More than 130 people across two cohorts have now joined the a16z New Media Fellowship, an 8-week cohort of world-class creators, writers, marketers, podcasters, and behind-the-scenes operators.

Over the past year, the fellows have been busy. They’ve founded and raised money for new companies like Energy, Osmo, and Underdog. They’ve led launches at fast-growing startups like Thatch, Ghost, and Exa, and built new media properties like MTS, Substrate Magazine, and SOTA Wives. Others have taken on new roles at companies like OpenAI and Prime Intellect, and a few even joined our own New Media team. You can read about many more of them here.

Two cohorts in, we’ve learned that the best parts were the simplest: namely, good dinners and a vibrant, terminally online group chat.

So we’re launching the New Media Fellowship Cohort #3 and doubling down on exactly that.

NMF3 will run for eight weeks, starting in January 2027. Dinners will be hosted in both San Francisco and New York – but fellows can be based anywhere, and everyone will be in the group chat.

Subscribe now

Fellows will also have access to opportunities, resources, and ideas to explore with companies across the a16z portfolio.

Apply to join NMF3

This upcoming cohort already includes founding fellows like:

• Tyler Gold, Creators at X

• Samantha Trimble, Ecosystem at OpenAI

• Elsa Johnson, Editor Emerita at The Stanford Review

• Austin Lyons, Founder of Chipstrat

• Rosie Sargsian, Narrative at Ramp

• Eli Goldfine, Writer at The Supercycle

Applications are open now. We’ll be reviewing applications on a rolling basis. You can apply here.

If you have any questions, email ggoldberg@a16z.com. We look forward to hearing from you!

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.

Replit for Business launched today! To hear more from Amjad Masad, CEO of Replit, see his post here on the rise of the self-driving company.

One thing I love about Amjad is, sometimes he really says the thing and articulates something that other people weren’t brave enough or clear enough in their thinking to express. He has one of those today, as part of Replit’s idea of an Agent Habitat as the way that more businesses are going to become self-driving companies. What Amjad says that’s so on the money is one great word: chaos.

“Today, most agents live somewhere chaotic. They run in chat windows, personal accounts, and one-off scripts, holding whatever credentials and data someone handed them. Nobody decided what they could touch, and nobody can see what they did.”

Chaos! It’s one of the most underrated concepts in disruptive technology adoption. One of the best tells that a technology will actually change the world is, “Does it make a chaotic entrance?” And then the most important question thereafter, as Andy Grove put it, is, who is going to let chaos reign, and then rein in the chaos?

Throughout the history of computers and the internet, the most interesting, disruptive and ultimately consequential technologies are the ones that enter workplaces in a chaotic sort of way. What Replit really gets is that this joyful chaos can’t be extinguished: it has to be reined in just the right amount, so that all of the bottoms-up enthusiasm and ownership stays alive within an optimal habitat. Because the same excitement and ambition driving the initial chaos is also what’s creating deep organizational alignment, from the bottom up inside businesses, that makes these companies come alive.

Agents are joining a long tradition of technologies that started as chaos and got reined in just the right amount:

Computers

Many people working in tech today have no memory of this, but the early wave of adoption for personal computers at work was an incredible “let chaos rein, and then rein in chaos” sort of story. Not because IT wanted it that way! But because that’s just how it happened.

Like ChatGPT a few years ago, PCs mostly arrived because people brought them to work and plugged them in, allowed or not, because they helped them get things done. Once the secret was out, everyone wanted one. People lied on their procurement forms, claiming they were buying an oscilloscope, or a typewriter, or even a desk. People wanted PCs so badly that they were just necessarily going to arrive in the office in a pretty chaotic way.

IT departments did not like this chaos! From a 1983 article, “Guerrillas Shift Power Balance In The Office”:

With a personal computer, Keen notes, the enterprising employee can often bypass the company’s main computers to calculate profit-and-loss projections or track sales calls. More important, the personal computer can be linked to the main computer to siphon off desired information.

If a corporation isn’t careful, says Keen, the proliferation of personal computers will lead to “shadow” data processing departments that perform corporate calculations hidden from top management’s view.

Despite this knee-jerk reaction to regain control over these rogue PC users, smart business leaders understood how valuable this behavior was. Employees were staking a claim on new ways to work, and taking fierce ownership over them. Furthermore, this new behavior was illuminating and reinforcing the alignment inside the company that actually mattered. You can’t become an order of magnitude more productive overnight, just by adding a new tool, unless you know what to do, and have the alignment you need with the business to go out and do it.

What they needed was the right habitat for people to have computers at work. Which meant a bunch of learned IT practices and workplace “harnesses,” for people to actually use PCs in a way that kept the bottom-up productivity intact without sinking the organization into chaos.

Spreadsheets

Meanwhile, the PC’s original killer app - the spreadsheet - was itself creating a new kind of chaos.

Spreadsheets, in another sort of way, also have a claim to be the original “workplace agent.” Suddenly all kinds of people who didn’t consider themselves “programmers” could model how they felt their business should run, calculate how it was doing, and project what was possible. Spreadsheets caught on so quickly that they became the single most important reason why people wanted their own PCs.

This may be difficult to imagine in 2026, but “rogue spreadsheets” back then created a very similar challenge to “rogue agents” today. You had all of these people who were now able to create real-looking business artifacts and projections, and while some of them were phenomenally valuable, you also got a lot of spreadsheets that were, essentially, hallucinations. They could be horribly documented, hard to audit, full of formula errors, and usually exactly one employee actually understood what was in them: the person who’d made it. Mitch Kapor, the original spreadsheet king, put it to Steven Levy in 1984: “You can just randomly make formulas, all of which depend on each other. And when you look at the final results, you have no way of knowing what the rules are, unless somebody tells you.” Chaos!

The answer, again, was a habitat where chaos could be reined in a little bit. Data access rules inside organizations, conventions around locking cells, version histories, standardizing formulas, and auditing output. And then inside that habitat, people continued to use that individual empowerment and autonomy that the spreadsheet gave them, and they’ve been indispensable to organizations ever since.

Email

Yes, even email entered the workplace more chaotically than you might’ve realized.

Back before email, companies had the mailroom. It looked unimportant, but it was one of the most politically important functions in the company. The mailroom was where the company could control the flow of information inside the organization. Who could reach out to whom, and who could view, copy, or circulate what. All of it ran through the mailroom.

And then email came out, and suddenly, it became really easy, at least in principle, for anyone to contact anybody inside the company. Chaos! This was, in complete seriousness, a bit of a crisis of authority and control for executive teams. And it prompted huge anxiety around secrecy and information control: what happens if employees accidentally forwarded sensitive information? What if they shared it with someone they shouldn’t? What if they did it on purpose? It was considered a reasonable idea, for a minute, to have to get managerial approval to send an email.

Thankfully, again, we created habitats for company communication where employees could reap all of the benefits of instant communication and applied initiative, with the necessary permissions and access control to rein in the communication chaos just enough.

So there are two lessons here from the history of computers. One is, someone has to rein in the chaos a little bit, and create the habitat for the behavior to thrive in an optimal way. The second lesson is that the chaos is an important indicator in the first place for a kind of behavior that’s really valuable: the kind that creates deep alignment inside the company, for it to do its best work.

The self-driving company

In all of those examples from earlier in the computing era, why was the initial chaos such a strong indicator of something important and valuable? Because people were staking a claim on something they immediately found valuable, and wanted to homestead. People felt they’d discovered something deeply aligned with their individual ability to contribute, and with the organization’s needs.

Businesses are made of people who care. That was true before computers and before agents, and it will be true when companies become “self-driving,” as Amjad puts it. A successful business needs employees who are empowered and aligned with the company’s goals, and a way to turn that alignment into real artifacts at work.

When disruptive technology comes along that makes people radically more productive, the first people who really come alive are the people who know what to do. There is no substitute for these people. They know what needs to be done, and now they have a tool - whether it’s the PC, the spreadsheet, or now a fleet of agents - that lets them get it done 100x more productively than they could before. I believe Replit has an example where one employee at Sumitomo Mitsui Finance and Leasing took over a development process that was supposed to take six people six months, and did it themselves in one week. That one person who knows whats going on can now just do it themselves.

And, to a first order of approximation, the business succeeds based on how successfully these people are able to translate their individual alignment with the business into overall alignment on the part of the whole org, through the artifacts that they produce, and by the followers they attract.

Code is the ultimate “aligning artifact” that someone can produce, because it literally is the business logic of the organization. But it’s an inherently chaotic thing to generate, if suddenly all kinds of new people are creating it from the bottom up. That chaos is an important sign that something important has happened - just like people bringing their own PCs and creating “shadow IT departments”, or people creating their own spreadsheets or sending their own emails. And it’s a good indicator that you need a habitat for that behavior to thrive. You don’t want to put it in a cage. Nor do you want to let it run completely free. But you want it free-range.

To become self-driving, a company has to become aligned. And nothing creates alignment quite like expanding the opportunity set of who can contribute code. Soon, you’ll be able to measure a company’s alignment by the richness of its agent habitat. Replit and its customers already get this. Most companies don’t yet. In hindsight, it’ll be pretty obvious, once the self-driving companies have pulled ahead.

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.

Here is a message from Joel, Zane and the a16z infra team that we sent to our portfolio companies yesterday, regarding the current cybersecurity situation. Stay safe out there!

The previous week should make it obvious that things have changed. Vulnerabilities are being discovered and exploited at a rate that has never been seen before. While there has been a flood of new vulnerabilities AI agents are being adopted at scale multiplying the potential risk. Agents not only help automate the attacker lifecycle, every AI agent is an insider threat. Your environments are designed to withstand attacks from the outside, not from within. As such, we are suggesting company leaders consider the following drastic changes:

1. Significantly increase resources for security work. The average enterprise allocates roughly 10% of its tech spend to security. Consider doubling that. There are likely thousands of vulnerabilities that need fixing in your environments. Security through obscurity no longer works, all that can be exploited will be exploited.

2. Establish a zero tolerance policy for new features and product releases. Re-examine launch and release schedules and consider delays to ensure all vulnerabilities are patched before deploying new features or products.

3. Double tap the zombies. Find all old legacy environments that have been neglected and turn them off. Labs, test environments, zombie CI/CD projects etc. Turn it off for now. If it isn’t tied to revenue, kill it.

4. End the risk acceptance process. Over the last 10 years the risk acceptance process has morphed from a way to accept risks to enable the business, to a way to justify never fixing security issues. Adopt a zero tolerance policy for security issues and risks.

5. Going forward, scale up agents not people. The threat is moving at machine speed and you’ll need to respond in kind. Push your security team to automate as much of their detection and response process as possible. Security teams are risk averse and some have been slow to adopt new technologies. Push them to adapt.

This is general advice, Our team is also available as a resource to your security teams for any follow-up or technical questions.

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.

Welcome to this week’s charts. Before we get to them, we have something important to share first. Here’s a16z’s and , with some timely advice we sent yesterday to our portfolio company CEOs, on the current cybersecurity emergency: Patch or Die

Stay safe! Now onto our regular charts.

Attention is All You Need

Charts has wondered about whether Netflix is Having Its Blockbuster moment, and while that may or may not be happening, the streamer’s recent report can only add fuel to the fire:

Netflix is still growing, but topline growth has decelerated to just a shade under ~12%, which is the lowest it’s been, since the last time Netflix was supposedly reaching its peak back in ‘21-’22.

