap Revit leaves open. And every project that runs through the platform produces labeled data — here’s what the drawings showed, here’s what the issues were, here’s which ones became change orders — that makes the system more accurate over time. Nobody replaced SAP by building a better ERP. They built around it until SAP became the plumbing nobody touched.

3. Go after the services budget directly — the work that happens on top of Revit and has never been addressed by software at all. We believe MEP design may be the clearest opportunity here. Most of what MEP consultancies do isn’t done in Revit; it’s done by teams of engineers, often offshore, manually translating building specifications into drawings and documentation that has to be checked and revised at every step. This work has never been software-addressable before, but with AI we believe there is an opportunity to go after some of these workflows.
   
   This would unlock demand that was previously impossible to serve. Firms of many different kinds are sitting on backlogs of work they can’t take on because they don’t have the people. AI doesn’t just help them do existing projects faster; it lets them say yes to projects they would have turned away. Endra is building in this space. They have built an AI-native MEP design platform that ingests building models and automates fire, electrical, and mechanical system design — work that previously took months of manual effort, done in minutes.

Deep diving on MEP – where we’re seeing the conditions align first

MEP design is a $150 billion global services market that’s a large part of the broader AEC space. While Revit serves this area, most of the work is actually completed by services firms on top of or outside of Revit. The largest of these firms have tens of thousands of employees across dozens of countries. While these firms are critical and there is skill involved, on a time basis, almost all of their work is rote: rule-based, code-constrained. Think of the design of where to place fire alarms, power circuits, or HVAC units throughout a large buildout. Not exactly the kind of work dreams are made of.

Most of the revenue in this space flows from billable hours — manually producing designs like drawings, calculations, and documentation that a licensed engineer has to touch at every step. It’s so time-intensive that many of these firms have huge backlogs of latent demand they can’t address because they don’t have the people to do it. AI is the unlock to grow these services businesses, but even more importantly increase the rate at which society can build.

In some respects it’s quite similar to other large white collar spaces that have seen rapid AI adoption like legal, accounting, or financial research. A massive, rules-based market with legacy tools that weren’t built for the job, workflows bounded and deterministic enough to automate reliably, and a supply constraint creating genuine urgency.

The Business Model Shift That Makes This a $100B+ opportunity

You may be thinking ok, this sounds interesting, but how big can this really be? Autodesk’s Revit product is as close to a monopoly as you can get, and it’s doing a few billion per year in revenue.

The reason this is more than a niche software opportunity comes down to what AI is actually unlocking. Across all three layers, the bottleneck is the number of qualified people available to do the available work, and the speed at which they can operate. In other words, these are capacity-constrained fields in the world of atoms, which can actually be made much more efficient with the right software.

What makes the transition easier than it sounds is that large firms have already modularized this work. Unable to hire fast enough, most of the biggest firms have offshored the repetitive, lower-complexity tasks to large teams overseas. That work is already measured in output rather than hours, already priced as a discrete deliverable, and already sits outside the core workflow. AI doesn’t need to change how firms think about this work. It just changes who, or what, is doing it.

When AI expands capacity rather than just improving tools, the pricing model changes fundamentally. The opportunity in construction isn’t charging engineers for a better Revit. It’s aligning on a share of incremental fees, for efficiency gained, for speed to build, for the change orders prevented, or for the risk eliminated from a developer’s portfolio. These opportunities are orders of magnitude larger than the significant but not enormous budget line Revit already occupies.

This pricing shift is already happening. The first AI companies in this space are creating win-win arrangements with customers on results driven per project, not simply on a per seat basis. They’re measuring success in dollars of change orders prevented and percentage points of margin recovered.

We have to solve this problem

Right now, somewhere, an engineer is waiting for their Revit model to load. Or manually completing tasks that would cause a casual AI native observer to vomit. The consequences of these legacy tools are creating billions of dollars of waste and slowing GDP growth.

This is happening on every major construction project in the world, every week, in every city. The data centers going up to run AI. The hospitals being built to serve aging populations. The housing that cities desperately need.

The entire built world is still running on software built in 1997. That’s about to change. We’re investing accordingly.

If you’re building in this space, reach out to jschmidt@a16z.com, dh@a16z.com, cgoggins@a16z.com, and zabie@a16z.com

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Michael Spyker returns to a16z news! Last November Michael wrote a great two-part series that gave tech founders and investors a primer on all things natural gas. Today he explains everything we need to know about the energy crisis unfolding every day the Strait of Hormuz remains closed. Enjoy:

Think of it like this. The Persian Gulf is the us-east-1 of the global energy trade, and the Strait of Hormuz is the most critical global energy dependency. The consequences of a prolonged closure are not limited to higher gasoline prices; with enough time the pain will cascade into food, freight, fertilizer, petrochemicals, power, manufacturing, and eventually political stability.

Tech people understand hidden centralization better than almost anyone. A stale DNS record, a busted base image, secrets that weren’t rotated; a single failure can bring a ‘distributed’ stack to its knees overnight. That’s more or less where we are today with global energy markets. There is no failover for the Strait of Hormuz.

Entrenched

In a perfect world, nobody would consider the world’s reliance on energy produced in the Middle East to be a strategic oversight. By most metrics, it was a prudent global capital allocation decision in a vacuum.

A good shale well in the Permian Basin in Texas may recover about 0.5 million barrels of oil (mmbbls), while a strong well into Saudi Arabia’s largest oil field (Ghawar) can recover 100 mmbbls or more. The Middle East’s geology is uniquely advantaged; massive conventional oil accumulations that are highly overpressured and simple to produce have resulted in to the most prolific oil fields in the world.

Oil produced in the Middle East is some of the lowest cost on earth, and government ownership over the resource means their economic profits, which are ultimately used to fund their fiscal budgets, are far higher than the economic profits generated from a typical US shale well on a per-barrel basis (if the US producers generate any profit at all).

But a slightly higher breakeven didn’t stop the United States from taking the Permian Basin from zero, to outclassing Saudi Arabia’s crown jewel within a decade. When people say “shale was a massive net benefit to society”; this is what they mean. The US added millions of barrels per day of oil production that really shouldn’t have been “next in line” – all facilitated by the unmatched risk appetite of the American capital markets.

If we rewind to the 1970s (it’s important), we’d find ourselves at the founding of the Saudi Basic Industries Corporation. It was then that the Persian Gulf States began to appreciate the need to focus on monetizing the entire molecular value chain. This was accelerated by the 1998 oil price crash which made it viscerally clear that downstream integration was important for fiscal survival.

When China’s demand for conventional energy began to grow in the early-2000s, the Gulf States answered. Selling plastics feedstock to China allowed the Gulf to diversify their commodity exposure, but it made Asia reliant on Middle Eastern crude and liquids (spoiler: this eventually comes full circle). As China and Asia developed industrially, their refineries became reliant on, if not specifically designed to ingest Middle Eastern grades of oil. This was a symbiotic relationship – when refined, the finished product basket of Middle Eastern crude almost exactly matched the demand profile of local markets.

As Asia’s industrial buildout matured, US shale entered the scene, and its rapid production growth only sharpened the Gulf’s strategic imperative. If the US could flood the market with light sweet crude, the Gulf States would have to become more calculated. The Gulf countries needed to convert their hydrocarbons into whatever the world was short of – and leverage their scale and cost advantage to become reliable suppliers of those molecules.

All this resulted in Gulf Countries growing their LNG, LPG & condensate volumes, over brute-force raw crude. These lighter hydrocarbons find their way into India where LPG is widely used as a cooking fuel, Singapore where condensate is split into aviation fuel and plastics feedstock, South Korea where condensates are turned into specialty textiles, and even Taiwan where ~15% of their power grid is fueled with Qatari LNG. These are deep, hard to break, and not “pauseable” dependencies. For 55 days a year, Taiwan keeps its lights on thanks to Qatar. 400 million Indians use LPG to cook dinner every night. In Japan the taxi cab fleet runs on Middle Eastern autogas.

Kind of like how CUDA makes it impossible to move off the Nvidia stack – the Persian Gulf’s deliberate actions to build, internalize, and profit off the critical industrial processes that the world needed, has made their energy and materials output irreplaceable. Sure you could migrate to ROCm, but you won’t, and neither will Asia rebuild their ~38 mmbbls/d of refining capacity.

The Strait Jacket

And here we are today – the Persian Gulf is totally paralyzed, and with its attrition, a significant share of the world’s energy supply is completely stranded. Across every key hydrocarbon/commodity category, ~15% or more of the world’s production is unable to reach its normal-course buyer. And not just crude oil, its key industrial gases, refined fuels; even metals, fertilizers, and lubricants. Invisible things that are critical to a functioning earth. Remember, these are commodity markets, where pricing is set at the margin – and the loss of the Middle East is anything but marginal.

The impairment is being felt physically already in markets where the Gulf makes up a large share of global seaborne trade, like condensates and petrochemicals – and felt financially where the absolute volume loss is catastrophic enough to curb-stomp global pricing. The fallout has largely been limited to Asian countries that are reliant on just-in-time delivery of natural gas and crude oil. Substitutes, if available at all, are in the Americas and months away by ship – and if the Strait’s closure was to last that long, the world would no longer be facing a regional commodity deficit, there would be very real, unsolvable, global energy shortages.

Worst of all, nobody knows what’s happening. Visibility into activity through the Strait is deteriorating. Vessels are now transiting with AIS (Automatic Identification System) transponders switched off, leaving satellite imagery and fragmented media reporting as the only way to gauge true activity.

By our math, only 17 (maybe 18) vessels have crossed the Strait in the past 7 days, total! At normal levels, we’d expect ~50-60 vessels transiting daily, each way. Even counting the Iranian shadow fleet activity, and the new pay-to-cross scheme through Iranian territorial waters, activity through the Strait is still 95% lower than normal.

Saudi Arabia is the only country with a meaningful contingency plan by way of the East-West pipeline that runs across the map and terminates in Yanbu. Saudi Arabia has been successful in ramping their Yanbu export terminal to ~5 mmbbls/d, leaving them with ~2-3 mmbbls/d of crude and product exports that are impaired. Aside from Saudi Arabia, the UAE has diverted an incremental ~0.5-1.0 mmbbls/d to their Fujairah Terminal, and Iraq has routed ~0.2-0.3 mmbbls/d of crude through Turkey to be exported.

While the infrastructure on the map below may not look like much, the infrastructure on the map below represents ~15-20% of global liquids supply, depending on utilization, and ~10% of this supply can’t be exported any other way. The Al Basrah Terminal is ~3.8 mmbbls/d, and Iraq’s backup option couldn’t even handle 10% of Al Basrah’s volume. Kuwait is almost completely trapped, the same for Bahrain and Qatar. All these facilities, under normal circumstances would effortlessly load crude piped to the Gulf from inland fields onto tankers making the same, mechanical global laps. To return to normal would mean at minimum, ~20 tankers and gas carriers would have to pass each way (in & out) each day. Today we’re at ~2 tankers moving each way.

Insulated, Not Immune

Beside the price at the pump, the pain of the Strait closing hasn’t been fully felt by the West. Because North and Central America are net exporters of crude oil, liquids, and products, we don’t have barrels that are “unmatched.” But in a world where everyone is desperate for barrels, North America’s export position, and distance, can only act as a natural buffer for so long.

Prompt cargos from the US Gulf Coast have already begun trickling into Asia as the arbs got too generous to pass (that’s the point!). China has begun restricting product exports already, which leaves their buyers (Australia) scrambling. New Zealand, Australia, and Europe began with a broad SPR release which buys them ~2 weeks. The time for the Strategic Petroleum Reserve is indeed now, but it only buys time, and stock draws are not a fix to the Strait’s closure, which still desperately needs to come.

Financial Manipulation or Physical Desperation?

Rumors of the US Treasury intervening in the oil markets have spread like wildfire. Many watchers are questioning why West Texas Intermediate (WTI) crude oil is trading at such a discount to Brent crude (the European benchmark), and Omani crude.

Normally, Brent crude trades at a nominal ~$5/bbl premium to WTI, but last week that spread reached $20/bbl, and Omani crude changed hands at a $60/bbl premium to WTI! While dislocations that extreme might screen as manipulation, the physical crude market is, well… in extreme dislocation. These spreads are indicative of a healthy market, not manipulation – and the explanation lies in delivery timing.

Front-month WTI cannot reach Asia for ~60 days! The delivery window for prompt WTI contracts begins May 1^st, and those barrels still lie behind a 25-45 day tanker voyage to reach buyers that need oil today. This has put a significant premium on barrels that can be had sooner. Similarly, as physical flows abruptly change, the cost of freight has surged, with Argus’ VLCC freight index indicates it costs ~$10/bbl to move a barrel of crude today, compared to ~$5/bbl at the end of last year.

These little trifles make WTI less desirable – worse delivery timing, more expensive freight, and worst of all, a crude oil quality that Asian refiners aren’t designed for. Looking down the forward curve, the WTI-Brent spread normalizes to ~$5/bbl through late-spring, suggesting that prompt-month dislocation was indeed just a symptom of the most vital shipping lane on earth being blockaded.

Omani crude is loaded east of the Strait of Hormuz, and just days away from Indian and Asian refiners. Another hint that desperation is driving the dislocation – the spread between Brent 1^st line & Dated Brent reached a 5-year high at $7/bbl. The 1^st line spread represents the premium paid for near-term barrels over planned cargos weeks in the future. Logically, it makes sense that when ~15 mmbbls/d of oil production is turned off, people pay a lot for barrels available now.

Disruption Disbelief

When Russia invaded Ukraine, the market initially priced in a loss of ~2–3 mmbbls/d of global supply (<3%). That loss never fully materialized. Russian exports were rapidly re-routed to China and India, and production recovered within weeks. Physical supply remained largely intact, even as flows were reshuffled. The beauty of the global oil market is that it’s fairly malleable; as long as oil can reach the water, it will be worked into flows.

Despite a near-zero loss of Russian production, WTI crude traded >$120/bbl during the spring of 2022. The difference is duration. Many believed that broad global sanctions would make Russia’s production losses durable, and that Russia would not succeed in evading them. The market was ultimately wrong, and their supply losses were not durable.

Today, there is an observable physical stoppage — a material disruption in both production and exports. Unlike 2022 when production simply found a different home, today there’s a very real 10-15 mmbbls/d of oil missing from the global oil market: a number so large that it’s almost impossible for analysts to underwrite any solution that isn’t a speedy one. Unlike Russia, which easily found buyers for their crude outside of Europe, today’s supply shock is seeing a hostile third party come between the willing exporters of crude, and potential buyers.

While Saudi Arabia and the United Arab Emirates are using their alternative export terminals (Yanbu and Fujairah) to deliver as much supply as possible into the market, Strait crossings still remain impaired. Despite this, the implications of a prolonged closure are so severe that even with an immediate, visible, and unresolved blockage in the Strait of Hormuz limiting >10 mmbbls/d of production, the oil market is, all things considered, remaining relatively calm.

Glut-be-Gone, How the World is Backfilling Supply

Going into 2026 the US Energy Information Agency (EIA) had expected the global oil and liquids market to average ~2.5 mmbbls/d oversupplied. That narrative has quickly disappeared, and the world is beginning to draw down its strategic crude oil and petroleum product (i.e. diesel) stockpiles.

Last week, the International Energy Agency (IEA) announced its member countries would facilitate a collective release of ~400 mmbbls in liquids stockpiles. That covers roughly 1 month of lost Gulf supply. Coupled with the un-sanctioning of Iranian and Russian oil on water (~100 mmbbls), the world has backfilled ~45 days of lost Persian Gulf production (without accounting for transit times).

While Gulf producers initially diverted output into onshore tanks and floating storage, full field-level shut-ins quickly became unavoidable. By our math, the Gulf has shut in ~12.5 mmbbls/d of liquids production (including NGLs and refinery gains).

These are field-level losses, not export disruptions — barrels that didn’t come out of the ground. Compare that to 2022, when Russia invaded Ukraine; the cumulative loss of production in the 24 weeks following their initial military action was ~95 mmbbls. Just 24 days into the Iran war, and cumulative liquids production losses already tally ~285 mmbbls. In the forecast scenarios shown below, the black dashed base-case assumes a March 28^th resumption of full tanker activity, with our field-level production resumption forecasts moderating ramp times.

All said, even in a nearly ‘immediate opening’ scenario, where the world normalizes in 24 hours, there will be ~700 mmbbls of liquids that everyone expected to be produced in 2026 but weren’t. Going into 2026, markets expected to exit the year with global liquids inventories building by ~11%, instead, in the best case, inventories will likely decline by ~9%. Every week that the Strait’s opening is delayed, means another 60-85 mmbbls drawn from stocks; or fewer than that, if the world starts to run out of petroleum inventories. The “more worse” case scenarios are fairly mechanical. Every day the strait remains closed means fewer barrels available to consume; and further demand destruction, if not outright higher prices.

Chemical prices around the world are already going vertical, and with Qatar’s LNG trains offline it’s set to get worse. Condensate, a lighter form of oil, and primarily produced with natural gas, is a key building block in the petrochemical supply chain. Condensates, when refined, produce a high naphtha yield, molecules that then get turned into plastics.

Butadiene, the input for rubbery goods like disposable gloves, garden hoses, or even sneaker soles is up 140% YTD in China. Polyethylene terephthalate (PET), what soda bottles and takeout containers are made of, is up 45% YTD in Germany. Toluene, the base for nail polish remover, and various adhesives is up ~70% YTD in South Korea.

It can be easy to reduce the energy argument down to “oil bad, solar panels good,” but the non-obvious uses of crude oil and liquids are far deeper reaching than many appreciate — and easy to overlook. It’s equally easy to say “we need more energy” (an exceptionally brave take from the conventional energy analyst), but the closure of Hormuz should draw attention to the importance that hydrocarbons play beyond trucking and shipping.

The industrial processes that we rely on to be executed safely elsewhere rely on Gulf energy. The food that keeps the world alive relies on Gulf energy. Syringes, sequin blouses, antifreeze, car bumpers, LEGO^® bricks, carpets, diapers – these consumables are all downstream from crude oil production; and together account for more than 15 million barrels per day of petroleum consumption.

While the developing economies in Asia have the most exposure to the Strait of Hormuz, the developed economies of the West have certainly become used to the cheap manufacturing and other consumable exports from Asia. The early signs that higher petroleum prices are going to ripple through the system are showing up in plastic and other chemical feedstocks in Asia. As fertilizers get more expensive, and barrel deficits become more obvious, if the Strait of Hormuz doesn’t open, the carnage will no longer be contained to the East. The world’s petroleum trade is international, and with enough time, desperation, and lost barrels – global liquids flows will crisscross in response to the highest bidder. Today, time to ship between the US and Asia has been an effective circuit breaker. But when a fuse does blow, price eventually becomes the only clearing mechanism.