Just to briefly restate the premise: the issue for Netflix is some combination of competition from other streaming options, but even more so, competition from an entirely different medium, i.e. short-form video. If people increasingly lack the attention-span to watch episode-length streamed shows, and prefer the Cocomelon-styled rapid frame-switching of reels, tiktok and shorts, then who is going to be the net-new buyer of the ever-more expensive premium streamer?

It’s a fair question, and apparently, there is no good answer, just yet (which is also perhaps why Netflix is developing shorts of its own, in addition to its foray into live sports).

But, there’s another issue for Netflix, and not just Netflix, but really for all streaming and/or content. There’s apparently an upper limit for how much media we can actually consume, and it seems like we’ve just about hit it:

According to some data from BCG, US weekly media consumption tops out at ~26hrs/week, and has actually been in moderate decline since peak lockdown.

In other words, if you’re competing for attention spans, either as a content-creator or app maker, then you’re increasingly playing a zero-sum game with your competition. Relatedly, a lot of the media growth recently has come from “background” consumption, i.e. listening/watching in your car, at your desk, or otherwise while doing something else. Put another way, in the battle for attention spans, content has resorted to doubling-up with other activities to stretch total consumable hours to something longer than 24hrs in a given day.

The takeaway here is really just a continuation of a theme: So many apps and so much content, and yet so little time.

GPUs, Still Not Obsolete

With all the recent FUD banging around the timeline, it bears reflecting on a tiny slice of some actual hard data.

First, GPU rental rates continue to climb:

Every model of GPU in the series from Ornn Data, including the oldest A100 models, costs more to rent than it did at the beginning of the year.

That’s just not supposed to happen for assets that are supposed to depreciate from the moment they’re launched.

In fact, if you just compare a straight-line 6-year depreciation against the longer-history of available pricing, you get a pretty sizable surplus, once again, for every model:

Prices are anywhere from 9% higher to 143% higher than they would have been, if they had depreciated steadily in-line with the supposed 6-year life of the chip.

To be fair, the older A100s and H100s spent more time below the line than above it, so over the longer haul, pricing for those two, at least, has under-performed straight-line depreciation. But the longer haul isn’t entirely fair either, because April of this year is when things really took off.

Token demand, per OpenRouter (which captures only a small slice of the picture), began its steep upward inflection in the Spring, and after a brief end-of-summer plateau, has once again leapt forward in the recent read:

Token consumption for both open and closed models continues its exponential path, with proprietary models even making a modest share gain last week.

So, if Spring ‘26 and forward is the new normal, then to repeat where we started: the new normal has been awfully good for GPUs (and token demand), even the ones supposedly too old for new tricks.

Now, are these indexes perfect proxies for actual token spending and GPU pricing? No, of course not (for both better and worse). But, they’re certainly not nothing, and the story they’re telling isn’t FUD.

Doing More With The Same

One persistent, but still under-the-radar, observation about all this Capex and reindustrialization more generally, is that it’s still somewhat narrow, especially when it comes to rebuilding the domestic manufacturing base.

Generally speaking, you’d think that if there was a big surge in demand for a thing, followed by a big surge in prices, then the market response would be to increase supply. Increasing supply can come via two levers: more capacity and more utilization (or some combination of both). If the customer demands more, then either build more factories or crank up the ones you’ve got to 11. That’s pretty simple, so far as it goes.

When it comes to the key components of AI Capex, that’s sort of what’s happened, but not really, and not across the board.

• Computer capacity and utilization have both increased, leading to an even bigger leap in production;

• Communications had a big surge in capacity, with a drop in utilization, leading to a big lift in production (and presumably something of a glut in capacity);

• Semis added a lot of capacity, but no additional utilization, meaning all the production increases were driven by the former;

• Finally, the big outlier is electrical equipment, where prices have surged the most, but capacity has actually declined, leading to a modest drop in domestic production (despite an increase in utilization).

It’s a somewhat odd thing, that with the exception of comms, the domestic manufacturing response to a surge in demand (and prices) has mostly been to try and squeeze more water from the stone, rather than bring more supply capacity online—no sector more so than the makers of switchgear, etc. that’s required for the grid-glowup to power those GPUs and the electrification of everything.

Electrical equipment makers have basically said “Oh, there’s a historic backlog on electrical equipment and prices are through the roof? Yeah, we’re gonna sit this one out.”

Unsurprisingly, if domestic production lags behind, then the supply shortfall has to be filled from overseas:

The increase in imports for all four of these Capex-critical categories has far outpaced the increase in domestic production (as measured by their dollar values).

It seems that, for whatever reason, the cost and risk of bringing more domestic production online has simply outweighed the historic onslaught of demand. Or, perhaps, domestic firms simply don’t think they can compete with offshore alternatives. But either way, faced with the largest buildout since the railroads, the industrial base has mostly responded with a “do a bit more (and sometimes a little less) with what we’ve got” approach.

As an aside, and perhaps related to the core capacity problem, it’s not just the factories that are doing more with the same. It’s also the people working at the factories, where (by and large), the extra output is coming from existing workers working more, rather than more workers (on net):

Per data from ADP, weekly hours/worker for the AI buildout sectors have grown substantially—more so than other sectors, and generally more so than headcount itself (which has also grown, but not as much). In other words, rather than throw more bodies at the problem, manufacturers are generally throwing more at the bodies that they’ve already got.

Part of that might be a productivity story, where the same number of workers can produce more than before, but it’s also likely a deeper problem: there’s a shortage of people to do the work. That firms are bidding on a scarce supply of skilled labor is reflected in rising gross pay for manufacturing, and a whopping 5pp wage-growth premium for job-switchers over job-stayers—when job-switchers are getting big relative pay increases, it usually reflects a bidding war.

The doomer cry for an AI Jobs Apocalypse has quieted somewhat since the early days, but it’s a heckuva thing if one of the main bottlenecks to hastening the compute-revolution turns out to be a shortage of people to do the work.

SaaS Prove-It Cyber

Charts has variously observed that contra an undifferentiated-SaaSpocalypse, the market has actually been rather discerning. For software that has real or perceived tailwinds, performance has been pretty strong. For everyone else, it’s been a bumpy ride.

For most of the year, one of the standout software categories has been cyber . . . but over the last couple of weeks, the bloom has come off the rose, at least a little:

Cyber stocks in the $IGV have still had a very good year, but since October they’ve sold off substantially—with the rest of the market, only more so.

Why have cyber stocks sold off? Only Mr. Market can say, but perhaps perceived tailwinds aren’t as compelling as they once were, and actual tailwinds are the new hurdle.

As it turns out, for however much cyber is supposed to benefit from the AI-driven proliferation of cyber risk, actual revenue acceleration is few and far between:

Of the biggest cyber and cyber-adjacent names, revenue growth has only increased a few percentage points (and in some cases, has gone backwards), even as performance has been upwards of ~125%. Palo Alto Networks is the revenue growth outlier, but most of that growth has been acquisition-driven.

Now, just because growth isn’t there yet, doesn’t mean it’s not on the way. Maybe it is, and maybe it isn’t, but in a SaaS-Prove-It regime, Mr. Market may not have the patience it once did.

For what it’s worth, according to the cyber companies themselves, there is some evidence that AI may be helping things along:

These KPIs are non-standardized, and there’s no real longer-term series to infer a trend, but it’s at least a valiant ~5-shot Codex effort to see what Cyber has cooking from an AI front.

>800% agentic traffic in just 9 months . . . that sounds like a lot? ~200% ARR growth for SentinelOne’s Purple AI and Prompt Security (whatever those are) isn’t bad, either, but presumably it’s still too small to move the overall needle by much.

Only time will tell whether AI turns Cyber into a growth machine (or a replacement cycle), but like all the other software cos, Cyber will have to prove-it, soon enough.

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.

We’re excited to announce the fellows who will be part of the inaugural a16z Ops Engineering Fellowship, an 8-week cohort of leaders who are transforming their organizations with AI.

We received thousands of applications from talented people around the world. Today, we’re thrilled to share more about the inaugural cohort of 50 fellows who are leading operations engineering at companies like Ramp, Decagon, Luma AI, Opendoor, and ElevenLabs.

See a full list of the fellows here!

Many sit at the intersection of operations and engineering, scaling operational rigor within and across teams. They’re leading the charge by building the internal tools that consolidate the tooling stack, designing the data and context layers that agents draw on, deploying agentic workflows with human checkpoints, and driving adoption. Folks like Ben Booi and Rushaabh Saliya at Opendoor, Chris Jones at Shopify, Dilan Shah from Cresta, Jordan Shapiro at Handshake, Kevin Chu from ElevenLabs, Saya Iwasaki at Notion, Christina Huang at Cognition, and Keith Gliksman at Tennr.

Some are world-class COOs and chiefs of staff redefining how AI gets used across organizations. They’re deciding which workflows to automate first, the KPIs that prove it’s working, and are leading the teams that put it to work. Folks like Caroline Ingeborn at Luma AI, Gabriella Garcia at Warp, Brandon Gell at Every, Rohan Shah at Ramp, Matt Stehle at Formation Bio, and Meera Shah at Glean.

Others are taking how software gets built to the next level, creating new products and leading the teams that ship them, like Claire Vo at ChatPRD, Seb Goddijn at Yaklabs (ex-Ramp Glass), TK Kong at Concept, Matt Portnov at Decagon, Owen Williams at Stripe, Andrew Qu at Vercel, and Max Nardi at Runlayer.

Congratulations and welcome, a16z Ops Engineering Fellows!

Sign up here to stay in the loop about future fellowships and a16z opportunities.

Stay in the loop

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.

Today we’re announcing that we’re expanding the scope of our new fund beyond crypto startups to the broader category of frontier startups: companies building new technologies and creating markets that may take years to develop. To reflect this, we’re renaming a16z Crypto to a16z Frontier. This is an expansion: we will continue to invest in crypto and related ventures and advocate for sensible crypto regulation.

When we started investing in crypto, it was a world of its own. But as blockchains have become more widely used, the lines have blurred and the infrastructure has gradually receded into the background. Wall Street and fintech companies now regularly launch products built on blockchains. Does that make them crypto companies? The label matters less now. It makes more sense to define companies by what they do rather than by the technologies they use.

Our expansion also reflects an approach that has guided many of our investments for years: backing technologies that look futuristic, well before they reach the mainstream. When I joined a16z in 2013, I made these kinds of long-horizon investments a focus. One theme was new computing devices. The growth of smartphones made cameras, sensors, and chips better and cheaper, enabling new devices like drones, AR/VR headsets, and wearables. That theme led to investments in Oculus and drone startups like Skydio and Zipline.

The other was crypto, which I saw as a new wave built on social media and the open source movement. Social media allowed networks like Bitcoin and Ethereum to gain traction from the outside in, as people formed communities without first needing mainstream acceptance. Open source provided the model for software that anyone could inspect, modify, and build on. Blockchains extended that model from software to services and networks. This theme guided our investment in Coinbase and the creation of our crypto funds, as well as later investments in projects like Solana and Uniswap.

Now we are in the middle of another major tech wave centered on AI. The buildout of AI models and infrastructure will continue, but the tech is already sufficiently advanced to spawn a new generation of products and services. No one knows exactly what that generation will look like, but we’re beginning to catch glimpses. We’re especially interested in the new products and markets these advances will enable, including ones that don’t fit into existing categories.

Foremost is the emergence of native use cases: brand-new activities you couldn’t do before. Early applications of new technologies tend to focus on doing old things better. To date, AI has mostly been used to do existing activities — like coding and other knowledge work — faster, better, and cheaper. But soon native uses will open up entirely new, unexpected experiences. At first, many of these experiments will look strange and unserious, like toys and hobbies, but over time some will grow into mainstream products. The key to telling the difference is to understand their potential rates of improvement, the factors that can bend deceptively marginal growth into an exponential curve.