If you want to chat about energy projects, powering data centers, or anything else on topic, you can reach Michael at michael@htmog.com.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Sakina Arsiwala is a product and growth leader and former Twitch executive who helped scale Google, Youtube, and Nextdoor internationally. She is currently an a16z New Media fellow.

The YouTube Lesson: Content is a Geopolitical Weapon

Years ago, I PM’ed Google’International Search and then led YouTube’s international expansion, launching in 21 countries in just 14 months. I wasn’t just localizing a product; I was building local content partnerships and navigating the GTM minefields of legal, policy, and access. Most recently, I managed Community Health (Trust & Safety) at Twitch. Along the way, I founded two startups.

Today’s AI landscape feels eerily prescient of that early Google & YouTube growth. My career has taught me one truth: Globalization is not a product feature; it is a geopolitical negotiation. The hardest lesson was that distribution was never purely technical. Growth depended on local partners, cultural translators, and trusted community figures who bridged the gap between a global platform and a local audience.

I’ve lived through the GEMA blackouts in Germany, where a music rights agency essentially opted a whole nation out of the pan-EU YouTube rollout. I’ve lived through lèse-majesté arrest warrants in Thailand, where I couldn’t fly through the country because, as the face of YouTube, I risked arrest for content deemed defamatory to the King. I saw Pakistan bring down the entire internet to block a single video. I remember when our offices in India were physically attacked because a global algorithm collided with local sanctity.

What we were really navigating wasn’t just policy or infrastructure. It was trust friction.

In every market, someone had to absorb the cost of figuring out what was safe, acceptable, and valuable before users would engage. That cost compounds. And over time, it becomes a kind of Trust Tax, paid upfront by a small group, and amortized across everyone else.

Those same fault lines are now reappearing in AI—only now they are harder, faster, and more visible. A recent impasse between the U.S. federal government and Anthropic has triggered public debate, while OpenAI faces increasing scrutiny over its public-sector partnerships. We are witnessing a shift where user adoption is no longer just about utility—it is increasingly influenced by ideology. In this climate, trust is fragile; a single perceived erosion can trigger a rapid, mass migration of users.

While Google is doubling down on its embedded trust strategy, leveraging the familiarity of its existing ecosystem of Workspace and Search to bridge the gap, the global landscape is fraying. The EU’s stringent guardrails, China’s frantic development race, and the rising tide of AI Nationalism have put the world on alert.

The lesson for 2026 is clear: Institutional trust and cultural permission are now inseparable from the product itself. You cannot build an Intelligence OS without a bedrock of Trust.

This is the Sovereign Wall—the structural boundary where global intelligence collides with local control. But from a product perspective, it manifests as something more immediate: the Trust Wall.

Every global AI system now expands until it collides with it—the point where adoption is no longer determined by capability, but by whether users, institutions, and governments trust it within their own context.

The internet was borderless. Intelligence will not be.

Subscribe now

The Sunset of the Explorer Era

The first billion AI users were explorers and tech-optimists. But the Explorer era is over. For the last three years, we have lived in a time of prompt-craft and digital alchemy, visiting destination apps like ChatGPT or Claude as if they were digital shrines to witness the miracle of generation. In this era, the only metric that mattered was Model Parity: Who topped the latest benchmark? Who has the most parameters?

But as we cross into 2026, the campfire of the Explorer era is flickering out. We are no longer building toys for the curious; we are pivoting toward Intelligence Operating Systems,the invisible, ubiquitous pipes that must power the life of a solo-entrepreneur in São Paulo or a community health worker in Jakarta.

These users aren’t Explorers. They are Utility-Seekers. They don’t want to chat with a ghost in the machine; they want a tool that navigates the friction of their reality. This is the real Crossing the Chasm in the quest for the next billion users. And it is here, at the edge of the map, where the Silicon Valley Global API dream crashes into the hardest reality of our time: The Sovereign Wall.

The critical shift is this: AI adoption is no longer primarily a model problem. It is a distribution and trust problem. Frontier labs will continue to push capability forward, but the next billion users will not arrive because a model scores higher on benchmarks. They will arrive because the intelligence layer reaches them through institutions, creators, and communities they already trust.

2026 Reality: AI as a National Infrastructure Problem

In 2026, the challenge is no longer about making the model smarter; it is about making the model permitted. The Sovereign Wall is the point where General Intelligence meets National Identity. Across the world we are already seeing the early contours of this wall: data localization requirements, national AI compute programs, and government-backed model initiatives from India to the UAE to Europe. What began as cloud infrastructure policy is quickly evolving into intelligence policy. It is where a country refuses to be a Data Colony and demands that the intelligence powering its citizens lives within its own Sovereign Vaults, speaks its own lore, and respects its own borders.

When you see the CEOs of Google (Sundar Pichai), OpenAI (Sam Altman), Anthropic (Dario Amodei), DeepMind (Demis Hassabis) on stage with Prime Minister Modi at the India AI Impact Summit 2026, you are seeing the Sovereign Wall in real-time. PM Modi’s M.A.N.A.V. Vision (Moral and Ethical Systems, Accountable Governance, National Sovereignty, Accessible AI, and Valid Systems) is a clear signal: if frontier labs attempt a Direct-to-Consumer land grab, they will be regulated into obsolescence. Trust is the only currency that bypasses these borders.

The Weak Network Effect Problem and Why It Forces a New Playbook

Unlike social platforms, where each additional user increases value for every other user, AI’s value is largely local. My thousandth prompt does not directly make the system more valuable for you. The data flywheel improves the model, but the user experience remains personal, not social. AI is personal. It can be emotional. But at its core, it is a utility.

This creates a structural problem: AI does not benefit from the kind of compounding social network effects that powered the last generation of platforms. Without a native social graph, the industry defaults to a high-burn loop, chasing early adopters, power users, and the tech elite. That strategy works in the Explorer Era. It does not scale to the next two billion users.

And more importantly, it breaks down at the Sovereign Wall. Because when network effects are weak, trust does not emerge organically, it must be imported.

The Shift: From Network Effects to Trust Effects

If AI cannot rely on social network effects to drive adoption, it must rely on something else: Trust networks. This is the critical shift:

From Customer Acquisition → to Intermediary Empowerment

YouTube scaled because it rode on existing human networks of trust. AI must do the same. Instead of trying to build direct relationships with billions of users, the winning strategy is to:

• empower the people who already have those relationships

• leverage the trust they’ve already earned

• and distribute intelligence through those channels

Why This Matters

In a world shaped by the Sovereign Wall:

• Distribution is constrained

• Direct-to-consumer is fragile

• Trust is local, not global

Without strong network effects, AI cannot brute-force its way to scale. It has to flow through trust. AI doesn’t have network effects. It has trust effects.

The Solution: The Intermediary Era

How did YouTube actually gain a foothold in international markets?
Not by being a better video player. Not by localizing UI strings. We won because we became the platform of choice for people who already had local trust. In every market, adoption didn’t start with YouTube. It started with Identity Anchors: people and communities who already held cultural authority:

• A Bollywood fan page curating rare Shahrukh Khan clips for diaspora communities in Dubai

• Anime superfans in the US building deep lore ecosystems that mainstream media didn’t serve

• Local comedians, teachers, and remix artists translating global content into culturally legible formats

These creators weren’t just uploading videos. They were contextualizing the internet for their audience. They acted as Trust Brokers bridging the gap between a foreign platform and a local user. YouTube succeeded because it became the invisible infrastructure powering these Identity Anchors.

The Missed Insight: Direct-to-Consumer Hits the Sovereign Wall: Most AI companies today are still operating with a Direct-to-Consumer mindset:

Build a better model → expose it via chat → acquire users directly.

This works—until it doesn’t. Because in high friction markets, users don’t adopt technology directly. They adopt it through someone they trust.

YouTube didn’t scale globally by convincing billions of users one-by-one. It scaled by powering the people who had already earned the audience’s trust. That is what invisible infrastructure actually means:

=> You don’t own the relationship , => You power the relationship.

And at scale, that is far more defensible.

From Chat to Agents: Powering the Trust Brokers

This is where the shift from chat → agents becomes critical. Chat is a tool for individuals. Agents are leverage for intermediaries. If we take Anthropic’s Ami Vora’s framing “build for the most tired person”, in many markets, that person is the Trust Translator:

• the educator adapting foreign concepts

• the entrepreneur navigating local bureaucracy

• the community leader managing information overload

We win by solving their Trust Latency, the gap between global intelligence and local usability. That requires an Agentic Success Stack that actually does the work:

• For the Educator:
  Sora / GPT-5.2 transcreates curriculum—swapping an American football analogy for cricket, preserving meaning while aligning with culture

• For the Solo Entrepreneur:
  An agent that doesn’t just explain a Singapore tax form, but completes and files it via local APIs

• For the Community Leader:
  Contextual memory layered onto WhatsApp—turning 10,000 messages into structured action, preserving signal and community norms

Why This Works: Solving Trust Latency at the Last Mile:

To understand why this model scales, you have to understand Trust Latency. In many parts of the world, the bottleneck is not access to technology. It’s the time, risk, and uncertainty required to trust it. Technology isn’t adopted through ads, it’s adopted through vouching.

The mistake most AI companies are making is trying to pay the Trust Tax centrally—through branding, distribution, or product polish. But trust doesn’t scale that way.

The fastest path is to outsource the Trust Tax to those who are already paying it—creators, educators, and operators embedded in local contexts. This is someone who has already tried and errored on behalf of their audience figuring out what works, what breaks, and what actually matters in a local context. They absorb that risk so their audience doesn’t have to.

By empowering these Trust Brokers:

• CAC approaches zero: distribution rides on existing trust

• LTV increases: utility is locally relevant, not generic

• Adoption accelerates: trust is inherited, not earned from scratch

You effectively gain a global sales force you don’t have to pay for. One that is more credible, more efficient, and more deeply embedded than any centralized go-to-market motion. You are no longer building a product for users. You are building leverage for the people users already believe.

That is how YouTube scaled globally. And it is how AI will cross the Sovereign Wall.

2. Sovereign Vaults: The Geopolitical Moat

The endgame for Marc Andreessen’s Techno-Optimism isn’t fighting regulation; it’s productizing it. The race against China’s DeepSeek and Kimi isn’t won by ignoring borders; it’s won by owning the vault.

What is a Sovereign Vault? It is a residency-first, localized instance of a model that lives within a nation’s Digital Public Infrastructure (DPI).

• The Geopolitical Moat: By giving nations like India or Brazil digital sovereignty over their models, weights, and data, we fundamentally shift the balance of control. Intelligence is no longer mediated by foreign platforms; it is governed within national boundaries. This doesn’t “block” foreign adversaries outright—but it raises the cost of influence, reduces external dependency, and limits the surface area for control, extraction, or unilateral intervention.

• The Identity Anchor: By anchoring our models in local lore and legal reality, we build a moat that General Intelligence can’t touch.

• The Feedback Loop: Solving for the hyper-local nuances of a Malaysian tax permit isn’t a distraction; it’s a Model Accelerator. It provides the Cultural Elasticity our base models need to remain the world’s most capable intelligence.

There is a real tension here. The promise of AI is universal intelligence, but sovereignty pushes the ecosystem toward fragmentation. If every nation builds its own stack, we risk a patchwork of incompatible systems, uneven safety standards, and duplicated effort. The challenge for frontier labs is not simply to scale intelligence, but to design architectures that allow local control without collapsing the benefits of global capability.

Three Structural Shifts in the Intermediary Era

1. AI Distribution Will Move into Existing Trust Networks

AI will not scale through standalone apps. It will embed into messaging platforms, creator workflows, education systems, and small business infrastructure—because trust is already established there. Without strong network effects, distribution must ride on existing human relationships.

2. National AI Infrastructure Will Become a Default Requirement

Governments will increasingly require local model hosting, sovereign compute, or regulatory oversight for critical AI systems, accelerating the emergence of sovereign vault architectures.

3. The Creator Economy Will Become an Agent Economy
Creators will not just produce content—they will deploy agents that perform real work for their communities. These agents will act as extensions of trusted individuals, inheriting their credibility and distributing intelligence through trust networks.

There is, of course, another possible future. A single dominant assistant could emerge. Deeply embedded in operating systems, browsers, and devices creating a direct relationship between user and model that bypasses intermediaries entirely. If that happens, the trust layer collapses into the assistant itself.

But history suggests something more pluralistic. Even the most dominant platforms—from mobile operating systems to social networks—ultimately grow through ecosystems. Intelligence may be universal, but trust remains local. Whichever architecture ultimately wins, the underlying challenge remains the same: AI adoption is no longer primarily a model problem. It is a distribution and trust problem.

Conclusion: Niche is the Only Real Global

The greatest fallacy of the Explorer Era was the belief that intelligence is a commodity—a single, global API that would manifest identically in a Manhattan boardroom and a village in Karnataka. The Sovereign Wall has exposed a harder truth: intelligence may be universal, but adoption is not.

Nations and local institutions don’t want a black-boxed external system. They want control, context, and the ability to shape intelligence within their own boundaries. They don’t want the application—they want the pipes. The infrastructure, the security, and the compute that empower their own citizens to build for themselves.

Growth in 2026 is no longer about discovering a universal UX. It is about product elasticity—the ability for intelligence to adapt to local context, regulation, and culture without losing capability. If we continue to chase the global consumer directly, we remain a foreign layer—fragile, replaceable, and subject to the same disruptions I saw at YouTube.

But when we pivot to Intermediary Empowerment, the model fundamentally changes. We move from chat to agency, from convincing users to empowering Trust Brokers, and from fighting regulation to turning it into a moat. AI doesn’t scale through models—it scales through trust.

The winner of the AI race will not be the company with the smartest model. It will be the company that is most effective at making the local hero—the teacher, the accountant, the community leader—ten times more powerful. Because in the end, intelligence travels through systems, but adoption travels through people.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

The entire modern world runs on an invisible resource.

Every time a cellphone connects to a tower, a satellite beams data to Earth, a drone takes flight, or a soldier calls for support—they are relying on radio frequency (RF) Spectra. The RF spectrum is, in a quiet and underappreciated way, the central nervous system of the 21st century. It is the unseen highway that carries the data on which our economy and our national security depend.

But what is the radio frequency spectrum?

For the fellow RF engineers out there, it is the finite range of electromagnetic frequencies from 3 kHz to 300 GHz where we exploit physics to juggle everything from submarine pings to 4K Instagram streams without them crashing into each other.

But for now, you can think of it like a system of real estate in the airwaves. Every wireless signal occupies a specific band of frequencies, a tiny slice of the RF spectrum. So many systems require spectrum for different purposes that governments have to divide up the spectrum, granting licenses to dictate who can transmit what, where, and when. You can’t simply start transmitting on a frequency because you want to; you need authorization, and that authorization flows through federal agencies—in the United States, the Federal Communications Commission (FCC) and the National Telecommunications and Information Administration (NTIA). And the entire wireless economy runs on this essential system of allocation and coordination.

And the number of devices and systems that rely on spectrum is exploding. Beyond denser 5G (and, soon, 6G) networks, we’re seeing an unprecedented proliferation of autonomous systems: everything from commercial drones to self-driving cars to industrial robots on factory floors. In order to operate safely, each and every one of these autonomous systems depends on reliable access to RF spectrum. And this massive growth in usership is making spectrum a more crowded and complex resource than it ever was before.

And these systems need spectrum access on timelines that the current systems were simply never designed to support.

The systems that manage who gets to use which frequencies when and where are entering a slow-moving crisis. Not in an immediate or dramatic way, where it’ll stop working next week and your garage door suddenly won’t be able to open; but in the gradual, grinding way that most crises actually develop, where every year the gap between what the technology could do and what it actually does widens ever-so-slightly.

The bottleneck is the system that sits between demand and supply. The FCC and the NTIA have brilliant engineers, lawyers, and public servants. But they’re working with tools from a different century.

Airbase is the partner the U.S. government needs to build the wireless provisioning system required for the future for RF infrastructure.

A resilient, abundant wireless future

We founded Airbase because the greatest threat to our wireless future is a lack of access and coordination. At True Anomaly, Planet, and Boeing, we watched as cutting-edge innovations such as new satellite constellations, advanced autonomous systems, and critical defense capabilities were grounded not by physics but by paperwork.

We saw a world where the demand for data, and the number of devices, is exponential, but the infrastructure managing it is stuck in the analog era. We realized that if we don’t fix the foundation for spectrum provisioning, the next generation of technology simply will not come to pass—at least not on American shores.

We founded Airbase to solve three critical challenges:

1. Spectrum is Constrained: Physics imposes hard limits on spectrum access. There is a finite amount of radio frequency spectrum to go around. And the demand is exploding. From 5G to LEO mega-constellations, everyone is fighting for the same beachfront property. The traditional answer has been to slice the pie ever-thinner, or to operate on a first-come first-served basis. We believe that the combination of intelligent software and hardware provides a better answer.

2. Spectrum is Underutilized: Vast swaths of spectrum sit idle because of rigid, decades-old allocations managed by static databases and PDF forms. We are stuck orchestrating a dynamic, real-time world with administrative tools from the 1980s. We want to replace that friction with dynamic, software-defined agility.

3. Spectrum is Vulnerable: In an increasingly contested world, spectrum is a battlefield. Commercial and governmental assets are fragile and targeted constantly. They rely on frequencies that are often jammed or contested. We believe that without a real-time, intelligent layer to manage these assets, America’s critical infrastructure is vulnerable to foreign threats.

Airbase’s mission is simple: to transform spectrum from a static utility into a software-defined resource. We believe that bandwidth should be negotiated by machines in milliseconds and not by humans in months or years. And we believe that spectrum should be fluid, adapting to the needs of the moment rather than trapped within rigid licenses.

Now is the time

We are at an inflection point. The commercial space economy is taking off. The nature of warfare is shifting to the RF domain. The devices in our pockets are becoming hungrier for data.

The speed of innovation has lapped the speed of human coordination.

We founded Airbase to be the bridge. To bring the clarity of modern software to the chaos of the airwaves. To ensure that whether you are launching a rocket, deploying a 5G network, or defending the nation, you have the freedom of action to operate without constraint.

And most importantly, this vision of the future is becoming reality. Federal regulators are already using our software to analyze and coordinate spectrum access. Airbase is already working directly alongside the U.S. government, building tools for the problems that they face each and every day. And there is so much more to come.

The invisible infrastructure of our world is in crisis. And we exist to fix it.

Welcome to Airbase.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

If you don’t know it, pmarchive.com is a good repository of Marc’s old posts from years back. Today, we’re rerunning a 3-part essay, “The truth about venture capitalists”, a time capsule that takes us all the way back two years before a16z was founded, and gives a glimpse into why Ben & Marc were motivated to found a16z.