We’re lucky enough to be living through one of the most consequential curves in human history today. Technological progress is accelerating rapidly, and we’re already starting to see early signals about what a post-AGI world might look like. For one thing, software and increasingly hardware are becoming easier to create. Just as the growth of smartphones fueled an earlier computing cycle, we expect a proliferation of new devices to underpin this one. Some of these will be autonomous machines that can perceive, reason, and act in the physical world. Others will draw on advances in biology and medicine. Still others will fall somewhere, cybernetically, in between.

All these advances will present new ways of building companies. We’ll see more full-stack startups that combine software with hardware and services, as founders take advantage of new capabilities and seek new forms of defensibility. Even the idea of what it means to be a founder may change as we gain more superpowers and as AI agents become increasingly autonomous, able to act on their own. As these capabilities spread, they could change how industries are organized and even give rise to new ones.

AI will also enable entirely new kinds of networks connecting people and intelligent agents that learn, coordinate, produce, and transact. The large-scale networks we use every day were made for humans. Networks built for machines will look very different. Many will start as single-player tools and evolve into multiplayer networks. The potential here — for machine age networks — is enormous, and builders have barely begun to scratch the surface.

Blockchains will play a significant role in this transformation. Crypto has given the internet money rails that are fast, cheap, global, and programmable. Stablecoins are quickly becoming a standard tool for fintechs, established institutions, and others. And, because blockchains are programmable, both centralized and decentralized financial services can be built on top. Crypto can give new networks native economic capabilities, enable machine-to-machine transactions, and create new ways to bootstrap them.

These are examples, not boundaries. The defining companies of past waves were rarely the ones anyone predicted. As always, we expect founders to surprise us and create categories we haven’t imagined yet.

I’ll be leading the fund along with the firm’s founders Marc and Ben, and GPs Eddy Lazzarin and Guy Wuollet. Our operating teams will continue to work with our existing and future portfolio companies on regulatory, go-to-market, talent, marketing, and technical R&D.

We believe some of the most important technology waves are still ahead. Ideas that seem weird or far-fetched today will become indispensable, era-defining products of the future. If you are a founder building what comes next, please reach out.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Not an offer or solicitation. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. Some of the companies mentioned are portfolio companies of a16z. Please see a16z.com/disclosures for more information. A list of investments made by a16z is available at a16z.com/portfolio. 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.

In 1990, Eugene Brooks of Lawrence Livermore told the Supercomputing conference that nothing would withstand the attack of the killer micros. “Micro” referred to microprocessor, meaning a general-purpose chip that did a good job at most types of computation. They were also being sold by the millions per year into desktop PCs, delivering faster and faster performance each year.

Around that same time, competing specialized architectures, such as vector processors, massively parallel processors (MPPs), and Lisp machines started dying off en masse. ETA systems closed in 1989, Multiflow in 1990, Alliant in 1992, Symbolics in 1993, Kendall Square Research and Thinking Machines (not the same one as today) in 1994, and Cray Computer in 1995. Danny Hillis, the founder of Thinking Machines, remarked years later that “…ultimately these things were just going to be boxes of microprocessors.”

Interestingly, these specialized chips didn’t lose because they were superior at unimportant problems. The general-purpose microprocessor was actually outperforming them, not just on a wide spectrum of workloads, but on the very workloads the specialized chips were designed for.

And although the companies that made them didn’t make it, the ideas they brought lived on inside other chips. CPUs absorbed vector processing as SSE and AVX, and massively parallel designs influenced both multi-core CPU and GPU architectures. The pattern at the time was that the specialists came up with good ideas, but the generalist chips absorbed those ideas into a single architecture.

Today, we are in the middle of another heyday for the industry. AI workloads are different enough from traditional software workloads that the industry is in the midst of rethinking what computing should actually look like, with many different new chips and ideas behind them.

There is a massive explosion of new accelerators to serve AI workloads, and GPUs generate more revenue for chip companies than CPUs do (although CPUs are also in their own demand crunch right now). In times like this, it’s interesting to look back on the last hardware boom and see what has changed since last time (and what remains the same), to better predict what will happen next. Is this time different for specialized chips, or will they ultimately be absorbed into general purpose chips like before?

Why general-purpose chips won (for a time)

Why were the “killer micros” beating specialized hardware - at the very workloads the specialists were designed for - back in the 90s? The general-purpose microprocessor was able to leverage a virtuous cycle: generality drove volume, volume drove performance, which in turn drove generality.

1. Generality drives volume

The microprocessor was sold by the tens of millions into desktop PCs, being general-purpose hardware that could run every workload a PC had. The specialist hardware, on the other hand, had much lower volumes because it only ran one.

2. Volume drives performance

Those volumes translated into much lower costs, more buyers, and ultimately more money to fund development for the microprocessor. From 1986 to 2003, Hennessy and Patterson estimate that single-processor performance grew over 50% year over year, most of which was delivered by faster clock speeds. Moore’s law scaling delivered a free lunch during that period - that is, free for anyone who can continually fund leading edge fabs and large chip design teams to take advantage of every new generation. The specialists with their lower volumes could not afford the same cadence, and therefore fell behind. In addition, the generalists were able to absorb some of the best ideas of the specialists once the specialist workloads were better understood (see vector processing becoming a capability of CPUs).

3. Performance drives generality

The faster clock speeds largely enabled by transistor scaling were a rising tide that lifted all boats in the micro camp. The same program moved from a 386 to a 486 would automatically get faster since they shared an instruction set. For programs that required recompiles or new libraries based on new instructions or hardware capabilities, Intel and software vendors could afford to invest heavily in the operating system and compiler side and scale those out to their large pool of users. Specialized hardware, on the other hand, required its own software stacks and it was much harder for those companies to keep up with limited funds and lower scale. Each time the microprocessor got faster and cheaper, applications that previously weren’t feasible were unlocked. One interesting example was PCs with DVD drives that were originally shipped with specialist video decoder cards. Within about two years, these cards went extinct when the Pentium II could decode video in software.

Losing our free lunch

Unfortunately, the “automatic” speed gains from Dennard scaling were not to last forever. In 1974, Robert Dennard at IBM published that a smaller transistor was faster, cheaper, and lower power than its larger equivalent, so the clock speed could increase each generation while power stayed constant.

This matters for the dilemma around specialization, because we can drive performance improvements for workloads in different ways (for example algorithmic improvements, chips getting specialized/better at a specific tasks at the hardware level, and fundamental efficiency gains in process technology can all drive better performance).

The first piece of Dennard scaling to fall was the ability to scale smaller transistors to lower voltages, which was what unlocked the ability to maintain the same power level for more, denser transistors in the same chip area. This happened around 2005, as you can see from the chart below. As a result, smaller transistors couldn’t benefit from proportionally lower voltages. This voltage scaling was the thing that allowed us to scale clock frequency, so you can see frequencies also flattened out at this time. We can see that number of logical cores started rising around this time as well, to compensate for the flattening of clock speeds, however not every application can be efficiently parallelized, so it wasn’t the same “free lunch” as before.

The next thing to fall was cost per transistor, which happened in about 2012. Transistors did continue to shrink, but the cost of the fab and extra process costs grew faster than the transistor count per wafer, so each shrink still gave more transistors per chip but at the same or higher price per transistor.

There were other scaling challenges as well. Going back in time to 1995, experts at the time named the “memory wall” after the observation that memory bandwidth was increasing much more slowly than compute (and it never caught up). David Patterson estimates that NVIDIA GPU compute rose 80x from 2012 to 2022, while memory bandwidth rose 17x. One key factor for this difference is that compute scales with the area of the chip, whereas the bandwidth you can get into a chip scales with its perimeter.

Back in 2012, when transistor costs stalled out, I was a graduate student at Mark Horowitz’s group at Stanford. One of my cofounders, Omid Azizi, worked with Mark on research into sources of inefficiency in general-purpose chips, finding that a 720p encoder running on a general-purpose chip was about 500x less energy efficient than running it on a specialized ASIC. Now that the free lunch of Dennard scaling was over, was it time to return to specialized chips?

The rise of GPUs: revenge of the specialized, or the new general-purpose chip?

The GPU was one of the original specialist survivors of the CPU era. The GeForce 256 in 1999 did a single job very well: rendering triangles fast on a screen for gaming use cases. Unlike other specialized chips like vector processors and Lisp machines, it had a mainstream audience and therefore large volume, with millions shipping into gaming PCs. Rendering is massively parallel, demanding massive numbers of cores, something that 50% YoY improvement in CPU performance couldn’t alone deliver.

Once clock speeds stopped scaling in 2005, chips had to pivot from doing the same thing faster to doing more things at once. GPUs were already designed to do this, with their thousands of simple cores and very high memory bandwidth, a tradeoff over single-threaded speed. This made it a very attractive platform for very large scale workloads, from early machine learning to crypto to high-performance computing. Without the dramatic efficiency and performance that GPUs offer over CPUs for AI workloads, we couldn’t have trained and scaled trillion or even billion parameter LLMs, and would probably still be stuck limited to tiny regression models.

The GPU was able to dramatically scale its performance, but in a different way than its CPU predecessors did in the height of Dennard scaling. Bill Dally from Nvidia broke down where performance gains have come from in the past 10-15 years, at Hot Chips 2023 via the following graphic, showing inference performance gains over time.

For single-GPU inference performance, about 16X of the 1000X speedup game came from simpler number formats (e.g. FP32 to BF16). 12.5X came from new instructions, 2X came from leveraging sparsity (skipping weights in the computation), and only 2.5X came from process improvements from 28nm to 5nm. This meant that the process, which had driven massive performance scaling in the industry over 40 years, contributed less than the other major factors! Note that this analysis doesn’t include the massive performance scaling that also happened simultaneously around large-scale systems of many GPUs (since it is limited to a single GPU’s inference performance).

As AI continues to scale, GPU sales have actually risen past CPU sales. Perhaps at this point we should call GPUs the general purpose accelerator and CPUs the specialized chip!

Why the future will be even more heterogeneous (for a time)

Zooming back to today, inference is growing incredibly fast. According to a 2026 Goldman Sachs report, demand for tokens will increase by 24x from 2026 to 2030, yet available power will only increase by about 2x. We need to run these workloads more efficiently and performantly, because we will likely still have scarcity in the available physical infrastructure.

So what levers do we have left? Dennard scaling is dead, and each lever that produced the GPU’s 1000x gain in 10 years will yield diminishing returns. Going from 32 to 4 bits of precision was hard, but there isn’t that much room below 4 comparatively. Network scaling and algorithms will remain hugely important but also challenging in their own ways. The memory wall is a major limiting factor for inference workloads, especially in the decode phase which is highly memory bandwidth bound.

To help boost performance of inference, the GPU has been following in the footsteps of the CPU, by incorporating inference-specific hardware concepts directly inside the GPU, similarly to how the CPU once incorporated floating point units. Tensor cores were added in 2017, MXFP4 and NVFP4 in 2024 which are predominantly used by inference.

Is it time for specialized chips to make their return, delivering dramatically more efficient performance for specific workloads? There is a Cambrian explosion of new chips focused on inference, but one of the most interesting moves in this direction came from Nvidia in its large-scale deal with Groq. Groq, now known as Nvidia’s LPU, is an SRAM-centric chip which has a fundamentally different architecture from a GPU. GPUs keep their memory far from compute in HBM where they can store a lot of data, whereas near-memory compute chips like Groq, Cerebras, and d-Matrix put the memory directly on die next to the compute, which provides extremely fast memory bandwidth when data fits on the available chip memory. Both designs have advantages depending on the slice of the workload, and NVIDIA is now orchestrating inference across GPUs and LPUs in tandem in order to continue to push inference performance.