Just as we ‘go abroad in order to understand your home country better’, it’s similarly useful to go ‘abroad’ to other periods of time within our industry, in order to understand how we got to now - and appreciate how much we have yet to change.

Part 1:

A lot of people have opinions about venture capital—the pros and cons of VC, whether or not to take VC, which venture capitalists to take money from, how to get VCs to invest in your company, whether VCs are seasoned risk-taking professional investors or psychotic entrepreneur-hating sociopaths, etc.

Often these opinions are based on one individual’s specific personal experience with venture capital, and often based on someone’s negative experience—as is often the case, people who have negative experiences are more motivated to tell others than people who have positive experiences.

With that in mind, I will try to provide my hopefully broad perspective on the topic.

I’ll just say up front that I don’t think my point of view on this is any more valid than that of any of my fellow entrepreneurs—everyone’s experience is different, and this is definitely a topic where reasonable people disagree.

My experience with venture capital includes: being the cofounder of two VC-backed startups that later went public (Kleiner Perkins-backed Netscape and Benchmark-backed Opsware); cofounder of a third startup that hasn’t raised professional venture capital (Ning); participant as angel investor or board member or friend to dozens of entrepreneurs who have raised venture capital; and an investor (limited partner) in a significant number of venture funds, ranging from some of the best performing funds ever (1995 vintage) to some of the worst performing funds ever (1999). And all of this over a time period ranging from the recovery of the early 90’s bust to the late 90’s boom to the early 00’s bust to the late 00’s whatever you want to call it.

I’m starting to understand why I don’t have any hair left.

Subscribe now

The most important thing to understand about venture capitalists is that they are in business to do a very specific thing.

They raise a large amount of money—often $100 million or more—today, in order to invest in a series of high-risk startups over the next small number of years—usually 3 to 4 years.

The legal lifespan of the fund is usually 10 years, so that’s the absolute outer limit on their investment horizon.

They generally intend, and their investors generally expect, to have the returns from those startups flow back within the next 4 to 6 years—that’s their realistic investment horizon.

Within that structure, they generally operate according to the baseball model (quoting some guy):

“Out of ten swings at the bat, you get maybe seven strikeouts, two base hits, and if you are lucky, one home run. The base hits and the home runs pay for all the strikeouts.”

They don’t get seven strikeouts because they’re stupid; they get seven strikeouts because most startups fail, most startups have always failed, and most startups will always fail.

So logically their investment selection strategy has to be, and is, to require a credible potential of a 10x gain within 4 to 6 years on any individual investment—so that the winners will pay for the losers and in the timeframe that their investors expect.

From this, you can answer the question of which startups should raise venture capital and which ones shouldn’t.

Startups that have a credible potential to be sold or go public for a 10x gain on invested capital within 4 to 6 years of the date of funding should consider raising venture capital.

Most other startups should not raise venture capital. This includes: startups where the founders want to stay private and independent for a long time; startups where there’s no inherent leverage in the business model that could result in a 10x gain in 4 to 6 years; and startups working on projects with a longer fuse than 4 to 6 years.

Notably, there are many fine businesses in the world—many of them highly profitable, and very satisfying to run—that do not have leverage in their model that makes them suitable for venture capital investment.

By leverage in this context, I mean: the ability to make something once (a piece of software, a chip design, a web site) and sell it (directly or indirectly) to a lot of people (1,000 business customers or 10 million consumers)—which leads to the classic “hockey stick” revenue projection.

Venture capitalists shouldn’t, and can’t, invest in companies that don’t hit these criteria—not because they’re not good businesses but because their own investors wouldn’t stand for it.

There are also many fine entrepreneurs in the world who want their companies to stay small, or who don’t want to sell their companies or take them public. That is also well and good, and those entrepreneurs should not raise venture capital.

On the other hand, a business that is built for leverage that could be sold or go public in 4 to 6 years should strongly consider raising professional venture capital, for three reasons:

First, you get the cash to invest in the business and grow it at the speed required to realize its full potential.

It’s satisfying to say you don’t want to deal with VCs and you want to do it on your own, but if your business has the potential to get big, in my view you should take the cash to invest to make it as big as you can, and that usually requires more capital than you can raise from bootstrapping or from angels.

Second, you get that cash from a professional investor who invests in this kind of business as her full-time job and reason for existence in the world.

Most other possible investors in a high-growth startup will be much more difficult to deal with than a professional venture capitalist.

Third, in the best case, you will get help building your high-growth business from the venture capital partner you take money from (but see more on this in Part 2).

When a venture capitalist turns you down, it isn’t personal and it isn’t (usually) because she’s stupid. Instead, it’s often for one of these reasons:

One, she can’t see the leverage—she can’t see you getting to a sale or IPO with a credible prospect of a 10x return within 4 to 6 years. If she can’t see this, and 10 of her peers at other firms can’t see it, then you may want to revisit your fundamental business model assumptions and try to understand what’s missing.

Remember, it’s in her best interest to see the full potential in your business—she is looking for high-potential startups in which to invest.

Two, she thinks that what you’re doing is too early or unproven.

This is the one that drives entrepreneurs nuts. Isn’t the whole point of venture capital to make risky investments in unproven technologies and markets?

Unfortunately, that’s life—sometimes things are simply too early for venture capital. In that case, develop your idea further with bootstrap or angel funding and then take it back to the VCs later with more proof points.

Three, she isn’t convinced that you’ve assembled the right team to go after the opportunity. This usually means she doesn’t think your technical founder(s) are strong enough, or she doesn’t think your founding CEO is strong enough. Again, it’s in her best interest to see the potential in the team if it’s there—so if she and 10 of her peers pass on your startup because of concerns about the team, then you may want to rethink your team.

There are many other reasons in addition to these that a VC may pass on your investment that have nothing to do with you:

She loves it but she can’t talk her partners into it—which happens.

She’s fully committed and doesn’t have time to take on a new opportunity.

It would require traveling and she can’t or won’t do that.

You’re in a market she doesn’t know much about.

Or, she had a bad experience with a similar investment in the past.

The frustrating part is that she won’t always tell you why she’s passing—in large part because she wants to keep the door open to investing at a later date if things change (i.e. if it becomes clearer that you have a home run on your hands).

For that reason, whenever a VC passes and explains why, no matter how mean or unfair they sound, the best response is to thank them for their honesty.

Part 2

Comparing venture firms, and comparing partners within firms:

When raising venture capital, remember that venture firms vary wildly in style and quality.

For example, some venture firms are very entrepreneur-friendly. Others are notoriously brutal.

Interestingly, financial success in the venture capital profession does not seem to be correlated to entrepreneur-friendliness.

Individuals (partners) within each venture firm vary wildly in style, personality, knowledge, experience, ability to be helpful, drive, and ethics.

Personally I’d recommend being more focused on picking the right partner than picking the right firm.

This is slightly counterintuitive advice—and firm quality does matter—but the partner is the person you’re going to be working with. The other people at the firm you will see probably twice in the whole lifespan of your company.

Best of both worlds is to pick a strong partner at a strong firm, but be aware going in that even strong firms have weak partners.

Venture capital professionals arguably used to be a more homogeneous group: the founders and pioneers of the business, and their hand-picked proteges who had grown up as venture capitalists under close supervision and with rigorous training.

The explosion of venture capital in the late 90’s has led to a much broader range of people becoming partners in venture capital firms.

Many partners today have little venture capital experience but come out of an operating background (an executive role at a big company, or experience as an entrepreneur with their own startup), or come out of some other background (corporate attorney or executive recruiter), or come straight out of business school with no meaningful experience whatsoever.

There are pros and cons to working with any of these kinds of partners.

For example, VCs with operating experience are great when it comes to sitting down and talking about how to run a business, but sometimes they have less perspective (because their career was probably focused on one or two companies, whereas a professional VC has probably invested in 30+ companies), and they may have trouble keeping their hands off the steering wheel.

A VC with an executive recruiting background can be incredibly helpful at recruiting—one of the main areas in which a VC can add value (see below).

And a VC who used to be an attorney can be very helpful when you need to get a parking ticket fixed.

But there’s probably still no substitute for the VC who has been a VC for 20 years and has seen more strange startup situations up close and personal than you can imagine.

A venture capitalist’s ideal investment:

A venture capitalist’s ideal investment is the one that would be a huge success without her.

How much help, and what kinds of help, you can expect from your VC:

Assuming your startup does not fall in the category of a VC’s ideal investment, what kinds of help can you hope for from your VC?

First, it’s important to really internalize that the founders of a startup are the ones who have to make a startup succeed.

Odds are, nothing your VC does, no matter how helpful or well-intentioned, is going to tip the balance between success and failure.

In addition, VCs are—usually—incredibly busy people. Sitting on as many as a dozen boards, sourcing new investments, tracking the fast-moving technology industry, raising money and managing their LPs, and pitching in on their partners’ companies and their problems takes up a lot of time.

(Although every once in a while you will run into a VC who is lazy as sin—but that’s a topic for another post, when I’ve had more to drink.)

The best assumption to make is that your VC’s primary value add is the cash they are investing.

Then you’ll always be surprised on the upside.

Additional areas in which a VC can help include: recruiting, strategy, partner introductions, customer introductions, additional fundraising, and generally being a good sounding board and source of advice and industry knowledge.

Some firms run incredibly helpful programs such as forums in which new consumer Internet startups can interact with major advertisers, for example.

The only real way to find out how much help you can really expect from your new VC is to ask the founders of other companies funded by that same partner.

Finally, never expect the help to just happen unsolicited. If you want it, ask for it proactively.

You actually don’t want a VC who provides too much help without being asked. I leave the why as an exercise for the reader.

How VCs spend an awful lot of their time, and why you should feel sorry for them:

My friends who are VCs seem to spend a surprising amount of their time working with their failing companies.

The reason goes right back to the definition of a VC’s ideal investment: their winners are succeeding—they don’t need very much help.

Some of the best VCs in the industry spend most of their time on their successes, helping to boost them to higher and higher levels of success.

But generally, life in the VC trenches seems to consist of trying to jumpstart or otherwise fix fatally flawed startups.

Can you imagine how un-fun that would be?

Venture capitalists: soulless and rapacious capitalists, or surprisingly generous philanthropists? Or both?

Here’s something surprising about venture capitalists: their primary job is often helping very worthwhile nonprofits build larger and larger endowments to be able to continually make the world a better place.

The largest investors in many top-tier venture capital firms are nonprofits—particularly universities and large philanthropic foundations.

This is partially because such institutions are very patient investors and have very long time horizons. But this is also partially because many of the top venture capitalists feel a real sense of obligation and mission to help such vital organizations grow and flourish.

Traditionally this has been hard to see because venture capital firms have wrapped the identities of their investors (limited partners, in the lingo) in confidentiality agreements. But more recently, via certain SEC filings and disclosures by public universities, it has become possible to get a glimpse into the investor bases of some of the world’s best venture capital firms.

For example, it has been previously well publicized that you can see who Sequoia’s investors are by reading the SEC disclosure on the Google/Youtube acquisition.

You can see in that filing that Sequoia’s major investors include such universities as Amherst, Brown, Colby, Columbia, and Dartmouth—and that’s just into the D’s. Similarly, philanthropic foundations invested in Sequoia include the Ford Foundation, the Moore Foundation, the Irvine Foundation, the Rockefeller Foundation, and the Hewlett Foundation.

This is not unusual.

The best VCs get to improve society in two ways: by helping new companies take shape and contribute new technologies and medical cures into the world, and by helping universities and foundations execute their missions to educate and improve people’s lives.

Why we should be thankful that we live in a world in which VCs exist, even if they yell at us during board meetings, assuming they’ll fund our companies at all:

Imagine living in a world in which professional venture capital didn’t exist.

There’s no question that fewer new high-potential companies would be funded, fewer new technologies would be brought to market, and fewer medical cures would be invented.

We should not only be thankful that we live in a world in which VCs exist, we should hope that VCs succeed and flourish for decades and centuries to come, because the companies they fund can do so much good in the world—and as we have seen, a lot of the financial gains that result flow into the coffers of nonprofit institutions that themselves do huge good in the world.

Remember, professional venture capital has only existed in its modern form for about the last 40 years. In that time the world has seen its most amazing flowering of technological and medical progress, ever. That is not a coincidence.

How to make a VC’s head explode, in one easy step:

Point out to her that her compensation from carried interest should be taxed as ordinary income, not capital gains, since she’s receiving a fee for service and it’s not her capital at risk.

Part 3

The current venture capital environment in the United States is characterized by a very large number of venture firms (866, according to the National Venture Capital Association), investing an extraordinarily large amount of capital (over $7 billion in the first quarter of 2007 alone, according to Price Waterhouse).

Traditionally the venture capital industry was said to experience a “seven fat years, seven lean years” model—seven years of boom, followed by seven years of bust.

Following that pattern, the late 60’s/early 70’s were great (the “-tronics” boom—this is when Intel was funded by Arthur Rock), the mid-70’s were terrible, 1978-1985 was great (the PC!), ’86-92 was terrible, ’93-99 was fantastic, and ’00-06 was not so good.

As you’d expect, inflows of capital to venture firms during the lean years typically shrank dramatically—venture capital returns are terrible during the lean years, and who in their right mind wants to put more money into an investment vehicle with terrible returns?

This capital inflow shrinkage would then lead to a significant percentage of venture firms closing their doors (technically, not raising new funds—the venture capital firm equivalent of going under)—especially the newer, less proven ones.

Ultimately, capital inflows would shrink to the point where the remaining venture firms were managing a much smaller base of cash, which primed the pump for dramatic investment returns over the next seven fat years: less capital + a new wave of high-growth startups = explosive investment returns.

That is the cycle that has played out every time—except this time.

Let’s examine that $7 billion invested by venture firms in the first quarter of 2007.

Annualized, that is an annual investment rate of about $28 billion per year.

Pulling the data on venture capital investing by year over the last 10 years, we see that this is, as you might expect, substantially lower than 1999’s $54 billion and 2000’s $105 billion (what a year!)...

...but, higher than 1997’s $15 billion and 1998’s $21 billion.

And that rate of investment has been broadly consistent for the last several years—in fact, it’s been trending up.

Even when you adjust for inflation, venture capital funding is flowing into venture firms and out to startups at a higher rate in 2007 than in 1997 and 1998.

Those of you remember 1997 and 1998 will remember that those were true boom times for venture capital. The returns on funds from ’93-95 were extraordinary—some of the best ever—and limited partners were shoveling money into venture firms, leading VCs to fund new companies as fast as they possibly could.

Yet, despite disastrous venture returns on average from 2000-2006, the cash spigot from investors in venture capital firms continues to be wide open to a level where VCs can be more active in 2007 than they were in 1998.

While some older, stale venture firms have recently shut down—Sevin Rosen and Yankee come to mind—the rate of venture firm death has not been anywhere close to what you’d expect, and in fact many new funds have been formed and raised money in the last few years.

And the cash just keeps on coming.

Somehow we’ve ended up in a paradox: venture capital returns, on average, have been terrible, but contrary to historical precedent, the money keeps flooding in, venture firms keep going, and you have more money chasing deals than you did in the middle of the dot com boom.

How can we explain this?

In a nutshell:

Institutional investors who invest in venture funds—large university endowments, philanthropic foundations, and pension funds—began radically shifting their investment strategies in the early to mid 1990’s, and that shift has led to private equity generally, and venture capital specifically, becoming a permanent “asset class” for those investors.

I call this the “asset-classization” of venture capital.

Here’s how it works:

A large institutional investor like a university endowment runs its investment strategy with a top-down approach that says, we’ll put x% in stocks, y% in bonds, and so on—this is called asset allocation.

The actual details of which stocks, which bonds, etc. are less important—the big decision is what percentage of the total capital to put in which asset classes, because when you run a huge pool of capital, that’s what mathematically drives your returns. (You can’t put enough money into any single investment to really move the needle, at least not without being irresponsible, so you have to think in broad strokes—in terms of asset classes.)

Traditionally, such large institutional investors were quite conservative. An asset allocation that was perhaps 60% US equities, 30% US bonds, and 10% cash would not have been unreasonable.

Really daring institutional investors might have allocated some percentage to non-US equities, or (gasp) high-yield “junk” bonds.

This all started to change in the late 80’s and early 90’s when a group of advanced thinkers, such as David Swensen, then and now head of the Yale University endowment, crunched the numbers and realized that if they had a long-term time horizon (which they did—Yale and its peers are expected to be around for some time), they could generate higher returns by allocating more of their capital to so-called “alternative asset classes”—basically, anything other than public stocks, bonds, and cash.

This meant hedge funds, real estate partnerships, commodities, timber, leveraged buyout firms—and venture capital.

To read about this new strategy—and it is a fascinating strategy—pick up a copy of David Swensen’s excellent book Pioneering Portfolio Management, which describes his approach in detail. (Be sure to also pick up a copy of his book for individual investors—Unconventional Success—which explains why you can’t pursue this strategy without getting your clock cleaned.)

The institutions such as Yale and its peers that adopted a Swensen-style strategy did fantastically well in the 1990’s, and outperformed (technically, “kicked the ass of”) any institution that had an older, more conservative investment policy.

This predictably led a significant number of institutions to shift massively into alternative investments and venture capital in the late 90’s, just in time to get hammered by the crash of 2000-2002.

Here’s the interesting part: that hammering—by people who, say, only started investing in venture funds in 1999—has not resulted in a significant pullback on the part of institutional investors from venture capital.

Instead, venture capital has become an apparently permanent asset class of many large institutional investors—and increasingly, smaller institutional investors.

Those institutional investors are managing so much money—literally trillions of dollars—that even a very small asset allocation to venture capital represents an enormous amount of cash—tens of billions of dollars per year.

An organization called NACUBO (don’t ask) tracks asset allocation behavior of university endowments, and tells us that the average large ($1+ billion) university endowment had a 3.5% asset allocation to venture capital in 2006.

3.5% of a ginormous amount of money is a lot of money.

But it’s a small percentage of the total base, so apparently what’s been happening is that although returns on venture capital have been poor (technically, “sucking”) for the last several years, institutions that invest in venture capital are not taking that much actual pain on their overall asset bases, and they don’t see that many better alternatives, and so they’re sticking with their overall asset allocations and therefore sticking with venture capital.

You can argue that this is smart—that such institutions are very well set up for the next venture capital boom, and that they will do very well over the next 10-20 years with this strategy—versus the old approach of pulling out just before the sector was set to boom again.

You can also argue that this is not smart—that this is leading to more venture dollars chasing few good deals and long-run terrible returns for everyone. Particularly since historically, most of the positive returns for venture capital have gone to the top 10% of venture firms—or maybe even the top 10 venture firms—and most of the money going into venture capital as an asset class by definition is going into the other 90% of venture firms.