Inference is a highly heterogeneous workload, and different phases of inference benefit from different tradeoffs in the underlying hardware. Given the scaling challenges across transistors, memory bandwidth, and network bandwidth, selecting the right chip architecture for each segment of the workload offers large performance improvements, even if the underlying resources are no longer scaling like before.

Orchestrating inference workloads across diverse accelerators is what my company Gimlet is most focused on. By splitting workloads across different types of chips - such as NVIDIA GPUs and LPUs, or other hardware like Google’s TPU, Cerebras wafers, d-Matrix Corsair, Trainium, AMD GPUs, and other accelerators - we can see dramatically better performance per watt than homogeneous stacks alone can achieve.

Although specialized hardware failed in the past, the circumstances are different today. It’s easier than ever to develop optimized code for new accelerator types using agents. We have significantly more expertise and infrastructure for building new chips than we did decades prior (most chip companies don’t need to handle their own fabrication, for example). We are severely power limited with a workload (inference) that continues to explode in scale, with many of the low hanging fruit of performance already picked off the tree.

Splitting inference workloads across different types of chips is not an easy feat, and as much of the complexity is in the physical infrastructure as it is in the software layer (but on the plus side, physical infrastructure complexity might be a more durable “moat” than software can be). It’s important to select and deploy hardware in a very modular, configurable way, with a software stack that can leverage all available hardware and not leave stranded capacity earmarked for particular tasks.

I am bullish on the benefits of bringing together diverse silicon, because in this era of inference, delivering dramatically more or faster tokens with limited power is such a massive advantage that it’s worth the complexity. This phase will last years, although it may not last forever.

The pendulum of specialization

Tsugio Makimoto, who ran Hitachi’s semiconductor business, noted that general versus specialized parts are a pendulum that swings cyclically from side to side. While time to market, cost effectiveness, and operational efficiency benefit standardized parts, differentiation and supply-demand imbalance cause the pendulum to swing to the other side.

So where is the pendulum in 2026? In spite of the move toward heterogeneity, there are still tailwinds for the GPU. Fab and HBM allocations are very scarce and difficult for new players to acquire. AI accelerators (GPU and otherwise) need to be deployed for multiple years to be economical, while the workload continues to evolve. And GPUs will continue to absorb new inference features as they prove their worth.

At the same time, the specialists have many factors in their favor. One of the scarcest resources is power and specifically performance per watt of AI workloads, and specialized chips can maximize performance for a given workload. It’s getting easier to develop software for new hardware, as agents are shrinking both the time it takes to write kernels or even design new chips. We expect to see a continued increase of specialized chips to deliver better performance per watt for inference, especially in conjunction with GPUs for the portions of the workload that are not easily offloaded to new accelerators.

Although I expect the pendulum will likely eventually swing back to generalized architectures, the time of specialization is upon us and a very exciting time to work in infrastructure. Each era of computing (desktop, cloud, mobile, now AI) brings a new set of dominant workloads, and a need to rethink what computing actually looks like. Even after a pendulum swing, the novel, specialized hardware of today has the potential to become the mainstream hardware of tomorrow (and even if it doesn’t, its contributions can still continue to live on in future generalist chips).

Thanks for reading, and if you are interested in these topics, send us a note at hello@gimletlabs.ai! We are hiring for all roles and scaling out our inference cloud to deliver radically faster tokens to more users in the next few months.

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.

We’re excited to announce that Orchestra Bio has been acquired by OpenAI. To read more about the acquisition, read and ’s post about it here. For more general information on what OpenAI is up to in the Health space, check out our podcast with them here.

It took me eight months to build my first FPV drone. I procured the parts in January, received them in February, remembered I bought them in June, assembled and lit the drone on fire in July, re-assembled and crashed it into my neighbor’s roof in August, and then, and only then, at the behest of my wife, downloaded a FPV flight simulator.

A super-intelligent AI surely could have done better.

While far from superhuman today, the frontier models can now generate CAD, design custom PCBs, source COTS parts, and describe how those parts are to be assembled. And they’re only getting better. Hardware design benchmarks like CAD Arena, EE Bench, and my own LuxoBench show steady improvements; and more generally, you’d have to be crazy to bet against the Bitter Lesson and say hardware design will forever stay beyond the frontier.

Superhuman hardware engineers are coming soon, perhaps very soon. But what will they look like? Will superhuman hardware engineers surpass all humans in one fell swoop, or only in specific domains/disciplines? And what will hardware engineering look like before, during, and after the crossover?

My answer to each of these questions starts with one simple observation: that the interface between the digital and physical world is lossy. This makes it hard to “write” changes from the digital world of bits into the physical world of atoms, and hard to “read” measurements of the physical world back into a computer. This lossiness will shape when and how future superhuman hardware engineers will emerge — and means that even once the models can consistently one-shot electromechanical designs, they will still have a lot more work to do.

Write access: from digital designs to assembled atoms

When OpenAI released GPT-6 Astra, the online world erupted with demonstrations of its 3D design prowess. Tom Krcha used Astra to model a steam train, Thomas Ricouard made a walk-through of a photorealistic house, Anshu Chimala made a 3D game, and Sharif Shameem made the Palace of Fine Arts in San Francisco.

These demonstrations got hardware engineers wondering if the models could also generate engineering-grade CAD. Nim Cheema, co-founder of Dessn AI, asked Astra to create a DJ controller. Roberto Nickson made the top models design, engineer, and 3D print bridges. And I launched luxobench.com, where I asked OpenAI’s Astra, Anthropic’s Fable 5.1, and SpaceXSI’s Grok 4.6 to help me build an interactive desktop lamp.

We quickly found that — unfortunately — the models by themselves are still not quite ready for primetime hardware engineering.

There are many levels to building real-world hardware, and clearing each one requires different, and often entirely independent, skills.

At the 101 level, you have to decompose the product down into components, identify available commercial-off-the-shelf (“COTS”) parts, ensure their mutual compatibility, design the mechanical and electrical interconnections between them, ensure the final design is buildable, and then actually procure parts from online stores.

Even this step is hard enough, and can set back not only backyard hobbyists like yours truly, but also the most sophisticated engineering teams in the world. In Walter Isaacson’s biography of Elon Musk, he relays a story of a time in 2006 when a Falcon 1 hot fire was delayed due to a capacitor — a tiny electrical component that was only in stock at a store roughly 6,000 miles away from their launch pad in the middle of the Pacific Ocean. Getting COTS parts where you need them, when you need them, is hard enough.

At the 201 level, however, things get even harder: you now have to make custom parts, which means navigating the complex translation step from digital design tools to physical machines. Take, for example, CNC machining. After a mechanical engineer designs a bracket, they can’t just hit “print” and have the part materialize before them. Instead, they work with a machinist, who translates their CAD file into CAM (“Computer-Aided Manufacturing”) software, which writes the instructions that will actually drive the CNC machine.

In practice, this is an artisanal process that requires a master machinist to choose the right stock, workholding, tools, and cutting strategy to guide the tools as they cut through the block of titanium. This causes variability in the manufacturing process, and variability leads to errors in translation.

At the 301 level, you take those simple parts and start to bring them together into a complex electromechanical assembly. As you add more parts together, you add complexity and “stack” tolerances up — meaning, every part might individually meet its drawing, but still fail to fit into the assembly due to compounded and correlated errors. Those other parts are also likely built by different people on other teams in the company, requiring requirements, schedule, and risks to be communicated across the organization to ensure that the parts can interface nicely together and are delivered when they are needed, but not before and not after.

At the 401 level, you realize that your design needs not only COTS parts and custom-designed parts, but also either inputs or particularly complex assemblies from other external vendors. The hardware world loves talking about vertical integration, but basically everyone relies on vendors — it almost never makes sense to build your own chip fab or copper mine. Even SpaceX, the champion of vertical integration, has 591 current employees with a “supply chain” or related titles, and is actively hiring 128 more.1 And once you have outside vendors involved in building custom parts, good luck!

And then, finally, at the 501 level, you have to do all of the above at massive scale. It is a fundamentally different task to build 1 of something and to build 1,000,000 of something: you need to worry far more about part and process variability, you need much more financing (much further in advance) for things like machines and inventory, you need to train the people (or program the machines) who will assemble, integrate, and test the product for you, and you quickly hit the limits of your vendors.

Each of these “writing” problems, from the 101 to 501 level, are all fundamentally solvable but need to be solved individually. As we’ll discuss below, superhuman AI might be great at pure design, meaning at defining the platonically ideal part for a particular job, but struggle with “design for manufacturing,” or finding a point design that meets the requirements while remaining easy to build/procure and inexpensive. The former is a relatively tractable task that can largely stay within a machine; the latter requires significantly more interaction with the real world, and should therefore prove more difficult for the models to master.

Read access: observing and measuring the real world

One of my friends buys 45 million chickens per week. But despite his efforts to buy the same spec from the same suppliers and run everything through the same process with the same machines, his week-to-week variance in saleable food products after processing (i.e., chicken breasts you can buy at the supermarket) has historically been as high as 20%.

Improving that number over time has taken him a herculean effort: extensive measurement and analysis of every step of his supply chain, plenty of old-fashioned vendor cajoling, and a very sophisticated data operation behind both. In one memorable case, he found a malicious supplier deep in his supply chain that was injecting its chickens with water before shipping them. All of the intake measurements for weight would pass, but that water was, of course, not meat, so he was left with less usable product than he intended to buy.

Everyone who has built hardware at scale, or otherwise managed a supply chain, has their own version of this story. It is exceptionally hard to know precisely what is going on in the real, physical world.

First, the real world is non-uniform, and constantly changing.

Physical objects have defects, change their properties based on environmental variables (e.g., temperature), and are subject to wear and tear. Even extremely well-tuned processes and measurement tools stray over time. In practice, this means that physical systems like supply chains and manufacturing lines constantly break, and must be designed with variance in mind — and you can only know when something breaks or has strayed if you invest significantly in measuring the inputs and outputs of the system.

Second, real-world data is sparse. You can’t get all the data all the time. This is a fact I’m acutely aware of, having worked on a number of satellite programs: our thermal team would have loved to put thermistors more or less everywhere on the satellite, but the software, structures, and production teams unfortunately also got a say.

And then when something goes wrong, you often realize that diagnosing the issue would have been significantly easier if you had collected some obscure piece of data — a picture of an electronics board at an intermediate stage of assembly, an extra sensor at a new location, and so on.

Third, real-world data is often expensive to collect. High-performance domains (e.g., F1 racing) and high-reliability domains (e.g., human spaceflight) often need hardware measurements to be extremely precise, and precision comes with cost in dollars and in time because precise measurements need even more precise machines. For instance, the CMMs (coordinate measuring machines) that measure flatness, perpendicularity, dimensions, and so on for machined parts are not cheap, and often painfully slow to procure and to operate.

Fourth, the real world has inherent latency. Physical processes take time — you can only thermally cycle an electronics board so quickly; mixing volatile chemical solutions or solid rocket propellant must be done with care; and mechanical fatigue testing takes time to play out. These processes can be accelerated, but only so far.