But regardless, it does seem to be the case.

And that’s why, from where I sit in Silicon Valley, there are probably 200 venture capital firms within 20 miles with likely over $20 billion of capital at their disposal chasing a very small number of good potential investments, despite terrible average returns for the asset class over the last seven years.

I love this country.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Robotics is keeping out the people it needs most. Over the last few years, the technology has advanced faster than most people expected, and yet real-world deployments remain stubbornly rare. That gap is not just a research problem. It is a people problem. The field has spent decades optimizing for a narrow profile of contributors, and now that deployment is finally within reach, it doesn’t have the builders required to get there.

If the goal is to see intelligent robots actually deployed and augmenting human labor at scale, then robotics needs fewer roboticists per capita. Not fewer people working on robots, but more operators, more people who care about customers and reliability, and more outsiders. It needs to look less like a research subfield and more like an industry.

While robotics is a broad field, this essay is primarily about robotic manipulators that learn from data-driven approaches rather than follow pre-programmed routines. This is where the gap between research promises and real-world deployment is currently widest.

The Moment Is Now

Most roboticists would agree real-world deployments remain the ultimate goal. The potential of robots is unlocked once they are doing useful work, repeatedly, in the messiness of the real world.

In practice, however, many robotics startups inherit the culture of research labs rather than of businesses. In robotics, novelty earns prestige. Reliability and customer obsession earn revenue. The two are rarely optimized together.

Intelligent robotic manipulators have been “almost there” for decades, and each generation felt one breakthrough away from robust generalization and real-world usefulness. Optimizing purely for research is rational when nothing is working. Hardware was expensive and bespoke, teleoperation was clunky and didn’t scale well, and robust and flexible learning-based techniques didn’t exist. If the core stack is brittle, deployments are premature.

What’s different now isn’t ambition. Robotics has always aimed for real-world deployment, but now, the tools are crossing thresholds required to make deployments viable. Pre-trained models provide strong priors for manipulation, and a relatively narrow post-training recipe built around behavior cloning and DAgger has emerged as a path to high success rates on targeted tasks. Vision-Language-Action (VLA) and Video-Action models (VAMs) are beginning to generalize beyond narrow demonstrations, showing early signs of adaptability across objects, environments, and task variations. At the same time, a new tier of low-cost hardware has emerged. Capable research-grade robotic arms are now available for a fraction of what platforms Franka or UR5e cost, moving robotic systems closer to economic viability.

Robots still fail in deployment. Today’s available models seldom meet cycle time requirements, they have limited memory, and our understanding of post-training and robustness is in its adolescence. But we’re no longer blocked by “nothing works.” The blocker is building systems that prioritize reliability, ease of integration, iteration speed, and unit economics. That is a different kind of problem, and it requires a different kind of builder.

Subscribe now

The Missing Application Layer

The intelligence stack in robotics is improving rapidly, but there is a missing layer between current capabilities and economic impact. In software, foundation models unlocked value because companies built application layers on top of them, products integrated into workflows, sold to customers, and iterated in production. Robotics needs an equivalent layer.

Even if autonomy worked perfectly for every use case, deploying a robot would still require custom hardware, robust telemetry and monitoring, purpose-built sensors, custom end effectors, remote teleoperation infrastructure, fleet management tooling, safety systems, and escalation pathways. Waiting for autonomy to be perfect before building these systems is backward. These systems are what make autonomy economically meaningful in the first place.

Labor is the product, not full autonomy. If a robot can deliver reliable labor, even with teleoperated fallbacks, it creates value immediately. Yes, full autonomy reduces the teleoperator time required to deliver that labor, driving down operating costs over time. But that is a margin improvement, not a prerequisite for deployment.

Reframing the problem this way changes how companies are built. Teams can deploy today, learn customer workflows, develop custom hardware and software stacks, and build operational expertise while riding the future wave of autonomy improvements. The moat for a deployment-layer company comes from relationships, domain knowledge, and integration depth within a specific vertical, not just model performance.

The application layer gap is not just a technical gap. It is a people gap. Building it requires operators who understand customer workflows, engineers who care about reliability over novelty, and product builders who can translate messy real-world constraints into systems that actually ship. These are not the people robotics has historically emphasized. They are exactly the people the field’s current culture screens out.

The Talent Bottleneck

The field of robotics has a gatekeeping problem that the current moment can no longer afford.

One of the things that surprised me most while building my last company was how quickly bachelors-level interns were able to contribute. The work was hard, but they excelled because they were living inside the latest papers in a way that most industry engineers simply aren’t. University labs and clubs have a kind of information velocity that’s easy to underestimate from the outside. New techniques propagate fast, everyone is tinkering, and the culture rewards trying things before they’re proven. The interns who came out of those environments arrived with deep intuitions about methods that were only months old. They got up to speed and started contributing faster than anyone expected.

That experience shaped how I think about the cultural problem in robotics. Interest in the field has surged over the last couple of years. Where do these people go? Hiring has grown, but it’s been narrowly concentrated. A handful of large humanoid companies like Tesla, Figure, and 1X are absorbing much of the available talent. Outside of those gravitational wells, hiring has largely remained conservative and credential-heavy. I’ve met countless operators, software engineers, product builders, and ML engineers from adjacent fields who want to work on robotics but struggle to find a way in.

Graduate degrees have become soft requirements to touch a robot. Experienced roboticists are treated as scarce specialists to be insulated rather than force multipliers who can mentor and scale teams. The result is slower execution and fewer entry points for new contributors, at precisely the moment when the field needs more builders. Communities like Hugging Face have lowered the barrier to entry, but hiring practices haven’t fully caught up.

The irony is that narrow specialization is increasingly a liability. Many of the most important techniques are only months old. Most of the practical knowledge lives in hands-on experimentation rather than textbooks. Deployment rewards applied judgment: knowing how to curate data, debug complex failure modes, and iterate toward reliability. That kind of judgment is not exclusive to people with graduate degrees.

Building in robotics is extremely hard. It involves hardware quirks and software failure modes you don’t encounter in pure software. But difficulty alone doesn’t justify insulation. SpaceX routinely gives early-career engineers responsibility for systems that land rockets. Every fast-moving technical field grows by absorbing outsiders who bring new tools, abstractions, and instincts with them. If robotics remains resistant to cross-pollination, it guarantees its own bottleneck. We should invest more in early-career talent and candidates from adjacent fields. Prioritize slope over intercept.

Deployment is the Forcing Function

Deployments cannot be treated as the final step that happens after “real innovation” is complete. They need to become a forcing function that shapes the next cycle of research. Real-world use exposes failure modes that research environments rarely capture, constraints around reliability, maintainability, operator workflow, integration with existing tooling, and cost. Those pressures should flow back into every layer of the stack: hardware design, data infrastructure, training and evaluation pipelines, control software, and the models themselves. Deployment is not the final stage of innovation. It is the mechanism that guides where innovation should go next and demands a different kind of talent to drive it.

Waiting for universal embodiments and universal intelligence is a convenient way to never deploy anything. The goal isn’t to start with the perfect robot. It’s to start with a robot that can deliver value today, even if narrowly. The minimum viable deployment may be constrained, heavily teleoperated, or limited to a small slice of a workflow, but it grounds progress in reality. Every deployed system forces clarity around reliability, cost, operator experience, maintenance, and customer needs, problems that cannot be solved in isolation. The scaffolding required for early deployments, safety systems, monitoring, logging, remote intervention, onboarding, and operations infrastructure is not wasted effort. It becomes the foundation that future autonomy scales on.

Deployment also expands the robustness manifold: the set of tasks and conditions in which robotic systems can create value. Research proves capability on narrow tasks, but deployment exposes where entire systems actually break: latency bottlenecks in software, weaknesses in data pipelines, hardware limitations, and workflow mismatches with real operators. Each deployment generates feedback that shapes how models are trained, how software is written, and how hardware should be designed.

For builders, this creates a compounding strategic opportunity. The companies closest to deployment are often the ones best positioned to ride the next wave of model capabilities. The playbook already exists in software. Harvey, the legal AI platform now valued at $8 billion, started out in 2022 on GPT-3, a model that hallucinated, needed manual review on every output, and couldn’t handle complex legal reasoning. Rather than waiting, they spent that window embedding themselves inside law firms: training models on each firm’s private work product, hiring ex-BigLaw attorneys to sell and build, and signing Allen & Overy before GPT-4 even existed. The model was mediocre. The relationships, domain knowledge, and firm-specific data were not. When better models arrived, Harvey absorbed them on top of a distribution moat that no competitor starting fresh could replicate. Their co-founder, Gabe Pereyra, captured it well: “Don’t build for the current capabilities of models today—build for where the models are going to be. Tackle more complex versions of problems so that when better versions come out, they aren’t solved as a side effect.“ Robotics companies solving operationalization, customer integration, and deployment infrastructure today are the ones that will compound on every model improvement that comes next.

Researchers who have carried the field this far deserve enormous credit, and they should absolutely keep pushing the frontier. The community requires a summative transformation, not an in-place one. We need to welcome a new wave of builders whose primary job is deployment, iteration, and integration to solve the operationalization problem. Deploying robots into real businesses is itself an exciting first-class engineering problem; it is distinct from research, and one that demands its own dedicated class of builders.

A New Age of Robotics

Robotics needs fewer roboticists. Not overall, but per capita.

It needs operators, designers, infrastructure engineers, people who sell, support, and scale systems, and young talent willing to own deployments end-to-end. It needs to be more open as a community and more ruthless about grounding itself in the real world. Creating value is the benchmark that matters in the long run.

Past waves of robotics pushed the frontier forward. This next one has the chance to turn that progress into real, durable value, but only if the field is willing to let more people in and treat deployment as the forcing function it actually is.

Robots don’t change the world until they leave the lab. It’s time to build.

If you’re building in this space, or seriously thinking about building at the application layer, we’d love to hear from you. The next wave of progress will come from people who are willing to iterate in the real world.

Thank you to Kyle Vedder, Advait Patel, Utkarsh Singhal, Vishnu Mano, Shiza Charania, Nicholas Wade, Connor Soohoo, and Tracy Livengood for reviewing drafts and discussion..

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Imagine you’re the CFO of your favorite CPG brand. You’re selling in Target, Walmart, Amazon, and the financials are looking good — but at the end of every month, the same retailers driving your revenue growth are paying you up to 20% less than what you believe you’re owed. You want to prove that you’re right but doing so requires an enormous amount of manual effort and time.

Welcome to the messy world of retail deductions.

Deductions are one of the highest-leverage problems in consumer goods. Historically the only solution was to throw people at the problem – teams of analysts or outsourced vendors sorting through hundreds of pages of shipping records and invoice line items, logging into dozens of retailer portals, manually reconciling records to dispute what they could. But even the best processes miss plenty of illegitimate deductions, evaporating revenue and profits.

Enter Glimpse. Glimpse uses AI to solve the deductions problem and bring modern tools to the back office of CPG brands. Glimpse’s system ingests deduction claims from every source — retailer portals, EDI, email, PDFs and consolidates them into a single source of truth. Glimpse identifies what’s worth disputing, and executes. Deductions are a hair on fire entry point for brands because resolving them can literally increase revenue and earnings by multiple percent. But interestingly this wedge product provides a data layer that Glimpse can use to solve a litany of other back office problems for brands.

The early results have been phenomenal. The business is growing vertically, serves over 200 brands, and has already helped billion dollar multinational firms discover millions of dollars in lost revenue.

Perhaps what’s most impressive about this company is the team. Glimpse’s founding story dates back to three best friend founders as engineering undergrads at Purdue. Akash, Anuj, and Kushal started with an entirely different problem and navigated a hard pivot post-YC to get to what Glimpse is today, at times even going without salaries, and that maniacal passion permeates everything they do.

We’re thrilled to lead Glimpse’s Series A. If you’re a CPG or retail brand, there is almost certainly money sitting in your deductions that you don’t know about — reach out to the team at Glimpse to find out how much. And if you’re a builder who wants to work on a real problem with a team that simply will not quit, Glimpse is hiring.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

This case study is designed for founders and builders to share lessons and insights from operating companies. See disclosures at the end of the content for more information.

As growth investors, we often say great companies start with their numbers. In Revolut’s case, they are required to publish annual financials as a UK company - and their numbers are outliers, to put it mildly:

• Revenue grew 46% to £4.5 billion

• Profit before tax grew 57% to £1.7 billion, realizing a 38% margin

• Retail customers grew 30%, having added 16 million in 2025 alone

• Revolut has Euro-wide penetration with no single country accounting for more than 25% of fee revenue

• Likewise, revenue is now distributed across 6 segments, with no single category responsible for more than 22%

• 11 different product lines exceeded £100 million revenue

• Return on equity (ROE) is a category-breaking 35% (despite over capitalization)

Revolut continues to grow quickly and efficiently—its rule of 75% (defined as revenue growth + net profit margin) is in the highest tier among modern and established financial institutions.

Perhaps more importantly, we believe Revolut still has ample room to run in its existing markets, on both customer growth and monetization. And that says nothing of all the new markets that are potentially ripe for the taking—Revolut just applied for its US charter and has truly global ambitions.

The point is: this is not your grandma’s neobank. Revolut has the potential to become one of the largest banks in the world. There is more ground to break to get there, but we believe the foundations are there.

Throat-clearing aside, let’s dig in.

Subscribe now

I. One of the fastest growing financial institutions in the world

Let’s start with the topline. Revolut is growing revenue at an incredible clip.

Together with NU, they stand in a league of their own relative to the rest of the consumer fintech industry (as demonstrated in the chart below). Since crossing USD $1B of revenue in 2022, Revolut has compounded at a remarkable 76% CAGR (or 70% in GBP) in the subsequent four years, making it one of the fastest growing companies after crossing $1B. That growth is particularly notable when you consider that Europe has a very mature consumer banking industry (unlike developing markets, where NU operates).

To put that in perspective, in 2022, Revolut had less (or about the same) revenue than any of Robinhood, Affirm, Sofi, Adyen, Wise or Chime. Now it has anywhere from 33% more revenue to nearly 3x more than any of those other well-known consumer fintechs.

II. Dissecting Revolut’s Growth Algorithm: A Six-Trick Pony (and Counting)

An important differentiator for Revolut is that it’s no one-trick-pony (any more). It has multiple revenue drivers firing, all at once.

Revolut started out by addressing an acute pain point for Europeans: FX fees. With Revolut, Europeans traveling in and out of the eurozone or sending money abroad no longer had to face payment delays or 5% fees charged by their banks.

What was once a single-product, geographically concentrated nice-to-have, has become a full-feature business and personal bank that now captures ~1 in every 3 new accounts opened in Europe (the main region where Revolut operates today):

In Europe, 1 in 5 working age adults now uses Revolut. Revolut’s euro-wide appeal reflects the company’s product velocity and execution, which has been a thing to behold.

Revolut has delivered a full-suite of both personal and business-banking features that are driving growth in very different markets across Europe. Importantly, Revolut’s product suite now increasingly attracts users within the eurozone that may not care about the original FX value prop at all. We’d call Revolut’s platform “feature complete,” except that Revolut keeps delivering new features, so that could be selling them short.

It’s also not just the number of new features and products that Revolut offers. It’s the quality of the execution. Customers love it. The company reported in 2024 that 65% of new customers came either organically or via a referral from an existing customer. And our survey research highlights Revolut customer NPS is more than double the industry average.1

Put it all together, and the net result is that users continued to compound at 30%, reaching 68M at the end of 2025.

To put that 68M users in perspective, JPM—the world’s largest bank outside of China—has ~85M consumer customers (~70M+ considered “digitally active”).

To be clear, JPM’s total AUM makes it orders of magnitude larger than Revolut, but from a sheer reach perspective, Revolut is no longer just a “challenger,”—it’s the real thing. Revolut has more users than Sofi, Robinhood, Dave, and Chime combined.

A full product suite has not only continued to attract more customers, but has created an increasingly diversified revenue profile:

The company publicly reports 6 main revenue streams:

• interest income;

• card payments;

• wealth;

• FX;

• subscriptions; and

• other income.

All six segments grew YoY, and no single segment comprises more than 22% of revenue.

The business is even more diversified than this would suggest, as every revenue stream may have multiple sub-products underneath (e.g. wealth encompasses both public equities and crypto). 11 different product lines exceeded £100 million revenue in 2025.

Importantly, 76% of the revenue comes from fees, up more than 4 percentage points vs 2024, while interest income accounted for just under 22%. That is the inverse of established banks that make 70%+ of their revenue from interest, and also one of the reasons why Revolut can earn premium ROEs (more on that below).2

Unsurprisingly, a diversified revenue base has translated into diversified ARPU growth, as well.

Since 2022, every reported revenue stream has grown and aggregate ARPU has grown ~65%, or a 18% CAGR.

The diversification point is important because it supports continued compounding and builds resiliency. In any given year there could be product lines that explode and ones that face headwinds (e.g. a decline in interest rates, as per last year). But, the net of it all can still drive strong ARPU growth, driven by attachment of new products and continued share of wallet gains in the core.

III. Top Tier Efficiency

Revolut has demonstrated rapid user growth, extreme product velocity, and diversified revenue, but we promised efficiency, as well.

In 2025, Revolut realized revenue growth of 46% and a net profit margin of 29%, resulting in a rule of X (growth + margin) of 75%. So much for Rule of 40!

The company’s combination of growth and efficiency puts it in rarified air—only a handful of companies have ever achieved Rule of 75% at $1B+ of revenue.

In fact, considering that Robinhood and Dave are expected to grow <30% next year per consensus estimates, Revolut may soon be standing alone at the top of the podium.

Efficiency is rooted in Revolut’s DNA. The combination of prioritizing self-developing its banking infrastructure, driving highly organic growth, and tightly managing costs, has resulted in realizing a net profit margin of 29%. Combined with its limited physical footprint, Revolut now has a meaningful cost advantage over incumbent banks that should compound as the company continues to scale.

To put a bow on it, the company is also demonstrating AI can drive even further operating leverage. Take customer service, for example:

In 2024, Revolut’s Assistant chatbot reduced resolution times by 80%. In 2025, those improvements continued with resolution times decreasing another 40%+ for retail and 50%+ for businesses—all while user NPS increased nearly 12 percentage points year over year. Revolut’s Assistant now resolves over 75% of customer queries.

All this efficiency results in Revolut delivering the highest ROE we have seen at scale in fintech (and it’s still continuing to improve). We’ve written in the past about the importance of ROE to bank valuations, and Revolut is a shining example of efficiency at scale.

Revolut’s 35% ROE is well-above other leading consumer fintechs, and ~3-4x higher than mature banks. Keep in mind that Revolut is ‘over-capitalized’ (i.e. its reported equity is higher than what would be needed under banking capitalization requirements), which means that its ‘true’ ROE may be even higher.