My favorite machine at Astranis, the Highly Accelerated Life Testing (“HALT”) Machine, was colloquially called the “shake and bake” — it combined vibration with rapid thermal cycling to stress parts as fast as possible. But a HALT campaign as a whole is hard to compress: no matter how fast you design new iterations of the board in between tests, you still have to run the whole test again once per cycle.

And, fifth, some information about the physical world is qualitative — and gathering such data is often the most incompressible step of real-world hardware design. Do your users like your product? How comfortable is the device to hold? How easy is it to build a component using your tool? Reckoning with qualitative, unquantifiable data is a part of any real-world build, and measuring that data to bring it back into a digital learning cycle is exceptionally difficult.

Data limitations like these are massive drivers in the spikiness of AI capability development today. The more quantifiable and machine-readable the results of a test, the faster the models can climb the hill. The reasons above have limited the frontier models to date — but how long those challenges can hold back the unstoppable force of the Bitter Lesson remains to be seen.

Hardware development in an AGI world

Hardware engineering is iterative: you design, build, test, and iterate, and each loop through this process takes time.

The design step is digital, and therefore very model-legible. Just as human hardware engineers use computer software to design the part or the assembly, superhuman AI can do the same — and given the growing power of computer use and the growing capability of AI in similar all-digital tasks, we should expect AI models to become superhuman hardware designers in the near future.

But designing hardware is only the very beginning of the hardware engineering process. Once you design the part, you still have to build, test, and iterate — and each subsequent step requires more physical-world interaction than the one before. Becoming a superhuman hardware engineer therefore requires much more than superhuman hardware design, and the frontier will likely be split across many distinct lines of effort.

First, hardware design needs better simulation tools. Every engineer is rooting for Navier, Inviscid, and Lanyon, because better simulation tools would allow hardware engineers to iterate at software speed. Engineers love that better sim tools may save them trips to testing facilities like cyclotrons, radiation sources, thermal-vacuum chambers, and wind tunnels, and executives love that they can save the capex needed to build such facilities in house.

It may prove difficult to convince hardware companies to bet on simulation alone, but as models get demonstrably better, engineers will start trusting the sims more and more, just as sims are getting faster and more efficient per unit of compute. Atoms are likely to stay in the iteration loop forever, but with better sims, atomic assembly can be confirmatory, not experimental.

Second, we expect demand to “shift left,” i.e., move closer to raw materials. As the models distribute superhuman hardware design skills more broadly, the world will need more wire harnesses from Senra Systems, control panels from Podium Automation, machined parts from Hadrian, and electronic components from Cofactr. The world will also need more software licenses, like Sift for telemetry, Instrumental for inspection, and Revel for testing — and, for that matter, more copper from Mariana Minerals.

Armed with superhuman intelligence and new tools, we also should expect to see more companies bring more of their bill of materials in house. Take, for instance, what Diode has done for PCBs.

In its current form, Diode is a way for professional electrical engineers to develop custom printed circuit boards faster and more efficiently. But what happens when everyone has the same skills as a professional electrical engineer? We should expect to see much more demand from a much broader set of users — all of whom will need to translate their designs into the real world, which means navigating supply chains, ensuring designs can be cheaply and accurately built, and more.

Third, the AI labs need to acquire “process knowledge.” You hear the legendary stories of a machinist who can tell how an operation is going just by listening to their CNC machine, or how PhDs at TSMC have spent decades perfecting everything in the semiconductor manufacturing process, all the way down to the polishing pads.

But end-to-end data that could allow models to approximate that kind of “process knowledge” is hard to come by. In Eric Jang’s great piece, “Smooth Exponentials for Robotics,” he estimates that robotics labs in 2026 are training their robots using, roughly — millions of hours of egocentric data, hundreds of thousands of hours of UMI gripper data, tens of thousands of hours of teleoperated robot data, and just hours of “high-quality, hardware-specific demonstrations for high-performance demos.” This funnel is unsurprising, and I expect it to persist even after the superhuman crossover.

The best answer to overcoming this data scarcity, of course, is to find creative ways to get it yourself — Periodic Labs is a great example. The company announced recently that they have set up an autonomous lab to run end-to-end physical experiments on things like superconductors, magnets, and semiconductor materials. This is the platonic ideal of AI-for-hardware data: design, work instructions, test execution, end-result measurement, and iteration all in one coherent loop.

But even the best test setup can only go so far. I suspect that the most incompressible step of the design process — meaning, the step that even a truly superhuman hardware engineer would struggle to expedite — is gathering feedback from the end users of hardware products. A superhuman hardware engineer could design one million new iterations of the Apple Watch and get them manufactured for cheap… but then how can it know what customers actually want?

Superhuman hardware engineers are coming fast

I expect the frontier models to quickly solve hardware design, which will shift the frontier to the remaining steps of hardware engineering (DFM, testing, iteration) that require greater interaction with the real, physical world. Advancing that frontier will mean solving a whole new set of challenges: improving the interface between the digital and physical worlds; finding ways to more quickly discover customer preferences; and building data pipelines that can approximate the process knowledge currently only held by master technicians.

The crossover between human and AI hardware engineering may prove uneven, as laid out above. But once here, its impact should prove immense: humanity will be able to directly use digital AI tools to bring about real, physical, material abundance — or, at the very least, drones for hobbyists that don’t light on fire.

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.

The human hand may be the most economically important machine ever built.

Much of the world’s physical economy—from manufacturing and electronics assembly to logistics and food production—still depends on people using their hands. Industrial robots have transformed highly structured production, where fixtures hold every part in place and robots repeat the same motion quickly and precisely.

But enormous categories of physical work remain hard to automate, because the real world is variable. Parts move. Materials deform. Objects slip. Connectors don’t align perfectly.

That creates an enormous economic opportunity. If robots can reliably perform the dexterous tasks people do every day, automation can expand beyond repetitive, engineered processes into a much larger share of physical work. Labor is already a constraint across many of these industries, and the value of a machine that combines human versatility with machine speed, reliability and uptime is difficult to overstate.

We believe solving dexterous manipulation is one of the largest opportunities in robotics. It’s why we’re excited to lead the Series A in Eka Robotics.

Intelligence Meets Physics

Modern AI is making extraordinary progress in helping robots understand what to do. Vision models can recognize objects and estimate their position. Language models can interpret instructions and reason about complex topics. World models can increasingly reason about scenes and plan actions.

But knowing what should happen isn’t the same as physically making it happen.

Consider inserting a connector. Vision can tell a robot where the socket is. But once contact begins, the important information changes. Has one edge caught? Is the connector beginning to bind? Should the robot push harder, rotate slightly, or pull back?

Humans answer these questions almost effortlessly through our sense of touch.

The human hand has roughly two dozen degrees of freedom, but mechanics are only part of what makes it remarkable. Our fingertips and palms contain thousands of specialized sensory receptors. Some detect sustained pressure and shape; others detect light touch and the first signs of slip, high-frequency vibration, or the stretching of our skin. Combined with proprioception from our muscles and joints, they give the brain a remarkably rich, continuous picture of what is happening at the point of contact.

That is why you can pick up an egg without crushing it, tighten your grip on a glass the instant it starts to slip, or find the thread of a screw you can barely see. We don’t calculate every movement in advance. We continuously sense the consequences of our actions and adjust.

But we can’t practically rebuild this biological sensing system receptor by receptor in a robot hand. Thousands of specialized sensors would add enormous complexity, cost, wiring, compute and potential points of failure. The challenge is therefore not to mechanically reproduce the human hand. It is to give machines enough physical information—and intelligence—to achieve the same outcome with a much simpler sensing and actuation system.

Eka has built a Vision-Force-Action model that combines visual perception with force feedback from its robotic hands. Rather than treating the model, sensing, controller and hardware as independent systems, Eka designs them together so the robot can use physical contact itself as information.

That changes the nature of manipulation. A grasp that starts to slip can be corrected. Contact with a surface can help locate an object. Resistance during insertion can change the robot’s motion rather than simply trigger an error.

The goal is to preserve performance as conditions change without requiring an engineer to retune the robot for every material, object or fit.

That ability to recover from variation is what can move robotics from controlled automation toward general physical work.

Learning Through Experience

Eka also takes a different approach to how robots acquire physical skill.

Much of modern robotics learns from human demonstrations, whereby the robot learns its “policy” for how to behave in a given situation. This has produced impressive progress so far, and we don’t want to be unappreciative, but physical demonstrations have two inherent problems. First, they take up human time; so you can’t practically show a human demonstrating one million different slight variations on picking up an egg. Second, it’s hard for humans to demonstrate all of the ways you can fail at picking up an egg, in a way that conveys useful information.

Eka trains its manipulation policies through reinforcement learning in physics simulation. Virtual robots can practice millions of variations in object position, friction and contact before a policy ever reaches physical hardware. They can experience failure modes humans rarely demonstrate and learn how to recover from them.

If this approach works at scale, the implication is profound: robot experience can increasingly scale with compute rather than human labor.

The historical problem is that simulation isn’t reality. Contact dynamics are notoriously difficult to model. Small errors in friction, compliance or actuator behavior can turn a successful simulated policy into a useless real-world one. Closing this sim-to-real gap has been one of robotics’ hardest problems.

It is also the problem Eka’s founders are unusually well suited to solve, and they have been demonstrating that the real-to-sim gap is falling:

Pulkit Agrawal is one of the leading researchers in robot learning and a pioneer in computational sensorimotor learning—essentially, how machines develop physical intuition through experience. His work spans reinforcement learning, force-aware control, dexterous manipulation and moving policies from simulation into the real world through his Improbable AI Lab at MIT CSAIL.

Tuomas Haarnoja co-created Soft Actor-Critic at Berkeley, one of the foundational algorithms in modern reinforcement learning, and later worked at Google DeepMind training complex humanoid behaviors in simulation and deploying them on real machines.

They have spent much of the last decade attacking the same fundamental question from complementary directions:

How can machines learn sophisticated physical behavior through experience—and have that intelligence survive contact with reality?

Performance Changes the Economics

Ultimately, none of this matters if the robot isn’t productive. Industrial automation has a brutally simple scoreboard: units per hour, error rates, uptime and cost per unit.

Eka reports up to 1,500 picks per hour in laboratory induction and bin-picking tasks, two and a half times faster than humans. More importantly, the underlying system has demonstrated manipulation across fragile objects, slippery industrial parts and contact-rich insertion tasks. Sustaining that performance through full production shifts is the next test.

If these capabilities transfer reliably across production tasks, the economics of automation change.

Today, many automation projects make sense only when the volume of a task justifies the engineering required to constrain the environment and tune the robot. Every new object, material or process can create another integration project.

A robot that adapts to those variations changes the equation. Manufacturers can introduce new products without starting the learning process from scratch, and tasks that were never worth a dedicated automation project become addressable.

The next major frontier in robotics isn’t simply machines that see and understand the world. It’s machines that can interact with it, adapt to it and perform reliably at economically meaningful speed. If dexterity becomes a general capability rather than something engineers have to painstakingly program for every task, the boundary of economically automatable work expands dramatically.

We believe Eka is building that capability.

We couldn’t be more excited to lead Eka Robotics’ Series A and partner with Pulkit, Tuomas and the entire Eka team.

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.

Agents Denied

Agentic commerce is a thing. Muse and others advanced the timeline to “agentic commerce is here right now—it was nice knowing you, old-school ecomm” but then more recently, some big etailers (like Amazon) responded, “not so fast—agent shoppers are officially instablocked.”

Now, the battle between what agents and bots can do vs. what the rest of the internet will allow them to do has been going on for some time, but the recent escalation does appear to show up in the data:

According to data from Cloudflare, the share of 403 “Forbidden” responses received by agents and bots has fattened considerably since the summer (and is ~double what it once was).