Growth doesn’t get much more capital efficient than that.

IV. Ample Room to Run: ARPU x Users

While Revolut’s results in 2025 are remarkable, we think there is still tremendous runway ahead. Referring back to the company’s core revenue growth algorithm (users x ARPU), both levers have a lot of remaining pull.

Many more users to acquire

Consider that the company reported having 68M users at the end of 2025. As noted above, that’s a lot, but it’s also <15% of the ~450-500M adult population in Europe (excluding Russia). That also excludes the growth opportunities in Australia + Singapore (existing markets), Mexico + Brazil (recent entries), the US (recently applied for bank charter), and more geographies to come.

The point is that Revolut has a lot more potential users to capture.

Plus, the current user-mix already demonstrates that the future is unlikely to look like the present. Unsurprisingly, Revolut skews young(er) and digitally savvy—we think these customers are representative of what the majority of the population will seek eventually.

As Revolut continues to capture a high share of first-time banking customers (and convince older demographics that banking can indeed be a pleasant user experience), their market share should continue to grow.

Importantly, ~25% of Revolut’s users under 35 treat Revolut as their primary account per our survey. We’ll get into that more in a moment, but leaving everything else aside, simply the maturation of that cohort ought to have a profound effect on the future of banking market share in Europe.

ARPU has a lot more potential expansion

The other vector of growth, ARPU, has even more room to run.

Share of wallet in financial services typically shifts over decades, not years. Revolut continues to earn the trust of its customers, as primary account users (per the company’s reporting) grew 45%, outpacing the overall user growth of 30%.

The steep growth of primary account users matters, because when it comes to ARPU, “primary” users are the big prize:

• Our research points to incumbent banking institutions (with seasoned customer relationships) can push their “primary account” share to 60%+.

• Likewise, Revolut primary account users self reported spending and saving ~2x as much money with their primary account as they do with any other open account—and spending tends to increase as customers age;

Simply put, more (and maturing) primary users can translate into higher ARPU, and if incumbent banks are any guide, the ceiling for Revolut’s growing “primary share” is quite high.

Another aspect of the growing “primary” relationship is the still mostly-untapped lending revenue opportunity for Revolut:

• As noted above, currently Revolut derives 76% of its revenue from fees, in contrast to the ~30% typical of mature banking institutions;

• At the end of 2025, Revolut’s Loan-to-Deposit ratio (LDR) stood at just ~6%, compared to ~70-90%+ for established banks (or ~4% if calculated using total customer balances). Lending balances grew ~2x in 2025 and can continue compounding for many years to come.

To be sure, disciplined lending growth takes time. But if the incumbent ceiling is a guide, Revolut will have ample opportunity to expand ARPU much further, by leveraging its balance sheet and providing better loan products to its customers. By way of comparison, simple napkin math for Barclay’s UK consumer and business banking line, implies an ARPU of ~£435 or ~6x what Revolut generates today.

Just to illustrate the point, take a look at where Revolut currently stands, by reach (i.e. penetration) and depth (i.e. primary bank share):

Revolut has plenty of runway to keep pushing up-and-to-the-right (or, in Ireland’s case, mostly up), both by expanding its user base and deepening those relationships to “primary.” The latter of which should continue to occur organically, as the younger cohorts mature.

V. Closing Thoughts: Not Just A Challenger Anymore

Revolut’s 2025 numbers matter not only because they are impressive, but because they paint a cohesive picture of a financial institution, and not just a “challenger” bank.

Customer growth remains exceptional, monetization is broadening, primary account adoption is rising, and profitability is strengthening even as the company continues to invest and rapidly scale. That combination is rare in financial services (and in any industry).

There is still execution ahead—particularly in lending, regulation, and expansion into new markets—but reading this annual report, we believe the question feels less like ‘whether Revolut can become a scaled banking platform,’ and more like ‘how large can that platform become?’

The company’s stated long-term goal is “100 million daily active customers in 100 countries.” The march is well-underway.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

To software CEOs, founders, boards, and the investor community: the comfortable middle is over.

Public markets have already repriced the sector, and for good reason. They are telling us software terminal value is not what it used to be. I do not know what moves every stock in the next quarter, but over the medium and long term I think there are only two credible paths to durable equity value creation.

• Path one: accelerate revenue growth by 10+ percentage points, year over year, through genuinely new AI-native products over the course of the next 12-18 months.

• Path two: rebuild the company to 40%+ true operating margins (ideally 50%), including stock-based compensation.

Strictly speaking, these initiatives are not mutually exclusive. But I suspect the 12-18 month plan needs to be one or the other. By the end of next year, everything between those paths of high growth and high profits will look like no-man’s land: growth pressure, persistent dilution, and multiple compression. CEOs today need clear initiatives to drive toward one of these as an output.

The Adjusted Jig Is Up

Public software has already lived through the first half of the transition. Growth rolled over and valuations compressed. But in most cases, true profitability still has not arrived.

Yes, free cash flow improved; GAAP margins improved some. Yet once you treat stock comp as a real expense instead of a permanent carve-out, much of the sector still sits in the difficult middle: too slow to deserve a premium growth multiple, too diluted to deserve a fortress multiple.

If topline growth is slowing down, we should be seeing more operating leverage, and again, while we’ve seen some, we haven’t seen enough.

The reality is that the time has come for bold management. And no, the “8% or 10% layoff” headline no longer counts. That is the weak form. The weak form trims the edge of the org chart and leaves most of the machine intact. The strong form is a redesign of the machine.

In the next 12 months, I expect we will get much more of the strong form. You have two options for how to do this, and the difference is in how you want to rebuild your company.

Subscribe now

Path one: accelerate growth off new AI products

Accelerating growth with new AI products does not mean bolting on chatbots or copilot interfaces, attached to the old SKU list.

It means new products that can move the company’s total growth rate by 10 points within 12 months. And, just as importantly, it means you need to speedrun rebuilding your company - including your executive team - so that if you do find product market fit, you will actually capitalize on the opportunity.

The first thing you need to do is identify which people are going to be your leaders that help you pull this off. This is going to be a 12 month death march and you need to find out who is willing to go through the pain with you. There’s good news, though: somewhere in your org, there are ~five people who are going to deliver you 100x the amount of value you ever thought possible. Your first job is to figure out who these five people are (no matter how junior they are on paper!), explain the urgency of the situation with them, and give them the career opportunity of a lifetime to rebuild the company with you.

What do you do with these people?

First, you put them in charge of unglamorous, but critical information-gathering projects:

• process-capture sprints around every high-value workflow;

• harvest SOPs, tickets, transcripts, requirements docs, policies, CRM notes, support logs, event data, and approval paths.

Create a living context layer, not a pile of static PDFs. Treat documentation as product infrastructure. Instrument evals around accuracy, exception handling, latency, and cost. Get these five people on that task, each with their own scope of responsibility, immediately.

Over the next month, watch your VPs like a hawk to see who is on the bus with that team, and who is off the bus.

This will tell you everything you need to know about which half of your execs to keep for your imminent company rebuild, and which half needs to part ways. At the end of the month, have those tough conversations with the VPs and directors that need to go. Replace them with the crack team who just completed your information gathering sprint, and other AI-native up-and-comers in the company who’ve earned their shot. Now you have a refreshed and energized exec team, ready for the fight.

Meanwhile, you’re going to put 50% of R&D on net-new AI products.

• Use four-person pods; collapse design, product and engineering into one working unit, start writing code on day one, and cap headcount, not compute. Reduce communication overhead to as close to zero as possible.

• Make sure all of your best product managers are as customer-facing as possible. They can’t waste a minute. Their job is pure product discovery; make sure they’re unblocked by any legacy anything.

• Your best engineers, meanwhile, are going to stay in the central engineering org, reporting directly to the CTO. Their job is to make sure that the company’s core engineering architecture can evolve as quickly as your vanguard PMs.

Mileage may vary by company here, but my advice is don’t put all your best engineering talent at the edges. It’s tempting to do so, but this will Balkanize your tech stack and create years of technical and organizational debt that will smother any early promising progress. Furthermore, with AI, you don’t need your best engineers to be on net-new product discovery; you just need people who ship and learn quickly. The best engineers should stay focused on the company’s overall tech architecture, but prioritizing the new stuff ruthlessly.

As a part of this sprint, your business needs to get really good at escalating contentious decisions to unblock progress. You will not pull off this transformation and successfully build new AI-native businesses in 12 months without making hard choices, every single week. So get good at this process, and make sure your newly reformed executive team is dedicating a LARGE percentage of their time (at least one whole day a week) exclusively to unblocking designers, PMs and engineers like the company depends on it.

In the course of unblocking your team, you’re going to figure out exactly what your new business model is. It will need to make money off of tokens / per use, rather than the old seat-based model. You do have some time; seat-based pricing isn’t going to go away overnight. But you need to take the challenge seriously: you can’t hand-wave a new pricing model and product interface. If an agent can’t consume and pay for your product autonomously, you probably aren’t there yet.

The budget for new spend is there. You can do this.

But remember that your customers’ first and most obvious source of AI savings is labor efficiency, which means seats are where they will look to take cost out. The new growth, by contrast, will increasingly sit in tokens, consumption, automations, outcomes, and machine-driven workflows.

If you are not in the token path, you are not standing in the fastest-growing part of the budget.

Not all companies are in a position to do this. You may assess your options and not see any credible line of sight to winning via Path 1. But if you do, and if you make it through this twelve month sprint, you will emerge as a focused, accelerating company with a new leadership team, and a “refounding moment” that your team will rally and reenergize from for years to come.

Path two: rebuild for 40%+ true margins

Software companies got very good at talking about free cash flow margins over the last decade. But if we are serious, we should stop excluding stock comp and pretending dilution is not an expense borne by owners. For companies that are not about to reaccelerate growth, I think the right target is 40% or even 50%+ true operating margins, including SBC, within 12-24 months.

Getting to 40%+ profitability requires more than 10% or 20% RIFs. It means flattening management layers, standardizing implementation, minimizing bespoke services, killing committees, raising price where you own the workflow or the switching cost, moving long-tail customers to higher floor pricing or letting them churn, and counting every share issued as a transfer from owners to employees.

AI should change the shape of the company. The cost structure should change with it.

This will take a similar level of effort to the first path. And even though your goals are different, you’re still going to aim for an AI-native company, whose engineers are maximally productive and efficient, in 12 months. From day one, you need to figure out what a smaller, but even more motivated and productive workforce looks like in your organization twelve months from now.

The first thing you’re going to do, counterintuitively, is radically increase the amount of budget you allocate for token spend per engineer. If your engineers are not spending real money on tokens, they are probably not pushing hard enough. A thousand dollars per engineer per month is not excessive; it is close to table stakes.

A useful working premise is that the ceiling on individual engineer output is moving much faster than most companies are organized to exploit. Some of the best operators already describe top engineers seeing order-of-magnitude productivity gains and managing 20 to 30 agents simultaneously. Whether 20x proves to be the edge case or simply the frontier, the organizational implication is the same: companies built for ten-person committees will lose speed to companies built for four-person strike teams.

Meanwhile, you need to prepare for a significant RIF - you already knew this.

You cannot just prune the leaves at the edges of the company: if you lay off a large percentage of the ICs in a company but leave the director and VP corps intact, you’re worse off than where you started. To be clear, it’s different from path 1; you’re not trying to build a “new” business. But you are still “refounding” the company along a new set of values around performance and shareholder mindset, so make sure you go on this journey with the right bench of leaders.

Another very important thing is that teams should be honest about which old moats are weakening.

• Data alone is usually not enough.

• Integrations are getting easier to reproduce.

• Workflow and UI advantages matter less when agents can move across systems more easily. Migration is getting easier.

Competitors will increasingly come after each other’s modules, not just the edges. That means core-business price pressure is coming, so prioritize your strengths that will help you maintain pricing power and customer retention.

It Can Be Done: The Lesson of Broadcom

There is one public-market case study of the strong form, pre-AI: Avago/Broadcom under Hock Tan. It is a harsh model. It is not a cultural blueprint for every founder. But it is a reminder that radical cost discipline, product simplification, and price realization are possible. The strong form exists.

Path two may sound defeatist, but not every software company has a right to path one. If that right doesn’t exit, path two is the only path to value creation.

The pivotal question

Founders should put one question on page one of every board deck: which path are we on?

+10 points of revenue growth from net-new AI products? Or 40%+ true operating margins including SBC?

Investors should ask the same question more forcefully than they do now.

Where is the AI product engine that can move the curve? Where is the re-architecture of R&D around small, token-rich, customer-near teams? Where is the plan to build the dual human/agent interaction layer? Where is the explicit roadmap to 40-50%+ true margins? Where is the plan to bring dilution down as a share of revenue?

If the answer is some version of “a little of both” or “we are evaluating options,” I would expect the market to keep applying pressure.

Founders: you need to pick a path, and you need to decide quickly who on your team you want on the bus with you. You have the opportunity for a new founding moment for your company, your new team and your investors. Grow 10 or earn 40. Build the next product wave or build the cash machine. No middle lane. Good luck!

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

1. It’s a tough job market for young people, knowledge-worker or otherwise

It’s common knowledge at this point that it’s a tough job market out there for young people.

The obvious culprit is AI, which has already vibe-coded all the entry-level white collar jobs into oblivion:

The unemployment rate for young, college-educated workers has deteriorated more so than the field, and certainly more so than young, non-college workers. (Note that this is indexed change, and not absolute levels.)

If it’s the college-educated that are uniquely struggling, then it’s case closed: AI did it. No one needs analysts anymore. White-collar work we hardly knew ye’. This is the end for knowledge work. Blackpills and UBI from here on out.

Well, not so fast. The reality is that it’s not just “knowledge work” roles that are hard to find, but you have to cut the data a little differently to see it more clearly.

Take a look at the employment rates for both college and non-college young folks:

Employment rates for young workers, whether college educated or not, have both been in roughly equal decline for the past two years.

When you look at it this way, young knowledge workers specifically don’t appear to be uniquely challenged. It’s just hard out for all young people. There just isn’t that much hiring going on.

Now, the employment rate measures the share of young people in the workforce, but it differs from the unemployment rate in two key respects: (a) it includes people who aren’t actively looking for work (who would otherwise be excluded from the unemployment rate), but (b) still excludes people who are in school (and therefore have a good reason for not seeking employment).

In other words, the employment rate is a more complete measure of employment, especially for young workers. While it might make sense to exclude non-searchers from the headline unemployment rate (because there are many reasons a working age adult might not be looking for work, some good, some bad), it makes less sense for young folks, where college is really the only good excuse.

And, back to the story at hand, if it’s the case that both college and non-college young people are struggling to find gainful employment, then it’s harder to blame AI specifically.

Generally speaking, entry-level non-college work is not the repast of easily-automated workflows—it’s retail, food service, hospitality, and blue-collarish roles, all of which require fully-functional hands and feet, that AI does not yet possess. The point being that it’s not just “knowledge work” that’s hard for young folks to find—it’s all the work.

So, if it’s not AI that’s making the job market hard for young people, then what is it?

Well, it’s hard to say for sure, and it may not be one thing, but first some context:

Going back to 1976, the share of 20-somethings in the workforce has declined ~12%.

Young folks, in general, are an increasingly smaller share of the workforce, and that’s been true for a while. A good chunk of the change is driven by the rising college attendance, but other structural shifts play a role too, e.g. the relative decline of blue-collar/manufacturing jobs that might have otherwise been a compelling alternative to college.

Again though, just because the workforce has been skewing older since the 70s, that doesn’t explain why the employment rate for young people has declined very recently—especially because the employment rate excludes people currently in school. It may be that college has become less-appealing recently to young people, and the workforce hasn’t adapted to that shift, just yet—that’s certainly possible, but some other data suggests that perhaps a different dynamic is at play.

One thing that’s changed quite recently is that the age of the workforce has increased fairly dramatically since 2025

The workforce had been getting steadily older by ~1 year every 5, but that pace accelerated ~5x in 2025, when the average age jumped a full year, in just a year’s time.

Some of that aging is explained by the lack of entry-level hiring, but it turns out that another part of what’s happening is that people are staying in their jobs (and in the workforce more generally) longer than they used to:

“Within-occupation” aging explains almost the entirety of the recently older workforce, at least according to Revelio’s data.

If job-switchers were getting older too, then workforce aging might reflect some demand-shift to more experienced workers (which is true in some narrow cases, but not more generally). The fact that nearly all of the aging is “within-occupation,” indicates that people are simply less-inclined to retire. Along similar lines, other data shows that “working lives” have extended by ~1.5 years, and that retirees are increasingly coming off the bench to reenter the workforce.

Putting it all together, and while we can’t say for sure, it’s possible that one of the reasons hiring is rough for all kinds of young people is that older folks are simply working longer than before. While it’s generally a mistake to think of the labor market as a zero-sum game, if the off-ramp gets extended, it stands to reason that the on-ramp may get backed-up, as well.

So, don’t blame AI for the lack of entry-level jobs—perhaps blame old folks for working too hard, instead.

Subscribe now

2. Learn to weld

Staying on the jobs theme for a moment, but now we’ve got some good news.

The change is subtle, but manufacturing jobs appear to be on the rise:

Manufacturing job openings increased slightly in the beginning of 2026, in contrast to white-collar roles, which continued their 3-year decline.

A rise in manufacturing roles should be welcome news for entry-level hires, which have gravitated towards blue-collar work in the past. What’s causing the manufacturing job boomlet? Well, it’s probably got something to do with AI-the-job-maker. Or rather, the infrastructure build-out, more generally.

We already know that data centers are driving hiring for construction workers, and specialty construction, in particular:

Non-residential specialty contractors and civil engineers—the kinds of people involved in industrial construction—have been in hot demand, in contrast to residential construction hiring, which has been in decline for over a year.

Construction and manufacturing are not the same, but you can’t build data centers (or upgrade the grid) without all kinds of heavy equipment, parts and machines, i.e. the sorts of things that manufacturers manufacture. It makes a lot of sense that surging AI buildouts might also increase demand for manufacturing.

By some estimates, the (re)industrial renaissance is going to need so many new manufacturing, construction, and operational workers that we’re not going to have enough to go around:

Goldman Sachs estimates an additional ~500K jobs are needed to achieve the buildout, leaving a ~78K “labor gap,” (even assuming that 100% of apprentice roles gets allocated to the sector).

The shortage of skilled tradespeople has been a thing for a while now, both as demand for “blue-collar” expertise increases, but also as the existing workforce gets older. Add a few hundred thousand new AI-infra jobs to the pipeline, one can only imagine that the shortage will get worse. Certainly that’s what Goldman Sachs thinks, at least.