Whether that’s a response specifically to “agentic commerce” is a separate question, but presumably that’s at least some part of what’s going on. Etailers understandably don’t want to be disintermediated, and not only lose their customer relationships, but also lose those customer eyeballs that are a substantial source of advertising revenue for the large ecomm platforms.

There are likely other reasons why the Amazons of the world are apprehensive about agentic commerce, beyond simply eroding their platform advantages. First among those other concerns are likely fraud and returns, both of which have been a rising plague for ecomm more broadly:

According to data from the NRF, return rates and the total cost of returns have basically doubled and tripled, respectively, since 2019. Likewise, reported fraud is pervasive, with the most common concern being “dispute abuse,” i.e. customers buying something and then later disputing the charge. If the card issuer takes the customer’s side, the merchant is SOL.

It’s not hard to understand why etailers would be concerned that agentic commerce could make a worsening problem even worse. “Oh, my agent bought that personalized varsity jacket with my name in sequins, not me, I’d never do that” or “I told my agent to buy socks at the best price, but I didn’t expect it would find the ‘10,000 pairs of argyles, wholesale special,’ so please send a return label and a forklift.”

OK, so fraud, returns, customer relationships, and ad monetization are all throwing sand in the gears of the agentic commerce wave, but presumably those are solvable problems with better tech and commercial ingenuity.

There’s another problem, however, that may be holding back the bot-buying-spree: it’s possible that consumers don’t actually want their bots to do their shopping for them (yet).

As of now (according to some data, at least), consumers seem broadly open to the idea of AI helping them find what they’re looking for, and especially helping them find the best price, but when it comes to actually making the purchase, consumers appear to be somewhat resistant:

~25% of respondents said they’d be comfortable letting AI buy them a variety of consumer goods (except for luxury goods), but 42% said they would not trust AI at all to make purchases on their behalf. That’s less than half, but it’s still more than one would expect.

It would seem, that for now, consumers—like the etailers—are also worried about mistaken purchases, annoying returns, and chargeback disputes. That, and it may be the case that people just like to do their own shopping, at least when it comes to the final purchase decision.

Along those lines, there is some clickstream data from Similarweb (albeit a little stale) that seems to bear that out: AI is good for a consult, but buying is (for now) a human affair. There appears to be an increasing share of “shopping journeys” that visit AI before hitting the actual checkout page, but really the vast majority of AI shopping sessions appear to be in the “middle phase” of the journey, i.e. between discovery and purchase:

It’s fair to take this data with a grain of salt (because that is the nature of clickstream data), but the implications are otherwise pretty clear: for now, consumers seem to be using AI predominantly to do their homework on products they’ve already discovered, rather than either (a) discovering them with AI or (b) using AI to make the final purchase decision.

That makes the case for agentic commerce a bit less straightforward, but it especially makes the question of AI advertising attribution even less straightforward. If someone sees a recommendation in their AI-led middle shopping journey, but then makes the final purchase decision in a different session, then who will know what moved the consumer spirit to which outcome and when?

Again, the data appears to end in June, so whether that’s still true, let alone will remain true, is an open question, but either way, the relationship between shopper and bot is more complicated than it seems.

Super-Pareto TikTok Shop

While on the subject of ecomm, and potential disruption (or not) from agentic commerce, it turns out that social commerce on TikTok Shop has an incredible right-tail skew, for both the social part and the commerce part.

According to data from Marketplace Pulse, 8.4% of TikTok Shop sellers generate 1,000+ videos and capture ~60% of overall sales. Make a lot more videos, sell a lot more stuff.

But the distribution of sales to sellers is actually even more extreme.

Again, according to Marketplace Pulse, ~60% of sales are driven by the top 1% of sellers, and ~93% of sales are driven by the top 10%.

It’s hard to know what to make of this really, other than simply to marvel at the persistence of (super) pareto distributions, but Charts will try to, anyway.

• On the one hand, it seems that making a lot a lot of videos is a good way to drive sales, which implies there’s something evidently “personal” about the shopping experience that would be hard for an agent to replicate (although perhaps if Muse live-streamed its journey, that could change).

• On the other hand, if agentic commerce really does take off, then what would it mean for this small, but mighty cohort of prolific content creators who drive sales through a medium that has no relevance to an agent whatsoever?

Of course, agentic commerce and social commerce do not have to be mutually exclusive. Retail is really big, and there’s plenty of room for more than one model. But, it does seem broadly true that the way we shop now is optimized for a human interface to a degree that “just add agents” may be underestimating.

Cure To High Prices Is High Prices

Charts has no intention of weighing in one way or another on the “inflation: will it, won’t it” debate, but it will take a moment to reflect on the old saying that “the cure to high prices is high prices,” because it certainly seems worthy of some consideration.

The basic premise is straightforward: if prices get high enough, they suppress demand—either for that thing, or some other thing(s) that demand can no longer afford because demand is paying so much for the now more expensive thing. As everyone knows, when demand falls, prices tend to follow . . . so if high prices eventually suppress demand, and suppressed demand leads to falling prices, then the cure to high prices is . . . high prices.

Simple enough.

In this case, most everyone knows by now that AI capex has led to a surge in demand for chips and electrical components that is causing the price of those components to rise substantially:

According to the BLS, the prices paid for semis etc., by manufacturers of computers and computer-like things have skyrocketed over 25% YoY.

Now, those high prices do not appear to be dampening hyperscaler demand for chips—if anything, capex forecasts have gone higher, in part because the components have gotten more expensive.

But that doesn’t mean high prices are having no effect on demand. Hyperscalers may still be on a chip-buying spree, but everyone else appears to be saving their hardware upgrades for another day:

Just taking PCs as an example, according to IDC, PC shipments have dropped substantially since the beginning of the year.

It’s not just PCs that are facing some demand headwinds, but iPhones, as well, all of which have gotten more expensive, partly as a result of the AI buildout. Now, is it definitely the case that high prices are the reason that fewer PCs and iPhones are shipping? No, but it seems like a pretty good guess. And if that’s true, then the old adage appears to be working its way through the system (and will perhaps be reflected in softer prices, soon enough).

Consider, perhaps, another example of high prices being the cure to high prices.

Many, including the State of Markets, have observed the rising prominence of debt, when it comes to financing the buildout. Increased demand for borrowing (among other things) has, unsurprisingly, led to rising prices on that debt:

Rates for bonds linked to the tech sector have meaningfully diverged from the broader universe of corporate issuance, beginning roughly in September of ‘25, when capex borrowing began to take off.

More demand generally means higher prices, so it all checks out. And if higher prices eventually mean less demand, well, that too also appears to check out:

After growing substantially throughout the first half of the year, AI-linked debt issuance has fallen off since its June peak, with September issuance checking in at barely half of what was issued in August (which was already less than what was issued in July).

Of course, there are other reasons beyond supply and demand that would cause debt pricing to rise and fall, but supply and demand surely matter. And in this case, it appears as though the high price of borrowing has indeed put a damper on the inclination to borrow.

If demand falls, does that mean the cost of borrowing will fall in turn, and score another win for the old saying? We don’t know yet, but we’ll probably know soon enough.

Zoomers Living Large?

How’s this for a narrative violation?

According to some data and analysis from PNC Bank, Zoomers are actually doing pretty well, from a financial standpoint. Not only are they doing well, but Gen Z is doing considerably better in 2025 than Millennials were doing 10 years ago (when Millennials were Zoomer-aged):

Zoomers enjoy real incomes that are 20-40% higher than what their Millennial elders suffered at every part of the income distribution. Point-in-time generational comps are imperfect, but it’s still not the result most would expect to find.

It’s also not just higher incomes where Zoomers are winning. There’s also data that suggests Zoomers face lower unemployment, have higher net worths, more liquid assets, and bigger retirement accounts, then the previous generation did at an equivalent stage of life.

Indeed, with all that extra cheddar, it would appear as though Zoomers are living relatively large:

According to PNC’s internal data, Zoomers spend more of their income than Millennials and leave less in the rainy-day fund, as well.

In terms of what Zoomers are spending their money on, they allocate more to retirement, shelter, and transportation, and substantially less to education and energy:

At the category level, Zoomers also allocate relatively more to gambling and devices, but relatively less to clothing and sustenance.

Save more for retirement (but keep less in the rainy-day fund), while also hitting the tables and the money line, but sparing the booze. It’s an odd combination, but it’s Zoomer money, and they can spend it how they please.

Well, perhaps, “please” is the wrong word, because whatever their material advantages over generations prior, the Zoomer mood is decidedly dour:

Since 2017 (with a partial reprieve during pandemania), the traditional “sentiment” advantage that 18-34 year-olds typically enjoy relative to the broader population has all but evaporated. The Zoomers are not as cheerful as young folks of yore.

Now, consumer confidence is a very noisy series that has lost even more signal since the pandemic, likely because of falling response rates (among other things), but even if it’s noisier, it’s strange (but possible) that the noise would be isolated to 18-34 year-olds. In other words, even if the headline “confidence” figure is less reliable, one would think the ratios within ought to stay relatively stable . . . and in this case, it appears they haven’t: Zoomers are no more confident than anyone else, which makes them unlike all the young people of generations prior (at least through the ‘70s).

Taking this at face value, it would appear that the data does validate the vibe . . . but however bad the vibes may be, it’s hard to blame finances. That doesn’t mean Zoomers are without other reasons to feel relatively less optimistic, but certainly when it comes to making and spending money, they seem to be doing just fine.

Why so serious, Young Zoomer?

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.

To hear more from Terry, watch the companion a16z Show podcast here and check-out the open roles at VGS where Terry serves as Chief Strategy Officer here.

We’re also excited to announce Securing America’s Infrastructure Foundation today. SAIF is a 501(c)(3) nonprofit, built to deliver AI and cybersecurity tools, hands-on support, and training to public institutions across America that otherwise lack access to AI tools. View it’s full mission here.

Agents such as Instinct, OAI’s Dots and Meta’s Muse are bringing personal assistants into text messages and familiar apps. The promise is no longer just better answers; it is getting things done. If my agent can organize my kids’ soccer schedule, it can also order their cleats. If it can plan a trip, it can book the hotel. An assistant that understands my life becomes the starting point for much of my spending.

But completing those tasks requires something intelligence alone cannot provide: permission to pay. If your agent doesn’t have your credit card number, any recommendation from an agent becomes a “to do” list for me. With payment credentials in hand, the agent can complete the task for me.

So, of course, there is a scramble underway for what products and companies are going to help agents pay: enter the Wallet Wars.

Below the waterline, as wallets get connected to agents, there are two battles going on: one is over who can help my agent pay, and the other (more consequential one) is over who lets it transact as me. These seem like the same question but they are not! And the difference between the two is important, both for the user experience of “personal payment flows” (think loyalty programs) and for the economics of payments.

This is not a new challenge

In the early years of the internet, people were skeptical that anyone would ever buy something online. Giving your card number to a website seemed risky and unfamiliar. But online shopping was enough of a draw that people learned to trust the process. If history is a guide, just like last time, people will have no problem trusting agents to make payments on their behalf.

Part of what makes an online payment seem scary is that, to the user, the 16 digit card number represents both the identity of the paying account and the authority to initiate a payment. In reality, we’ve worked for decades on the underlying mechanics to make payments easy, safe, and secure, by effectively separating these two concerns. In the industry we call this “payment tokenization”.

Whenever you trust someone to save your card information (whether it be a merchant, a service, a wallet, or something else), they often also create a payment token. In addition to making the whole process more reliable and trustworthy, they also do nice things like, when your card expires, your Netflix subscription doesn’t have to pause - the actors behind the scenes coordinate, “you’re probably still you”, and keep your shows going. Payment tokenization is the site of many “inside-industry” power struggles for how you pay, that the user never sees - and I won’t inflict those stories on you, but payments professionals are very used to these kinds of standoffs.

Now we have these things called agents, which are a further evolutionary step in separating “the identity of the person / the account that’s going to pay” from “the authority to initiate a payment.” And, while it presents some new and interesting challenges which we’ll go through here, we’re not starting from zero - we’re building on years of precedent of how internet payments already work.

The war to connect

The fastest way to give an agent spending power is to connect a wallet that already holds the consumer’s credentials. Stripe’s Link (300M users), Shopify’s Shop Pay (250M users) and PayPal (400M+ users) bring existing payment relationships to that contest, rather than asking consumers and merchants to start over.

For agent developers, the questions are straightforward: How many of my users already have a payment method stored with you? How many of the merchants support this wallet? Stripe and Shopify have moved quickly, powering payments for Muse and Instinct. But speed of integration is only one advantage. PayPal brings an established network of consumers and businesses. Perhaps agents can give PayPal a second act because the winning wallet needs not the prettiest checkout.

The war for reach

Because no wallet’s checkout button is accepted everywhere, wallets offer alternative paths to reach merchants. For example, Link gives an agent a single-use virtual card to make purchases on behalf of the customer. In this flow, a merchant does not need to integrate Link; instead the agent can present its own virtual card during checkout.

While this expands its merchant acceptance, it doesn’t make the purchase equivalent to using the original card. Instead, two transactions are occurring. One between me and Stripe Link (using my card in the wallet) and one between Stripe Link and the merchant.

This can lead to some sub-optimal experiences. For example, imagine your agent buys something from Walmart. Walmart sees an opportunity to strengthen their customer relationship, and tells you: “We’d love to sign you up for Walmart+!” But if you try to accept their offer, what happens? The whole experience breaks, because the credit cards don’t match.

Stripe is building the AOL of agentic commerce

AOL packaged a messy online world into a curated experience; rather than read the news on an unknown website, AOL fetched the news and brought it to you, inside AOL.

For agentic payments, Stripe is building a similar walled garden where transactions occurring inside Stripe’s ecosystem never need to exchange credentials. If your agent shops at a Stripe merchant, Stripe can already charge the card on file without an agent touching it.

When transacting at non-Stripe merchants, Link has been the most aggressive wallet to use its own virtual cards for payment. This means that Stripe is actually buying the item on-your-behalf, rather than the agent acting as you. The good news is that the transaction succeeds. The bad news is the relationship between you and the merchant has changed and this creates problems outside the AOL-like walls.

Suppose I tell my agent to use my Chase Mileageplus card to book a flight at United. If my Link wallet charges that Mileageplus account, but pays United with a Stripe issued virtual card, the money may move without delivering everything I intended (where are my promised miles?). Link’s own terms warn that underlying card offers, benefits and rewards do not carry through in certain separate-payment-method transactions.

Stripe can reproduce the payment without necessarily preserving the relationship. That is the distinction between payment interoperability and relationship interoperability.

The war for routing

Now imagine my agent can access Link, Shop Pay and PayPal, with several cards available across them.

Which should it use?

I might prefer my airline card for airfare, another card for groceries, or the payment method already saved with a particular merchant. I may prioritize rewards for one purchase and purchase protection for another.

As we use agents going forward, “buy it” is not a complete payment instruction. The agent needs to understand my preferences, the merchant relationship and the consequences of each payment route. And when those interests conflict, whose economics should it optimize, the consumer’s, the merchant’s or the agent platform’s?

We are headed to a world where an agent needs a payment policy, not just a payment credential.

The war for liability

With so many intermediaries between me and my purchase, who bears liability for mistakes? Suppose I authorize my agent to spend $500 on a hotel. It stays within budget but books the wrong dates. The hotel delivers exactly what the agent ordered.

Who gives me my money back?

We are re-writing the rules for liability as each of the participants define their role. For example,- the card networks have frameworks for scoped payment credentials that restrict spending by amount and merchant and preserve the cardholder rights to chargebacks. Most wallets treat transactions from a connected agent as authorized even when the agent misinterprets instructions. The agent too will have to show that it followed the guidance set by the user.

So we have overlapping permissions and recourse. Authorization, correct execution and purchase protection are different problems. Authorization is addressed by payment tokens. Execution by the agent and purchase protection by the wallet and card issuer.

For consumers, the question will be much simpler: when my agent gets it wrong, who makes it right?

The Future is Open Standards

As agents scale, the AOL-era will give way to open standards where agents preserve my relationship with the merchant, use my preferred issued card and ensure my transaction earns the right loyalty and benefits rewards.

The card networks announced a set of standards for Agentic Commerce that allows agents to safely receive my card and (more importantly) transact as me. Each network has its own world salad approach: Visa has Visa Intelligent Commerce, Mastercard has AgentPay and AMEX has ACE.

These standards from the card networks move us from the “AOL of agentic payments” to something more like the open web. The best of all worlds is a future where my agent can use my credit card, without having my credit card. How could we achieve this?

• I store my card in a trusted wallet

• Before a transaction starts, the wallet creates a scoped agentic token that can safely be handed to the agent and given to the merchant. The token can’t be used by another party or merchant and has built-in safeguards that match the prompt (“don’t spend more than $100”)

• The merchant receives the agentic token and can transact using this credential, which maps back to my original card and preserves all loyalty and cardholder benefits.

AOL was a helpful set of training wheels for getting everyone online, but the open web ultimately won. We think the same thing will happen with agents as well.

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’s small cities are on the front lines of a fight most of the country doesn’t see. A ransomware attack shuts down a county’s 911 dispatch system. A school district’s student records end up for sale on the dark web. A rural hospital pays a six-figure ransom it can’t afford because paying is cheaper than losing patient care for a week. These aren’t hypotheticals — they are a regular occurrence across the United States, and the institutions absorbing the damage rarely have the budget, staff, or technical bench to defend themselves.

I spent the last several years working on exactly this problem from inside the U.S. government, first at the NSA and then as Deputy National Security Advisor for Cyber and Emerging Technology in the White House. The pattern is consistent: the same institutions expected to protect the most sensitive data and systems in our communities — student records, patient files, water systems, emergency services — are the least resourced to do it. Meanwhile, the AI wave that is improving productivity, diagnostics, and cyber defense in well-resourced organizations is largely passing these institutions by. This isn’t because the technology doesn’t apply, but it may not be accessible, affordable, and practical for smaller institutions.

That’s the gap we built Securing America’s Infrastructure Foundation to try to close.

What it is

Securing America’s Infrastructure Foundation (SAIF) is a new 501(c)(3) nonprofit, incubated by a16z, built to deliver AI and cybersecurity tools, hands-on support, and training to public institutions across America that otherwise lack access to AI tools — think small municipalities, school districts and rural hospitals.

This isn’t a grants program in the traditional sense. We’re not writing checks and hoping the right thing gets built. Instead, the Foundation will deliver software, hands-on training, and cyber incident response at no cost to the recipient. The model is closer to “forward-deployed engineering” than philanthropy as usual: our experts show up, help stand up real defenses, and stay long enough to make sure the capability sticks.

Concretely, that means:

• A deliberately mixed pilot: We will launch with a pilot in two small cities in different geographic regions because the vulnerabilities we’re addressing — ransomware, digital exclusion, underinvestment in public IT — affect communities across America.

• Sectors: We want to focus on specific sectors - the nation’s 1,800 rural hospitals for example - that provide critical services to communities where they are the only providers.

• Cybersecurity: We’re focused on AI based cybersecurity to start, given the risks of more and more sophisticated cyber attacks and the way AI is a game changer for cybersecurity (e.g., using AI to find and fix vulnerabilities in code for devices like MRI machines or hospital monitors)

• A five-year path to scale: From that pilot, we’re building further. We aim toward roughly 5-10 anchor cities over five years, beginning with $25 million in committed funding from a16z’s founding partners.

Why cyber-defense and public benefit, first

There is a lot AI and advanced technology could do for the public sector, but our first-year focus is deliberately narrow because we know these are hard problems: cybersecurity, digital literacy, and responsible AI use for public benefit. As we evaluate recipients, factors like leadership buy-in, digital vulnerability, infrastructure readiness, and the diversity of communities we can reach will be key. A program like this only works if it’s built with the people it’s meant to serve.

Who’s building it

I’m leading this effort and will initially run the nonprofit as a volunteer, establishing its strategy and building the founding team. Over time, we will hire an executive director to lead the organization’s day-to-day operations and continue building the team on the ground. I am proud to be joined by board members who have demonstrated exactly the kind of expertise — across cybersecurity, technology, and public-sector leadership — that this work demands.

• Brian Barrios is Vice President and Chief Security Officer at Southern California Edison (SCE), one of the nation’s largest electric utilities, where he leads the company’s cybersecurity and physical security strategy. Before SCE, he was the Executive Director, Threat Management and Intel for IT for Southern Company and spent years at MITRE supporting national cybersecurity infrastructure and the FBI’s Cyber Division.

• Stephen Goldsmith is the Derek Bok Professor of the Practice of Urban Policy at Harvard’s Kennedy School, where his work focuses on how cities use data and technology to better serve residents. Having served as Deputy Mayor of the City of New York and Mayor of Indianapolis, he brings direct governing experience to that research and firsthand knowledge of what it actually takes to get new technology adopted inside a city government.

• Bob Lord has spent his career securing high-profile cybersecurity targets, serving as Twitter’s first security hire, CISO at Yahoo, and later as a Senior Technical Advisor at CISA focused on making basic cyber practices accessible nationwide.

• Vishal Misra is the RKS Family Professor of Computer Science and Vice Dean of Computing and AI at Columbia’s Engineering school. His research spans network systems, AI, and global technology policy, including work shaping India’s net neutrality framework. As part of the founding team at CricInfo (later acquired by ESPN), he pioneered live sports coverage online, and he later built a commercial GPT-3 application before ChatGPT’s release — reflecting a career spent bridging research with entrepreneurial deployment.

• Phil Venables spent 17 years as Goldman Sachs’ first Chief Information Security Officer (CISO) and later became Google Cloud’s founding CISO, where he led the risk, security, compliance, and privacy teams for one of the world’s largest cloud platforms. His experience securing a systemically important bank and then building security for a hyperscaler serving thousands of customers gives him a rare view into both the enterprise-grade defenses under-resourced institutions need and how to make them practical at scale. Phil is now a Partner at Ballistic Ventures.

Why now

AI brings opportunity and risks. For small cities and municipalities, it’s harder to take advantage of AI to make city services better or more secure. We intend to show up, understand their specific constraints, and build alongside them. We know these are hard problems and we will approach them with hard work and humility.

That’s what Securing America’s Infrastructure Foundation is for. We’re starting small and focused deliberately — two cities, one sector, one clear mission.

If you lead a municipality, school district, hospital system, or community organization that’s trying to hold the line on cybersecurity, we want to hear from you. Contact us at info@secureourinfra.org.

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.

Traditionally, calling your internet provider’s support hotline about an outage went something like this. A robotic voice greeted you as a valued customer, then made you spend several minutes on phone menus (“press 3 for technical support, 2 for internet, then 4 for all other issues”). You were put on hold, feeling forgotten, until a service representative finally picked up and asked for your name, account number and issue. If restarting your router didn’t fix the problem, you were put on hold again, only to tell another representative the whole story a second time before receiving a technician appointment.