To all the struggling young folks out there, we’re not saying the solution is “learn to weld,” but we’re not not saying it, either.

3. Rise of the AI native small business?

We’ve all heard of the one-person billion dollar company coming to a vibecode terminal near you. With agents orchestrating agents orchestrating agents, a lone determined founder now has the operating leverage to build and operate a massive company without a single (other) employee.

It hasn’t happened yet, of course, but the possibility isn’t so far-fetched.

What has happened, however, is that lots of new small businesses are forming, and while they’re not hiring much, they are spending a lot on technology:

• New business applications are elevated, but “high propensity [to hire]” business applications are in decline.

• At the same time, SMB payroll expenses are flat-to-down, while tech-service spending keeps trending up.

So, the data shows lots of new business formation, but it’s of the “non-hiring” type, and while small businesses aren’t adding much payroll, they are spending a lot more on tech.

Is this the early innings of AI-native solopreneuring? Who knows, but we can’t wait to find out.

4. Who do you think you are, Mr. Big Short?

Every now and again, the doomers wonder “but what will happen to all the old GPUs?!”

It’s a fair question, for sure. The Big Shorter himself (aka Michael Burry) said the hardware should depreciate in 2-3 years (and not the 5+ years that hyperscalers are currently modeling). If new GPUs come out so quickly, then they must make the old GPUs obsolete, just as quickly.

Checkmate, hyperscalers. The bubble is going to burst, any moment now.

Maybe one day, but not today, Mr. Big Short, not today:

According to data tracked by 3Fourteen Research, since late 2025, “on demand” GPU-availability has gotten scarcer and scarcer—for the latest B200s, it’s basically non-existent. If you need on-demand GPUs, your best bet is the GH200, which is still only available less than 40% of the time.

Perhaps next month will tell a different story, but for now, if older GPUs are going obsolete, someone should probably tell whomever is burning all that compute that they’re wasting their time.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Editor’s note: Today’s piece is from Turner Caldwell, Co-founder and CEO Mariana Minerals. You can view the original here. It’s a fantastic primer on where copper comes from, why it matters, and how we’re going to produce it in the near future. -AD

Located in southeastern Utah, Mariana Copper One is an operating copper mine and refinery currently producing high-purity copper cathode for the U.S. market. The site, which we acquired in late 2025, has been operating for over 15 years and includes an existing open pit mine, a heap bioleaching operation, a hydrometallurgical refining circuit utilizing solvent extraction and electrowinning, and significant expansion potential. Over the past 12 months, Mariana partnered with the existing site team - previously Lisbon Valley Mining Company - to develop and deploy advanced software and autonomous systems that will overcome past operational challenges and expand copper output.

Our path forward is clear: 

1. Deploy PlantOS at scale to maximize copper recovery and reduce copper refining costs throughout heap leaching, solvent extraction, and electrowinning.

2. Restart mining operations with autonomous equipment and orchestration via MineOS.

3. Integrate copper scrap processing into the refining circuit, leveraging PlantOS to manage feedstock variability and to put a meaningful dent in US copper scrap exports.

4. Scale combined output at the site to 50,000 metric tonnes per year from both geologic and scrap feedstocks (leveraging CapitalProjectOS to accelerate capital project delivery).

This is Physical AI at scale.

Why copper?

Copper is the circulatory system of the modern economy. Nicknamed “Dr. Copper” for its ability to diagnose economic health, copper is central to power generation, supply and conversion, AI datacenter build-out, residential, commercial and industrial infrastructure, electric vehicles, grid-scale energy storage, and advanced defense systems. Anywhere electricity is generated or consumed, copper is at work.

As our global energy, transport, infrastructure, and defense systems modernize, we need copper. A lot of it. By 2050, we must double global copper mining and refining capacity - this means mining and refining more copper in the next 25 years than in all human history combined.

Subscribe now

How is copper produced today?

In thinking about how we will produce enough copper to meet accelerating demand, it’s important to first understand how copper is mined and refined today. We’ll start with the high-level and provide additional details around Mariana’s mining and refining pathway in subsequent posts.

Primary copper supply (mined copper) typically comes from 2 types of ores: copper sulfides and copper oxides. These ores are mined via open pit or underground mining, then processed through 2 common pathways - pyrometallurgy (high-temperature smelting) and hydrometallurgy (chemical leaching and solvent extraction) - to produce high-purity copper cathode for downstream manufacturing.

Building a robust, sustainable copper business requires a deep understanding of both the mineralogy of various ore types and of the metallurgical processes involved in concentration and refining. These are inextricably linked - processing and refining pathways cannot be developed independently of the ores being processed. This is ultimately what leads to the bespoke nature of the mining industry: what works at one site with one type of rock does not work at another. This is a major driver of the complexity involved in expanding copper production capacity globally.

Where is copper produced today?

Copper production and trade are highly concentrated. A small group of mining countries - Chile, Peru, Indonesia, Australia and Brazil - account for the majority of global copper ore exports, reflecting the concentration of major deposits in the Andes and Asia-Pacific regions. On the demand side, China dominates as the world’s largest importer, with Japan and South Korea also serving as major buyers.

The copper trade has fundamentally shifted over the past decade. Mineral-rich countries in South America export ores and concentrates, while Asian economies import them for smelting and processing into refined copper for electronics, construction, and renewable energy technologies. China is increasingly dominant as the destination for copper concentrates due to its central role in global refining capacity and manufacturing supply chains. The United States and other developed Western countries need increasing amounts of refined copper due to electrification and remain largely importers of the refined units with limited local refining capacity.

While the U.S. has some large copper operations, in 2024 we still imported ~50% of our total refined copper needs. This is increasing: over the past 25+ years, our output of refined copper has dropped by ~50%. The chart below, which resembles one that compares U.S. electricity production to electricity production in China, underlines decades of underinvestment, declining domestic capacity, and growing reliance on foreign supply.

Satisfying domestic copper demand requires both mining and processing capacity. Even when copper is mined domestically or available for recovery through scrap, huge volumes are exported for smelting or refining abroad and then re-imported as refined copper. This processing bottleneck, combined with rising demand from electrification, renewable energy systems, and data centers, has created a structural deficit that cannot be solved by mining alone. The U.S. needs expanded domestic recycling, refining, and smelting capacity to close the gap.

The issue isn’t isolated to the U.S. Over the past 25 years, major Western copper producers globally have struggled just to maintain output, while Chinese companies have scaled globally to levels that today exceed the combined production capacity of their Western counterparts.

It is getting harder to scale copper output, not easier.

Miners extract their most economic resources first, targeting the highest-grade ore for extraction. As copper mines age, the grade of the remaining resource reduces. More than half of operating copper mines globally are over 20 years old, forcing operators to mine more material to produce the same amount of copper each year. This presents a structural issue that must be addressed with new mines and new mining and refining technologies to reduce costs.

The number of large copper discoveries has declined sharply since the 1990s, and the deposits we do find are deeper and harder to develop. The industry’s reserve growth mostly comes from expanding existing deposits rather than finding new ones.

The remaining undeveloped deposits are often located far below the surface, requiring advanced underground mining methods and massive capital investments. The Resolution Copper project in Arizona - one of the largest undeveloped copper deposits in North America - targets an orebody located more than 1,600 meters underground. Some of the project’s mine shafts extend beyond 2,100 meters (deeper than 5 Empire State Buildings stacked on top of each other!), requiring ultra-deep block-cave mining to extract the resource.

These trends mean new copper supply increasingly depends on deeper, more complex projects that take longer to permit, finance, and construct. The pace at which new supply can come online is slowing, while demand continues to accelerate (note – autonomy is a solution for this, too, but we’ll save that for later).

As mines get deeper and ore mineralogy becomes more challenging, cost and complexity increase non-linearly. Capital requirements balloon. Permitting timelines stretch. Operating expenses and optimization challenges compound. Meanwhile, the pool of skilled labor in the West shrinks, particularly among those able to engineer, execute, ramp production, and optimize metallurgical processing circuits, increasing the delay of bringing new copper production online (as we called out in our Series A announcement).

The time to invest in and scale copper production is now.

The global copper supply system is under mounting pressure. Existing operations are harder to sustain. New sources of supply are more difficult, expensive, and time-consuming to bring online. Meanwhile, demand accelerates from electrification, renewable energy, grid expansion, and industrial growth.

The combination of geological limits, technical complexity, and long development timelines means supply will struggle to keep pace. Market forecasts anticipate a substantial structural shortfall - some projections suggest a deficit approaching 10 million metric tons if new capacity fails to materialize quickly enough.

We’re faced with a bleak outlook: demand is accelerating while supply struggles to keep pace. The gap is widening. Closing this gap requires a fundamentally different approach to mining and refining - one built on autonomy, software, and operational intelligence.

Today, autonomous vehicles navigate cities. We launch and land rockets autonomously. Autonomous weapons will soon define the battlefield, and humanoid robots will soon deliver services in factories and at home. Mariana Copper One is the first mine in the world that will demonstrate what end-to-end autonomy will deliver in mining.

Mariana Copper One: the only autonomy-first copper mine and refinery.

Mining and refining operations require thousands of cross-functional decisions daily. Today those decisions are informed by fragmented data management systems and driven by human intuition. This leads to significant shift-to-shift performance variability and exposes the major challenge we face today: we don’t have enough people that know how to operate and optimize these complex systems.

MarianaOS breaks this paradigm.

MarianaOS is a unified software platform that connects process development, capital project execution, geologic modeling, mine planning, simulation, fleet orchestration, plant control, maintenance, and supply chain management to automate operational optimization across the value chain. This integration gives operating teams a single, real-time view of distributed systems - from fleet telemetry and process plant performance to market signals and supply chain management. Automated data pipelines ensure information flows reliably and continuously, eliminating manual reporting delays and enabling faster operational responses.

On top of this integrated data layer, machine learning and reinforcement learning optimize complex operational tradeoffs - drill patterns, haulage routing, blending strategies, plant throughput, reagent, and energy use - while continuously learning from outcomes. The result is a self-improving system that autonomously orchestrates decisions across the mine, refinery, and logistics network, increasing productivity and reducing costs.

MarianaOS uses simulation, predictive maintenance, and reinforcement learning to automate drill and blast operations, truck and shovel orchestration, and process plant control based on real-time data; it handles continuous optimization and data integration while engineers, geologists, operators, and planners validate models, monitor the operation, and execute essential, hard to automate tasks. The expertise and creativity of great people remain essential, but autonomy and agentic workflows enable us to mine and refine more while the skilled labor pool shrinks.

The Impact of Autonomy

MarianaOS will significantly improve mining productivity, efficiency, safety, and environmental performance at Copper One. This is increasingly critical as over 200,000 people in the current US mining workforce will retire this decade, while fewer than 400 mining engineering students graduate annually in the US. The data below shows actual results from the Garpenberg Mine in Sweden, which implemented automation and remote control over a 13-year period. The operator realized step-change improvements in productivity, equipment utilization, diesel consumption, and workforce injuries.

There is a massive amount of meat on the bone in optimizing mining operations, and while some mines globally have introduced remote controlled and semi-autonomous operations (like the example above), full implementation requires custom development with third parties and high latency interfaces between site teams and technology partners, can take over a decade, and still relies heavily on human-in-the-loop control. MineOS compresses this implementation timeline, accelerating the deployment of autonomous mining and support equipment, automating and integrating data pipelines between distributed systems, and driving human-in-the-loop operations to near zero.

Why Recycling Matters

Recycling copper scrap represents a significant opportunity to accelerate U.S. copper supply independence. Copper can be recycled repeatedly without losing its properties, making scrap an important secondary source of supply. Recycling already contributes hundreds of thousands of tons of copper to US supply each year, but a large share of domestically generated scrap is exported rather than processed at home.

The United States exported nearly ~950,000 metric tons of copper scrap in 2024, making it the world’s largest copper scrap exporter, with more than 40% going to China. The scrap exported each year, if processed domestically, would meaningfully reduce US reliance on imported copper and strengthen the resilience of the domestic supply chain.

At Copper One, we’re integrating copper scrap processing directly into our existing refining circuit. Our solvent extraction-electrowinning system is well-suited to process both primary ore from mining operations and secondary copper from scrap feedstocks. By tooling the refinery to handle both sources, we can put a very real dent in the copper scrap exported from the US annually, capturing domestic secondary resources that would otherwise be shipped overseas for processing and subsequent import.

This dual-feedstock approach aligns with a core thesis at Mariana that we’ve had from the beginning - modern refineries must be designed to process both geologic and recycled feedstocks. As we scale Copper One to 50,000 metric tonnes of refined copper output per year, scrap processing will be a meaningful contributor to our output and a proof point that recycling can be done economically in the United States. Enough excuses. Enough stasis.

We’re not just talking, we’re taking action.

The past decade has been filled with promises to rebuild US mining capacity - promises that rarely materialize into meaningful production. We need to break out of the pattern of never-ending feasibility studies and multi-billion-dollar first-of-a-kind projects that take decades to build.

Copper One is an operating mine and refinery. We’re not starting from scratch. We’re deploying MarianaOS at a site that has been producing copper for over 15 years, demonstrating that autonomy can unlock step-change improvements in safety, efficiency, and output.

Our plan at Copper One: deploy autonomous control of heap bioleaching, solvent extraction, and electrowinning to maximize copper recovery; restart mining operations with autonomous orchestration; integrate copper scrap processing to capture domestic secondary sources; and execute a major expansion to ramp output to 50,000 metric tonnes per year.

Mariana’s mission is to build the safest, most environmentally responsible, and most efficient minerals operations on earth. Copper One is where we prove that software-first mining can deliver on that promise - at scale. In the coming months, we’ll share more on our technology deployment and progress toward autonomous operations.

And the best part: we’re just getting started.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Reinforcement learning is becoming both the bottleneck and the unlock for the next generation of AI systems.

Over the past few years, we’ve seen remarkable progress in model capabilities driven by scale: more data, more compute, and better architectures. But as models begin to interact with the world, by using tools, navigating interfaces, and writing and executing code, the limiting factor is no longer just pretraining data. It is the environments, and their quality, in which these models can learn, practice, and be evaluated at scale.

Reinforcement learning (RL) environments are already becoming the next critical layer of the AI stack: a shift from static, human-annotated datasets to dynamic, engineered systems that generate high-quality training signals at scale. This transforms data from a labor problem into an engineering and research problem, and increasingly, a compute problem.

That’s why we’re excited to announce that we are leading the Series A in Deeptune, a company building leading RL environments for computer-use and code.

Deeptune is tackling a problem that sits at the heart of frontier AI development. As models move beyond text prediction into real-world task execution, they need structured environments where they can learn to fully control computers and perform knowledge work tasks. These environments must be realistic, measurable, and adaptable to rapidly evolving model capabilities.

This is extraordinarily difficult to get right.

Subscribe now

Designing high-quality RL environments requires a deep understanding of both cutting-edge AI research and the practical needs of top labs. Tasks need to be representative of real-world complexity. Evaluation needs to be precise and reproducible. And the entire system needs to evolve quickly as models improve.

Deeptune stands out because they’ve built exactly this.

The company has been working closely with leading AI labs, developing environments for computer use that are already showing up in benchmark improvements like OSWorld (a computer use benchmark), Terminal-Bench (command-line interface benchmark) and beyond. As models trained in these environments become more capable, we’re starting to see meaningful gains in how effectively they can fully operate computers (via a command-line or a Desktop interface), an early signal of what’s to come.

While there is still significant room for improvement in computer use benchmarks, SOTA models have made rapid progress on computer use over the past year. Opus 4.6 scores 72.7% on OSWorld, surpassing the human baseline of 72.4%, and GPT-5.4 reaches 75%.

We got a glimpse of this future with systems like OpenClaw: models that can meaningfully interact with software, tool use, computer use, and complete multi-step tasks. These are early signals, but they point to a world where models are capable of completing tasks end-to-end entirely on a computer. RL environments are what make that transition possible.

At the center of Deeptune is Tim Lupo, the founder and CEO. Tim is an exceptional founder with a rare combination of technical depth and product intuition. He works closely with top researchers across leading AI labs and has a deep understanding of what they actually need. This shows up in Deeptune’s ability to ship high-quality environments, tasks, and evaluation frameworks at an unprecedented pace.

Building in this space requires tight feedback loops with frontier research, and Tim has built those relationships from day one.

He has also assembled an outstanding team. Deeptune brings together researchers and engineers from leading AI companies and top labs, combining deep expertise in machine learning, systems, and applied research. It is a group uniquely suited to build the infrastructure layer that modern RL demands.

What makes Deeptune particularly compelling is that they are not just building tools, they are helping define a new paradigm for how AI models are trained.If the last decade of AI progress was driven by better datasets, the next decade will be mostly driven by better environments.

We believe Deeptune is at the forefront of this shift.

We’re thrilled to partner with Tim and the Deeptune team as they build this critical layer of the AI stack.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

In 1750 BC, a merchant named Nanni sat down in the ancient city of Ur and dictated the world’s first known customer complaint. Inscribed in cuneiform on a clay tablet the size of a human palm—a tablet that’s now held in the British Museum—Nanni unleashed his fury on a merchant named Ea-nasir, who had promised fine quality copper ingots and instead delivered “ingots which were not good.” “What do you take me for,” Nanni wrote, “that you treat somebody like me with such contempt?”

Nearly four thousand years later, Nanni’s feeling is universal. Everyone reading this has been Nanni at one point or another. You, reader, have been on hold for 45 minutes waiting to talk to a human who can solve your problem. You’ve been transferred to another department. You’ve screamed “representative” into an automated phone tree. You’ve rage-typed into an unintelligent chatbot that isn’t programmed to know what you’re angry about. The medium might have changed, from clay tablets to phone trees to ticket queues, but the emotion is ancient and unaltered: what do you take me for, that you treat somebody like me with such contempt?

And here’s the strange thing—as businesses have gotten better at everything else, customer experience has gotten worse. That’s one of the great ironies of the internet age. The internet helped commerce scale infinitely: one-click purchasing, next-day delivery, oceans of choices for practically any consumer good. But it made the experience of being a customer a lot worse. The more customers a business services, the less attention any individual customer gets: because while logistics can scale exponentially, devoted customer attention—real attention, to make sure you’re happy with the experience—can only scale linearly. That is to say, with the number of humans that the business employs.

For any business that sells goods online, the inevitable fact is that customer service is a cost center to be minimized, not a relationship to be deepened. Each customer becomes a ticket number, a case ID, a position in an endless queue. This isn’t because businesses today care less than businesses did 100 years ago or 4,000 years ago. It’s simply because devoted attention doesn’t scale.

Until now.

Subscribe now

There are two types of consumer businesses

You can think of there being two types of consumer businesses.

The first type is what we’ll call the scale business. That’s where the internet lives. That’s Amazon, Uber, Airbnb, your airline, your telecom, your bank. These are the businesses that serve tens or hundreds of millions of people every year. They’re simply extraordinary at logistics, price, selection, and convenience: and above all, they understand the art of scale.

But the sheer volume that they have to manage means that each individual customer relationship is thin. That’s a mathematical necessity; it’s just how it has to be when you’re dealing with unimaginable numbers of people. You get IVR phone trees, ticket queues, decision-tree chatbots. And the result is that however much people might need or enjoy the services that these companies provide, they hate the customer service that they receive—to the point that “dealing with customer service” is a near-universal signifier for “an annoying burden that I have to put up with.” And that makes sense: each day the equivalent of at least 366 human lifetimes are wasted just having to deal with the headache of being a customer. About a quarter of Americans say they’d rather shave their head than deal with customer service.

But for a certain small slice of people, there’s another type of consumer business. We can call it the concierge business. This is your Hermès, your Porsche, your luxury hotel, your executive physical, your Global Services hotline. And this type of business is organized around the figure of the concierge—a person who knows you, remembers your preferences, and anticipates your needs. “You liked the Bordeaux the last time; perhaps you’ll enjoy this Touriga Nacional?”

The concierge is the highest form of customer experience ever devised: ambient, proactive, personalized, continuous. It’s not episodic in the way that “customer service” is for the masses. You don’t only talk to the concierge when something goes wrong and you’re upset. The relationship is ongoing. You feel less like a harried customer and more like an old-school aristocrat waited on by a cast of servants.

And that feeling of distinction is a big part of the appeal. The concierge experience isn’t just functionally better, but socially better too. It signals that you are not part of the unwashed masses, navigating a phone tree. You exist in a separate, more civilized world, where someone knows your name and your preferences and treats your time as valuable. That sense of exclusivity is something luxury brands cultivate very intentionally: the concierge doesn’t just serve you well, it makes you feel different.

The difference between the concierge business and the scale business, by the way, isn’t a moral one. Bernard Arnault cares as much about making money as Jeff Bezos. The difference is just about ARPU. A dedicated, high-quality concierge costs real money—humans are expensive, of course, but they’re especially expensive when you account for the costs of training and institutional memory. At different ARPUs, “customer service” takes on an entirely different function for a business. For a scale business with a low ARPU, customer service is just a cost center; for a luxury business, customer service is everything.

Hermès can afford to give you a concierge, because you’re one of a relatively small number of people who are able to spend five or six figures a year there, and because the “personal touch” is at the core of what Hermès provides. United Airlines and Verizon can’t make the same call. So the rich get the concierge; normal people get the phone tree. That’s the basic deal for the consumers of scale businesses: you get the scale pricing, but you lose the intimacy. And that essential tradeoff has governed consumer business for decades.

But something fundamental is changing.

What happens when attention is free?

There’s a conventional narrative about what AI will do to customer service, and it goes something like this: AI will replace human customer service agents; companies will save money; headcount will go down. And I think that in the broad strokes this narrative is true. The portfolio company that’s at the heart of this shift is Decagon, and we’re seeing that their agents are able to successfully resolve a huge amount of customer inquiries, resulting in higher customer satisfaction and significant reduction in costs. Customer service agents can now focus on the most important problems. In the aggregate, this usually means that Decagon’s customers, massive scale companies like Mercado Libre or Hertz, can economize on customer service spend while creating a concierge-like experience.

But I think that the conventional narrative stops before getting to the really interesting second-order effects. And the second-order effects are where it gets really interesting.

Throughout economic history, we see a really reliable pattern play out. Alex Danco identifies it correctly as the Jevons paradox. When the cost of something drops dramatically, if demand for that good is sufficiently elastic, then we don’t just consume the same amount of that thing and pocket the savings. We consume vastly more of it.

And I think people are underrating the enormous amount of latent demand for “really, really good customer service”—for concierge-level service—in our scale-dominated economy. I think that every consumer of a scale business knows what I’m talking about. We want to be treated like people, not like case IDs.

Before, this just wasn’t possible in a scale business, for the reasons I mentioned. Everyone knows that better customer experience drives retention, increases lifetime value, and builds brand loyalty. United and Verizon don’t want to make you hate them. But they just can’t afford to provide excellent customer service at scale. The bottleneck is high-quality attention. You just can’t hire enough people, train them well enough, give them enough context, and keep them engaged with every customer until they’re as happy as can be. For any business without an absurdly high ARPU, the economics just don’t work.

But AI collapses the cost of high-quality attention. When you have hyperintelligent AI that is ultra-cheap, always on, and can run infinite parallel instances—each with a full memory of the customer’s history, preferences, and context—then the entire logic of customer engagement flips in an instant. The question is no longer “can we afford to pay attention to this customer?” It’s “what becomes possible when the marginal cost of a customer interaction approaches zero?”

The answer, I believe, is that every business becomes a concierge business. When attention becomes abundant, customer experience undergoes a fundamental transformation.

Think about what a great concierge does. They don’t wait for you to come to them with a problem; they notice that your payment didn’t go through and reach before you get frustrated. They don’t interact with you at moments of maximum tension; your relationship with them is ambient, ongoing, an integral part of the texture of your customer experience. And, above all else, they know you: they know your history, your preferences, your communication style, that you hate shellfish and prefer window seating and had a bad experience in Bermuda last year.

AI will make all of this possible at scale for every customer at every business. The “customer service” relationship stops being episodic, reactive, and generic, and becomes continuous, proactive, and deeply personalized.

That is the concierge model. And it’s what’s going to define the future of essentially every scale business.

And here’s what might be the most underrated part of this shift: when the concierge knows you and is always on, the distinction between “support” and “commerce” dissolves. A great concierge at Hermès isn’t just a fancy customer support rep; he’s a salesman in disguise. They know your taste so well that their recommendations feel not like a sales pitch but like a favor. That’s the model that AI could unlock for every business. Customer service, done well, has always been a revenue opportunity hiding in plain sight; but until now the costs of hiring a capable human have made it feasible only for high-ARPU businesses in the luxury sector. Now AI can make it available for every business. At the limit, “customer service” and “commerce” meld into the same thing.

I don’t think that all of this is going to manifest magically in the next month or the next year. But I think it’s clear that this is where the relationship between businesses and customers is going. It’s not just “better customer support reps,” but a different category of experience entirely.

The Decagon bet

And that’s why I’m so excited about Decagon.

We first invested in Decagon at inception, when Jesse Zhang and Ashwin Sreenivas—two repeat founders who, in a nice bit of serendipity, met at a ski retreat organized by a16z—set out to build something that went beyond the first-generation AI chatbot. From the earliest days, what stood out about Jesse and Ashwin was that they weren’t just thinking about deflecting tickets, or replacing the marginal customer service rep, or trading quality for slightly better margins. They were thinking about what the business-customer relationship should look like in a world where high-quality attention is practically free.

In the time since, Decagon has rapidly become a leading conversational AI platform for enterprise customer experience. They’ve grown from zero to powering concierge customer experiences for more than a hundred major consumer-facing enterprise customers—from Avis and Hertz to Block’s Cash App, Mercado Libre, and Oura Health. Customers are seeing deflection rates above 80 percent, meaning that the vast majority of interactions are handled without any human intervention. And, crucially, customer satisfaction scores are going up. When Chime deployed Decagon’s AI agents, they reported a greater than 60 percent reduction in contact-center operating costs and a doubled Net Promoter Score, the universal metric for customer satisfaction and loyalty.

That last detail matters enormously. The conventional fear about AI in customer service is that automation means a worse experience, that you’re trading quality for efficiency. But Decagon is showing the opposite. AI can actually deliver a much better experience than the legacy model, because they combine instant availability with deep contextual knowledge and near-infinite high-quality attention.

But the reason I’m really excited about Decagon isn’t just “better chatbots” or “better customer service reps.” The really interesting stuff, the really high-value stuff, comes from everything after the initial use case. Because Jesse and Ashwin aren’t talking about “cheaper customer service”: they’re talking about infinite, hypercapable, and hyperintelligent concierges for every customer at every business. They’re talking about a future where businesses can combine the massive scale of Amazon with the ultra-attentive concierge service of Hermès.

If the Decagon bet works—and the early evidence is very encouraging—then the AI concierge doesn’t just sit inside the customer service org like it has in the past. It becomes the primary interface between a business and its customers, in the same way that the concierge is the primary interface for existing concierge businesses in the luxury economy. That’s a much larger market than “customer service automation.” That’s the relationship layer for all of commerce.

Here’s how I’m thinking about the magnitude of what’s happening. In the 1990s and 2000s, email “replaced” physical mail. But because the cost of sending an email was zero and it arrived instantly, email didn’t just do what physical mail did but much cheaper. It assumed a totally different role in people’s lives. It created use cases that were inconceivable in the era of stamps and envelopes. The volume of correspondence exploded, and the nature of correspondence changed.

The same thing is happening right now with customer attention. If you think AI is just going to “replace” the call center agent, you’re stuck thinking about first-order effects. AI is going to create an entirely new category of customer relationship that was previously impossible, just like how email created an entirely new category of communication. When you can pay near-infinite attention to every single customer, every business becomes a concierge business. The concierge—that type of intimate, proactive, and personalized experience that is today an exclusive property of the luxury economy—becomes the default for all commerce.

Nanni, in 1750 BC, had to carve his complaint into clay and send it across Mesopotamia, hoping that Ea-nasir might eventually read it and feel shame. In 2026 AD, we’re still basically doing the same thing: firing our complaints into the void and hoping someone on the other end might care. But after four thousand long years of customer pain and frustration, I think that AI is finally bringing that era to a close. The future of commerce is concierge.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Editor’s Note: Today’s piece is from a16z Crypto’s Jason Rosenthal, on the essentials of setting up a sales team—send it to every early-stage founder you know. Prior to a16z Jason worked with at Opsware. You can read more from Jason and the a16z Crypto team in their excellent newsletter.

Founders obsess over product details: features, UX, performance, reliability, and documentation. Now imagine applying that same intensity to sales and go-to-market.

Technical founders sometimes have an aversion to investing in go-to-market and sales, believing that the best products sell themselves — but that’s not the case. And it’s especially true now that we’ve entered the enterprise adoption era of crypto. So building an effective enterprise sales organization is one of the most powerful motions for scaling a company after finding product-market fit. A world-class sales team can be just as important and effective as an outstanding engineering team, and it can unlock new levels of growth and operational excellence that competitors can’t replicate.

It can also be surprisingly exciting and satisfying — even for deeply technical, product-oriented founders — to build a sales organization. The reason is simple: success here is tangible and metrics-driven. You can literally count the number of marquee account wins and losses, see how this translates into increased revenue, and quickly separate the 100x performers from the also-rans.

Here’s a step-by-step playbook to help you build a world-class enterprise sales organization.

Step one: Build the right team

Sales Leader

The first and most critical hire is your sales leader. They lead all revenue efforts across the company and design the sales playbook: how the organization goes to market, how it operates day to day, and how deals are won. This role shapes the processes by which campaigns are fought and victories are secured.

You must give this leader a seat at the table. They should report directly to the CEO and participate in executive staff meetings alongside other senior leaders so they are fully integrated into company strategy and decision-making. Otherwise you’re not signaling that sales is a priority in the organization, or are siloing it too much as a function.

Founders often make two mistakes when building a sales organization. The first is starting by hiring junior sales talent to “explore the market” before bringing in a sales leader. This strategy almost always fails. The main reason is that any enterprise sales leader who’s worth their salt will want to hire and shape their own team.

The second mistake is waiting too long to hire this role because founders believe they themselves are the best salesperson for the company. Early on, it’s true that a CEO can be the company’s most effective salesperson. They know the product intimately, they have “the divine flame” of the company burning brightly inside of them, and customers love to talk to a founder and CEO.

But if the CEO remains the primary salesperson for too long, it prevents the company from moving quickly enough to capture the market opportunity. So the best time to make the first critical sales hire is once a founder believes that they’ve achieved product-market fit or are on the cusp of doing so. Waiting too long to fill this seat risks competitors getting ahead and defining the market on their terms, potentially boxing you out.

When searching for a sales leader, it’s critical to find someone who has previously written a playbook for selling new technology into a new market, rather than someone who may have successfully executed a playbook that was designed by someone else. This role requires both creativity and experience to design the right plan and process for your company.

Ideally, this experience includes bridging frontier technology and selling it to new customers and markets. Other core competencies: operating in regulated environments; and the ability to translate technical innovation into enterprise outcomes like ROI, performance, compliance, and reliability. These are all key criteria enterprise buyers often use when making purchase decisions.

Sales leaders also need the discipline to build an auditable, repeatable sales process — and the judgment required to recruit strong enterprise sales talent. The best sales leaders are missionary leaders who attract like-minded, hardcore enterprise sellers.

Subscribe now

Account Executives

The next most important role is the Account Executive (often referred to as an AE or account exec). The AE interfaces directly with prospective customers and introduces your technology into enterprise accounts. The AE maps the customer organization. They must identify the true decision makers and communicate the value proposition in a way enterprise buyers understand and will respond to.

Successful account execs typically have prior experience selling similar products into related markets. Because they often come with a deep rolodex, founders often assume that the right account executive will unlock doors through their pre-existing relationships. But this rarely works in practice. Enterprise decision makers change frequently.

What matters far more is demonstrated ability to close large deals, navigate complex organizations, and win against strong competition. You can easily assess this in interviews because salespeople have clear metrics and outcomes, and treat these wins as their scoreboards.

Account executives are commonly organized either by vertical or geography. How many AEs you should hire depends on two factors: the number of potential accounts that require active outreach, and the number of deals each AE can manage simultaneously.

One helpful tool for determining this is to run a bottom-up analysis: Estimate how many active deals you expect over the next twelve months and how many deals a single AE can realistically handle. From there, calculate how many account executives you need. While the exact number is market- and product-dependent, this approach usually produces a reliable first-pass coverage model.

Some words of caution: First, don’t let your closest competitors out-cover you. If they’re the first to reach the most important accounts in your market, they’ll shape the customer’s requirements and define the rules of engagement — leaving you stuck competing against their playbook, from messaging to positioning, instead of your own. Second, avoid hiring account executives before finding product-market fit. Otherwise most AEs will struggle to sell products customers do not yet understand.

Solutions Architect

Solutions Architects, sometimes called sales engineers, are the next essential role in building a winning sales organization. Their role is to translate the company’s technical capabilities into the customer’s business and operational outcomes. They serve as a bridge between customer needs and the product and engineering teams. Through this feedback loop between product, engineering, and customer needs, solution architects surface product gaps, identify future requirements, and ensure the product evolves in the direction leading customers’ demand. This virtuous cycle strengthens both the product and the sales motions.

Since the CEO knows the product better than anyone, they often end up spending a significant amount of time translating the product and value proposition to customers. But upon hiring a solutions architect, the CEO can offload a significant amount of early sales process and sales calls. For this to succeed, the CEO must train the solutions architect directly so they can communicate value as effectively as the founder; this sometimes involves shadowing the CEO in initial sales meetings, or sharing sales training decks

A strong Solution Architect paired with a great AE and a compelling product is a lethal combination.

Solution Architects work closely with enterprise customers to determine how your company’s product can integrate new capabilities into their existing technology stacks. In the case of crypto, where significant amounts of money and value are transferred across blockchains, institutional customers almost always rely on complex existing infrastructure. The solution architect plays a critical role in helping enterprise customers understand how everything fits together, as well as relaying product requirements back to the company — ensuring that the right interfaces and integrations are built.

It may seem obvious, but successful solution architects must have strong communication skills. They spend roughly 70 percent of their time with customers and 30 percent translating insights internally. Solution architects should also be technically oriented, although they do not need to be hardcore developers.

I recommend hiring the first solution architect at the same time you bring on board your first account executive, since they often work closely together. Don’t worry about having one-to-one mapping between a solution architect and an AE; one solution architect can often support multiple AEs at the same time.)

Sales Development Representatives

Sales Development Representatives (SDRs) generate qualified leads for account executives through email, social, and phone outreach. They perform early qualification — triaging, basically — before passing leads downstream to the account executives.

As a company starts to scale revenue, sales development reps become an increasingly important source of new leads. Historically, SDR roles were populated by junior sales hires aiming to become AEs. But with the advent of AI-assisted sales tools , the SDR role now has a more quantitative dimension. SDRs can use data to understand which outreach channels and specific messages are converting best, and adjust their approach accordingly.

The first SDR should be hired only after the head of sales, first AE, and first solutions architect are in place – again, all after product-market fit. Because even if SDRs generate leads, without a mature process to convert them, those leads will be wasted.

Sales Operations

The head of sales operations ensures the machine runs efficiently. This role ensures the team follows processes, maintains accurate forecasts, and designs the sales funnel so leadership can allocate resources based on data rather than just intuition. This role usually reports to the head of sales.

The head of sales ops should be hired only after you’ve already filled the positions covered above (head of sales, sales rep, solution architect, sales development). This person’s role is to scale and tune the machine — not to build it. If you don’t already have the essential capabilities in place for lead generation, technical selling, and relationship management, then the machine your head of sales operations is meant to optimize will be incomplete.

The sales ops role is usually interrupt-driven: There’s always a fire somewhere in the pipeline. It’s essential to hire someone who can parallel-process multiple streams of incoming data and information, yet also dive deep into the details when necessary. Throughout all of this, they must remain calm under pressure.

Sales ops will not fix a broken sales motion or weak product. If deals are consistently lost due to weak differentiation, late entry, or product gaps, the root cause lies elsewhere. If you’re losing deals versus the competition; or are second or third into the accounts that matter; or have a value proposition that isn’t resonating with the market, then don’t fool yourself into thinking that investing in sales ops is going to solve those problems.

Customer Success

Once you’ve won a hard-fought deal, it’s critical to ensure the product is properly implemented and hitting the key business metrics for which the customer purchased the product in the first place. The customer success function ensures customers achieve the outcomes that justified their purchase.

The customer success goal is retention, expansion, and referenceability: ensuring customers stay with your product; upselling and expanding their use of the product as fitting; and serving as references when the customer is highly satisfied over time

A customer success associate will also regularly channel deep product insights back to the product and engineering teams. They identify product gaps and critical third-party integration needs, like which external systems the product must connect with to ensure customer success. They also surface sources of implementation friction that you can address through ongoing product development, making future deployments faster and more seamless.

Successful customer success associates must be very customer service oriented, taking on a client-facing approach each day with a can-do attitude. They also must be excellent multitaskers, juggling multiple issues and clients simultaneously. When finding customer success associates, there’s a huge talent pool to draw from in B2B SaaS — basically every software company from Box to Oracle to Salesforce and so on that has a recurring revenue vs. on-premise software model. Employees from these kinds of organizations will have seen all kinds of escalations and team configurations so will have lots of experience to bring.

Hire the first customer success employee once you’ve landed your first lighthouse customer — not just any paying customer but the one that is aligned best with your product/ market and future growth — and that deployment starts to scale. Otherwise, you risk your product and engineering team getting sucked into a full-time customer support and satisfaction role, which is not the highest investment of their time.

Step 2: Building a repeatable sales process

Sales organizations, like software development processes and factories, perform best when they are built on a clear, repeatable process. Each member of the sales field organization viscerally understands the playbook and knows the role they need to play. Doing so is what separates good sales teams from great sales teams.

Ask yourself:

What exact steps should your business development reps, account executives, solutions architects, and customer success associates follow to give your company an unfair advantage in winning deals?

This question covers everything from how you first approach a customer, to the written and presentation materials required at each stage of the sales campaign. The goal is to ensure your company is setting the agenda, defining the requirements, and boxing out the competition.

Once you’ve defined this process, it must be instrumented. Monitor its health at every step, including in your weekly forecast calls, so you understand where each deal stands and what the true win probability is. At the early stage of a sales buildout, companies often fail to distinguish which deals matter most. Not every deal is created equally, and understanding this can be helpful for sales forecasting and then building your pipeline to your customers.

You should invest significant effort in winning the deals that shape the market and influence future buyers, especially in enterprise sales where later customers look to early, sophisticated buyers for validation. Winning these lighthouse accounts early on is critical, both for your revenue goals and company evolution — losing them also creates openings competitors are eager to exploit.

Frame the decision as either a once-in-a-lifetime opportunity, or an existential threat to avoid, prompting urgency and a clear motivation to buy today. This requires a deep understanding of your prospective customer’s strategic business priorities and the ability to map your product capabilities directly to solve those problems. That might mean reducing existential competitive threats, improving audit and compliance posture, cutting costs, or enabling a high-return new line of business.

If you’re in a competitive dog fight with another company, lay traps and landmines for them. If you possess capabilities competitors lack, ensure those capabilities become central to the customer’s decision-making process. Your competition will find themselves unable to achieve the same results or demonstrate that capability. You can even define terminology and key metrics on your terms, so that when the competition sells after you, you’ve biased the customer to your worldview.

Creating internal champions for your product is equally critical. In most enterprise deals, there are at least three key decision-makers: the end user (who will encounter your product day-to-day); the manager accountable for results; and the executive sponsor. If you can become the preferred choice for all three of those buyers, your chance of winning the deal is close to 100%. If, however, you have a great relationship with the executive sponsor, but the end user and the manager prefer another product, that’s a much more difficult deal to win. You must think about what steps you can take to win each one of those parties.

You may also encounter detractors within an account — someone who prefers a competitor’s product because they have a relationship with that company, or someone who has worked with a sales rep from the competitor before. In this case, it’s important to think through how you can turn them into a champion. If that’s not possible, attempt to neutralize their influence and impact on the overall decision by finding other champions, sharing data or references they can’t argue with, and so on.

Every enterprise sales organization experiences the pain of lost deals. When this happens, it’s important that you run an after-action review process where you review the deals that you lost, and try to understand what went wrong. Were there product gaps that caused the loss that you previously didn’t know about? What are the steps that you can take to fix them? Understand whether the loss stemmed from product gaps, late entry into the sales cycle, or execution failures, and feed those insights back into product and sales processes.

Finally, you must ensure that your entire sales organization is well trained on your sales process. The best way to do this is to put every new hire through a training boot camp, where you drill them on the key attributes of the product and the technology. Walk them through the sales process so that they know what to do and when. Most importantly, every single member of the team should walk through a simulation of a competitive deal. This way they know what it feels like to operate under pressure and competitive tension, and have responses and strategies in place.

Step 3: Avoiding common mistakes

I’ve already mentioned a couple of common mistakes: believing that your technology is so great that it can sell itself. It won’t. Get over it. Second is making your first sales hire a junior sales account executive who can “explore opportunities” rather than focusing on a strategic high impact ahead of sales.

Next? Assuming that a bottom-up, developer-led approach to sales will work within large enterprises. Many of the last generation’s successful companies — including Stripe and Twilio — were able to gain early momentum by selling to developers first. The same has been true in crypto. But to penetrate the largest and most complex accounts, relying solely on a developer-led approach simply will not work. Focusing on developers can be amazing for developing early technical champions within an account, but it’s still an absolute requirement that you master the process of selling to managers, executives, and navigating complex procurement processes.

Finally, you can’t allow the competition to be first into the most important and strategic accounts. When this happens, you’re giving your competition the opportunity to define the rules of engagement and the key feature requirements and capabilities for a given sale. You must be first into the most important accounts so that you’re able to set the rules of engagement and make everybody else follow them.

Step 4: Creating a culture of improvement

Creating a culture of continual improvement is crucial in building a winning sales organization. With constantly changing technology and market cycles, the best sales teams relentlessly incorporate new data and learning into their processes. If a sales organization isn’t improving, it’s falling behind.

At the core of improvement are three questions:

1. Why should the customer do anything at all?

2. Why should they act now?

3. Why they should choose you?

The organizations that consistently refine and strengthen their answers to these questions win disproportionately. Continual improvement also requires not just winning but retaining the lighthouse customers who set the tone for the rest of the market. The largest and most sophisticated buyers create the signals that future customers follow, positioning your company for long-term success.

Another critical element is a rigorous win-loss analysis each quarter. Understanding exactly why deals were won or lost allows insights to compound across sales execution, product development, implementation, and positioning. The questions to ask:

• Over the past quarter, which deals were won and which were lost?

• For the deals you won, why did you win them?

• For the deals you lost, why did you lose them?

• What improvements can you make to increase your win rate?

• What improvements do we need across the company to improve the sales process?

• What changes should we make within the sales team to improve outcomes?

• What product improvements would help increase the win rate?

• How do we improve implementation?

• How can we improve market strategy and positioning?

• Are we incorporating learnings from these win-loss analyses into our process?

• Are these learnings helping results compound quarter over quarter?

• How strong is our sales hygiene?

• Is the team consistently following the sales process?

• Is the team executing the sales plan with force and precision?

These questions may seem basic, but many sales organizations never bother asking them in any consistent way. These learnings compound over time and steadily improve results. Maintaining strong sales hygiene is equally important: Ensure your team is following the process with discipline rather than assuming their compliance.

Finally, continual improvement means studying how your smartest customers use your product. Reducing friction and increasing ROI for them creates insights that feed back into product, marketing, and sales motions: a virtuous circle for both your customer and your company.

***

The first lesson is accepting that technical products don’t sell themselves. The second is figuring out sales. But it’s these details in between — the right hires; the best lighthouse customers; continual improvement feeding back into your sales process, product, and people — that will really help you win.

a16z crypto

your guide to the next internet from the team at a16z crypto

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

With AI, startups and their customers have focused their attention on net-new capabilities and the products they’ve enabled. Think shiny new voice agents, workflow automation tools, and text-to-app platforms.

While there have been and will be many exciting businesses in these categories (we are invested in several!), AI will have a massive impact on something far less glamorous and far more valuable: helping organizations get more from the large amount of software they already run. To ask a question that sounds almost disrespectful until you’ve spent a week in a Fortune 500: why do people still use SAP (and ServiceNow, and Salesforce) at all?

The short answer is that SAP, or any major legacy system of record, captures critical data across the businesses that use it. But on top of that, the business has customized it and built a set of specific procedures and roles on top of it, much of which is not actually documented anywhere. Moving off has been painful, expensive, and time consuming - often requiring an army of consultants, years of time and hundreds of millions of dollars. Upgrading from SAP ECC to SAP S4HANA can cost $700M, take 3 years, and require a team of 50 from Accenture. And then after the move, the software is almost only useful for generating read-only reports that are impossible to manipulate.

That is until now. AI unlocks that opportunity to upgrade, customize, replace, and frankly better access and use the data captured in these systems of record.

Ultimately, the destination with AI might not be to “replace SAP/ServiceNow/Salesforce,” but to make them more programmable and approachable. The winners will be the platforms that (1) plug into transformation budgets with measurable risk and timeline reduction, then (2) expand into day-to-day operations as the trusted control plane for work, gradually unbundling the legacy UI into composable, governed, AI-assisted actions and thin apps. In other words, the systems of record endure; the interface, automation, and extension layer becomes the new software frontier.

Subscribe now

SAP is Painful, And Yet We Still Use It

To set the stage here, let’s share a bit about SAP and what it does. On the surface, these systems are hard to navigate, painful to change, and somehow still the backbone of how the world’s biggest organizations operate. Consider what it looks like to use SAP!

But that “somehow” is the opportunity.

The uncomfortable answer is that, beneath the ugly UI and the endless configuration, these systems are very powerful: they encode the canonical data model of the business, the permissions and controls that keep it compliant, the workflows that make it operable at scale, and the integrations that connect dozens (or hundreds) of downstream processes. They’re not “apps” in the consumer sense, they’re accumulated institutional memory expressed as tables, roles, approvals, posting logic, and exception handling.

Replacing this is more than just expensive; it’s risky. And the more a company has invested – custom fields, workflows, pricing rules, reporting logic – the more the system becomes a moat of switching costs and a competitive advantage. That’s also why extensibility is so powerful: every enterprise is unique, change is constant (new regulations, new products, new org structures), and these platforms survive because they can be bent to fit reality. The challenge is that the same extensibility that makes them valuable also makes them brittle: each customization becomes a future upgrade landmine, each workflow a maze, each screen a tax on every human who has to use it.

This brittleness shows up everywhere. User satisfaction with CRM remains mixed despite broad adoption, and heavy customization in ERP is consistently tied to timeline and budget overruns. Workers are drowning in fragmented workflows - digital workers toggle between different applications ~1,200 times per day (about 4 hours a week lost), and 47% of digital workers struggle to find the information they need to do their jobs. Large-scale “transformations” routinely stumble; one estimate has roughly 70% failing to meet objectives. The spend tied to this friction is enormous: the software implementation/system integration market alone was about $380B in 2023.

The process and pain here presents an opportunity for AI to change the way this software is implemented and used. The easiest way to understand the opportunity is to follow the lifecycle of the suite: first you implement or migrate it, then you live inside it every day, and then you build on top of it as the business changes. In each phase, the job is turning messy human intent into correct, auditable action against systems of record.

Let’s consider how AI can improve how we use legacy software systems at every stage.

Implementation

Let’s start with implementation - the riskiest, most budget-sensitive phase and the one with the clearest payback. Concretely, that looks like turning messy discovery (meetings, docs, tickets) into structured requirements, then auto-producing the implementation workstream: process and field mappings, config and code, test scripts, cutover plans, and migration playbooks - plus the data cleansing and validation needed to go live. This is hard to get right: German supermarket giant Lidl once famously scrapped its effort to transition onto SAP after spending $500 million.

Companies here build copilots, project management tools, and other software to help with migrations and implementations. Here are some examples of startups working in the space (Andreessen Horowitz is invested in some of these companies):

• Axiamatic is an AI “assurance” layer for ERP: it builds a knowledge graph from project artifacts and flags hidden failures in requirements/change management via Slack/Teams to de-risk and accelerate S/4HANA programs (partnered with SAP Build; baked into KPMG/EY/IBM motions).

• Conduct is a code- and process-mapping copilot that generates a semantic layer and technical documentation across ECC→S/4, with Q&A over custom tables/APIs to speed internal takeover.

• Auctor does agentic implementation delivery for SIs/pro services, auto-capturing discovery into structured requirements before becoming a system of record for SOWs, design docs, user stories, configs, and test plans.

• Supersonik helps with AI-powered product enablement for channels/MSPs and customers - vision and voice agents that teach inside the real UI, reducing SE headcount needs and enabling reseller-led implementations/expansions.

• Tessera’s AI-native SI manages enterprise transformations end-to-end - connecting into a customer’s existing ERP instance, evaluating how it’s implemented, and then flagging/automatically remediating what needs to be changed during migration.

These companies create value by making transformations faster, cheaper, and less risky. They do this in a few key ways: catching problems early in requirements and change management before they snowball, compressing timelines (where a single slipped month can cost millions, turning messy project data into structured knowledge so internal teams can take ownership faster, and reducing reliance on large SI teams through automation of mapping, documentation, testing, and enablement.

We see room for more startups building tools that work with existing partners rather than against them. Specifically:

• Implementation agents that share in outcomes and risk (think requirements tracking, config comparison, cutover simulation, code generation, and drift detection)

• Semantic documentation tools that keep knowledge current and accessible

• Enablement agents that turn training and channel rollout into a repeatable product

Because startups can alleviate enterprise-level burdens, they can price to the delay avoided, and sell into the transformation budgets CIOs and CFOs are already spending, displacing bloated SI engagements in the process.

Usage and Maintenance

Next, after a software suite has been implemented, using it means navigating the mess of a UI these software suites have today. Day-to-day work spans dozens of screens, role turnover resets know-how, and a long tail of edge-case workflows never gets first-class product treatment. Users spend time hunting for fields, mirroring data between systems, and asking ops teams to “just run this report.” The result is slow cycle times, avoidable errors, and persistent training burden.

The opportunity is for AI to wrap legacy systems with a friendlier, more capable “system of action.”

Companies in this category build tools that help teams get more out of the systems they already use. In practice this looks like a copilot that lives in Slack or as a browser sidecar, that can answer “Where can I find X?” or “How do I do Y?” using semantic search , and then take safe actions (create a case, post a journal entry, update supplier terms) via APIs when available. These tools can also chain together compose multi-app workflows (“pull last quarter’s POs from SAP, check contract terms in Coupa, draft a variance note in ServiceNow”), with human approval steps, audit trails, and granular RBAC. The best ones track adoption, time saved, and error rates.

A lot of the work that matters in enterprises still isn’t exposed cleanly through APIs - it lives in screens, thick clients, VDI sessions, and half-documented admin consoles. That’s why modern “computer-use” agents are such an important complement to API-first copilots: they expand the reachable surface area of automation to the last 30–40% of workflows where there simply isn’t a reliable endpoint to call. The core capability isn’t “clicking buttons,” it’s reliability under mess - agents that can perceive the UI, anchor to stable elements, recover from pop-ups and layout drift, and checkpoint progress so they can resume safely mid-flow. When paired with validation (diffs, reconciliations, sandbox runs) and enterprise controls (SSO, secrets, least privilege, audit), this turns what used to be manual work into governed, repeatable automation - ticket triage, period-close steps, customer updates, pricing changes - even in the parts of SAP/ServiceNow/Salesforce that vendors never built for automation. APIs make the happy path fast, and computer-use makes the long tail automatable.

Companies like Factor Labs and Sola are already deploying these agents in production, replacing BPO spend and helping large organizations automate tasks at scale.

Extensions

Finally, even if you make SAP/ServiceNow/Salesforce easier to use, your business will keep changing which means your system of record will have to as well. New products, new policies, new acquisitions, new regulations, and a long tail of workflows that will never justify a core-module project mean constant work to keep software relevant to the actual state of your business. Historically, teams have had two options: customize the suite (and inherit the brittleness tax) or build one-off apps (and struggle to integrate, govern, and maintain them). This is the third wedge for AI: shipping small, governed experiences on top of systems of record, fast, while keeping the core clean.

Building net-new tools and automations on top of legacy estates becomes the “Lovable” layer over unloved software. The pattern starts with a unified data-and-action plane: read from systems of record via APIs and events (and safe UI capture where needed), normalize into a semantic model of business objects (orders, vendors, cases), then expose a governed set of actions with RBAC, approvals, and audit.

On top of that plane, teams ship focused experiences that feel modern and purpose-built. Instead of sending a procurement analyst through 12 SAP transactions to onboard a supplier, you give them a single “Vendor Onboarding” thin app that gathers documents, checks duplicates, routes approvals, and posts the right records back to SAP. Instead of asking RevOps to open five Salesforce screens to update renewal terms, you give them a spreadsheet-speed editor that can bulk-edit, validate against policy, preview impact, and then commit changes with a full audit trail. Instead of yet another “portal project,” you give frontline teams a command palette that can answer questions and execute the handful of actions they do every day (“create return,” “extend credit,” “open a Sev-2,” “post accrual”) across multiple systems, without spelunking through 20 tabs.

These extensions also unlock cross-system workflows and automations that no single vendor would ever prioritize: event-driven triggers like “if invoice posted AND variance >3% → draft an explanation → route for approval,” or “if ticket reopened twice → create problem record → assign owner → update customer,” with human-in-the-loop checkpoints where it matters. Over time, the most valuable deployments turn into reusable “intent packs” — quote-to-cash, vendor onboarding, period close — that encode not just what to do, but how to do it safely in your environment.

Platforms like General Magic’s Cell make the building blocks for architecting these bespoke workflows tangible: you upload OpenAPI specs so every endpoint becomes an action, then embed a native command bar with a single script tag that executes real API calls, backed by analytics, multitenancy, security guardrails, and RBAC, so the work shifts from rebuilding another UI to composing the right actions and policies on top of systems you already trust.

What does the endgame look like?

We think legacy systems will mostly persist, but they’ll stop being the surface area where work happens. ERPs, CRMs, and ITSM suites are too embedded to be ripped out on a typical software cadence; they evolve slowly and remain systems of record. What will change is the user-facing “system of action” that sits on top: AI will become the default interface for discovering how the system works, executing workflows across it, and shipping small, modern experiences that bypass legacy UI altogether. In other words, the bridge becomes the highway.

Enduring software in this category will look less like a chatbot and more like an operating layer: a unified data and action plane with a semantic model of business objects, plus guardrails that make AI trustworthy in production. If you’re an end user, instead of learning which screen, field, and transaction code to use (and then re-learning it every time the UI or process shifts), you describe the outcome you want and the system gets you there. You’ll ask a couple clarifying questions, be shown a preview of what it’s going to do, and then the tool will execute with the right approvals and an audit trail. Closing the loop looks like “create a return and notify the customer,” “open a Sev-2 and pull the last three related incidents,” or “onboard this vendor, collect the docs, route approvals, and set payment terms” — actions that today require hopping across SAP, Salesforce, ServiceNow, and spreadsheets. This gives us fewer errors and reversals, less dependency on tribal knowledge, faster cycle times, and dramatically lower training burden because the interface is intent-driven, role-aware, and self-serve by default.

The moat compounds from real usage: every successful workflow becomes a reusable intent, every exception becomes a guardrail, every migration artifact becomes living lineage, and every integration deepens the graph of how the enterprise actually runs. Over time, the “AI layer” becomes the place teams go to understand change impact, prevent drift, measure ROI, and ship new workflows, even when the underlying systems remain the same.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.