AI is already transforming support interactions like this one. Instead of grappling with a robotic voice and several service representatives, an AI agent can now pick up right away. It can see your account, check whether other customers in your area are also affected and either resolve simple problems directly or book a technician. No more “all our service representatives are currently busy” and no more hold music. Until recently, this was the technological limit.

But now, new capabilities are allowing agents to follow threads and execute tasks over longer horizons. We believe this will allow AI companies to go beyond individual customer interactions and rethink how to address complex customer journeys that run over weeks and months. AI will guide insurance customers from an initial call to payout, mortgage borrowers from pre-approval to closing, and enterprise leads from their first demo to signature. This will not only create better outcomes for those customers, but also let the AI vendor get paid for the resulting revenue expansion, not just for cost savings.

Value Only Appears At The End Of The Relay Race

There are many kinds of businesses where creating value for customers takes many steps, over weeks and months. Let’s take the traditional plaintiff law case preparation journey as an example. Prospective cases often arrive as leads by phone, where intake specialists gather the key facts and pass them to an attorney who evaluates whether to take on a client. Once a firm signs a case, investigation and evidence collection follow: Paralegals order police reports, try to get in contact with witnesses, and send spoliation letters to defendants to preserve evidence while keeping clients updated on progress. Firms then wait for client medical treatment to conclude before drafting a medical chronology and handing information to attorneys to finalize the demand letter.

This process is what we historically would have called a “relay race”: a long journey with risk of drop-off or value leakage at each handoff. Potential clients usually call several personal injury firms and often sign with the firm that gets back to them the quickest. So cases can drop out of the funnel if clients only reach voicemail or if the qualification decision is slow. Once a client is signed, evidence collection involves many parallel threads at once and firms may not have the persistence to call a potential witness for the fifth time after the previous four attempts went unanswered. In addition, labor constraints can keep firms from reviewing every piece of evidence in detail, especially for seemingly routine, low-value claims. A firm may accept the first offer when a detailed read of the doctor’s note would have surfaced indicators of a brain injury, increasing settlement value. Finally, cases can run for months, require numerous contributors, and involve information distributed across documents, stakeholders, and systems. This makes it nearly impossible to maintain the full picture on every client. As a result, important bills may not be reflected in the demand letter or treatment gaps may remain unexplained, driving lower settlement value.

The Long-Running Agent Opportunity

Until recently, AI agents mostly handled bounded tasks (“Send this follow-up email”, “Summarize that call”, “Create this document”) and narrow interactions. For instance, customer support has been a great early AI use case because the problem space is well-defined, making high reliability achievable, while still allowing for customization based on company-specific policies. Interactions are brief, outcomes are verifiable, and if AI agents encounter edge cases they cannot handle, existing support teams are usually still in place for escalations.

Now, long-running agents are able to pursue goals across dozens of steps, modalities, and longer time horizons, allowing them to take on relay races from start to finish. In the plaintiff law example, a long-running agent could be assigned to a case from the initial intake call with the goal of maximizing the client’s recovery. The agent would save and retrieve context on the case as needed. It wakes based on events or deadlines, proactively dispatches sub-agents to contact witnesses, update the client, and interact with attorneys across phone, email, or text. It then consolidates findings, plans subsequent moves, drafts documents for review, and drives the case towards the best possible outcome. Across clients, it keeps learning, as each case generates data which helps the agent process the next case better than the current one.

The agent addresses key value leaks observed in today’s relay race. At the start of the journey, clients are no longer lost to slow case evaluation as the agent is available regardless of the hour or weekday and can assess cases nearly instantaneously. Evidence collection is no longer bounded by human capacity, as the case agent can spin up an additional sub-agent at any moment to place the fifth, tenth, or even twentieth phone call. As agents are nearly infinitely scalable, even seemingly routine cases get worked fully and each piece of evidence is reviewed in detail, surfacing the potential brain injury that might otherwise have been missed. And finally, the agent keeps a complete record of the case, from witness phone call transcripts to police reports and medical records, making sure no information gets lost.

Getting Paid For Outcomes

Beyond closing these leaks, owning the full customer journey can change the economics for the agent vendor. As agents help deliver outcomes with higher quality at greater speed, they can drive revenue uplift, not just cost reduction for customers. More importantly, as impact is observable and attributable to agents, vendors can participate in the upside they create by pricing based on work products or outcomes. In plaintiff law, for instance, a single case is a discrete unit of work and the value of the resulting outcome, the settlement amount, is clearly identifiable. In addition, outcomes can be linked to agents more easily as they drive a larger share of the final work product. Vendors may then share in topline expansion, reaching contract values previously unattainable by solely focusing on cost reduction.

Admittedly, switching to outcome-based pricing is not frictionless. In plaintiff law, most states prohibit non-lawyers from sharing in legal fees, so pricing would need to align with settlement values without taking a direct share of them. In other verticals, vendors may face objections from procurement teams who do not yet have precedents for outcome-based contracts. And even though agents move closer to fully owning outcomes, attribution debates will likely surface initially. Successfully deploying long-running agents will therefore be a commercial challenge as much as a technical one.

Still, getting paid for outcomes on high-value customer journeys is not a new idea.

The Case For An AI-native Red Ventures

Red Ventures (RV) is a great historical example. RV was founded by Ric Elias and Dan Feldstein and pioneered digital end-to-end outsourced sales operations for long-term consumer contracts. As an example, for insurance policies, RV would collect details on customers when they were searching for these products (by owning sites like homeinsurance.com), reach out, walk them through coverage options, and close the policy. By fully owning this journey, RV collected a payment from the insurer for each policy it closed. This used to work extremely well: according to S&P Global, RV generated ~$370M in EBITDA on ~$1.4B in sales in 2022.

Now, this model seems structurally under threat. At the top of the funnel, consumers are changing how they discover products. Instead of relying on Google search or comparison sites to explore and compare products, they can now ask a personal agent like Meta’s Muse or Instinct for product recommendations directly, lowering the value of owning traditional exploration sites. Further downstream, instead of hundreds of humans running linear funnels, long-running agents can work across far more complex journeys. They will do so across products, and communication modalities, in parallel, in any language and at scale, with infinite persistence. RV’s EBITDA decline to an estimated ~$160M in 2025, driven by customer churn and softer ad spend, provides an early indication of the model’s vulnerability as AI will transform the industry further going forward.

Three Ways To Market

There are different ways to bring long-running agents to market.

(1) Selling the Platform: Under this motion, the vendor sells the long-running agent as software, supporting the enterprise in its existing operations. The plaintiff law case agent integrating with an existing team and their workflows is a good example. Selling a platform allows for revenue upside as there are often no comparable software or outsourced services contracts in place anchoring prices. This, however, also extends sales cycles as new budget lines need to be created, and it is harder to clearly attribute outcomes to an agent that interfaces with several departments across the enterprise.

(2) Selling the Service: Here, the vendor takes over an already outsourced service to deliver a desired outcome more effectively, just like Red Ventures owning the entire customer acquisition journey for an insurance carrier. This approach shortens sales cycles as the ROI math for buyers is simple and budget lines for BPOs already exist. However, existing, pre-AI pricing models may not reflect the true value the long-running agent can deliver and many high-value journeys are run in-house today, limiting the reach of this approach.

(3) Going Full-Stack: A full-stack approach means vertically integrating and becoming an AI-native player. For instance, instead of selling a long-running agent to a debt collection firm, you could buy debt portfolios and collect them yourself, allowing you to fully capture any AI upside created and address attribution problems. However, this also means owning parts of the service like licensing or compliance, which are outside the actual customer journey, adding distractions and operational complexity.

It is too early to tell which of these approaches will prevail, but we are convinced that long-running agents will win the relay race at scale and expand the market addressable by AI from a share of labor budgets to a share of revenue. Plaintiff case preparation is only one example and we expect to see agents drive outcomes across other high-value journeys like mortgage origination, insurance claims processing, complex enterprise sales and countless others. If you are working on capturing this opportunity, we would love to hear from you!

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.

Across the United States, specialty medical practices are under growing financial pressure. The cost of running a practice is rising faster than revenue.

While the discussion around AI in healthcare has mostly focused on its potential to lower costs and increase efficiency, AI can have a meaningful impact on the other side of the ledger: revenue. Prosper AI conducted a study to quantify the revenue opportunity tied to where practices can introduce AI to provide more consistent, reliable care and improve practice economics in the process.

Prosper AI leveraged data from interactions across patients and insurance companies accounting for more than $3 billion in care managed, and combined that data set with specialty-specific data from the Medical Group Management Association (MGMA). Prosper AI then conducted original research, a phone-access study called CallIQ, to assess the current state of patient access, covering approximately 2,000 practices across 29 states; the specialties included dermatology, ophthalmology, gastroenterology, OB-GYN, primary care, and orthopedics.

Key Findings

Our research identifies four actionable levers within practice leaders’ control, representing an estimated annual revenue opportunity of approximately $125,000 per physician.

The Growing Operational Squeeze in Specialty Practices

Specialty practices face compounding margin squeeze as rising costs collide with stagnant reimbursements and severe administrative overhead. Heading into 2027, this squeeze is likely to intensify as inflation persists. Meanwhile, Centers for Medicare & Medicaid Services (CMS) announced that it has no plans to renew a temporary 2.5 percent boost in Medicare reimbursements set to expire at the end of 2026.1

The expiration of these temporary payment cuts will cascade across specialty practices and add additional pressure on practices. Crucially, these Medicare cuts impact all insurance reimbursements, as commercial health plans benchmark their contract rates directly against Medicare fee schedules. These cuts come at a tough time for practice leaders; 84 percent of medical groups report higher operating costs than a year ago, with an average increase of 11 percent.2 Meanwhile, inflation-adjusted Medicare physician payments have declined 33 percent since 2001.3

The $125,000/Physician Revenue Opportunity

CAPTURE DEMAND: Convert missed calls to appointments

As basic as it sounds, reaching a provider’s office by phone remains a fundamental breakdown in patient access. Prosper’s proprietary CallIQ data analyzing 2,000 practices reveals an 18 percent daytime missed-call rate, with answered callers facing an average wait of 1.5 minutes, and 20 percent stranded on hold for over 5 minutes. After hours, the missed-call rate rises to 81 percent.4

Every missed call – and long wait time – creates the possibility of a lost patient or unfilled appointment slot. This adds to a growing problem. Our data shows that 19 percent of provider schedules go unfilled across practices; at some individual practices, as much as half of available capacity remains unused.5

The impact of calls that go unanswered can extend beyond a single lost appointment. Poor phone responsiveness can create a negative patient experience that affects reviews, referrals, and the likelihood that patients that need care come back for it.

AI-enabled patient communication delivers 24/7 coverage, enabling practices to handle more inquiries, convert higher call volumes, and maximize provider capacity. When scheduling is fast and frictionless, patient behavior shifts — elevating provider appeal and actively driving patient acquisition. Practices that are easy to reach and schedule actively grow practice volume by attracting patients looking for accessible care.

At Piedmont Plastic Surgery and Dermatology in North Carolina, with Prosper AI handling its calls, the practice saw a 17 percent lift in appointments booked in August 2026, compared to August the year prior. Across their 37 providers, appointments booked rose from 19,209 in August 2025 to 22,524 in August 2026.

Wait times during peak hours fell from seven minutes to zero, and the practice was able to serve 50 callers simultaneously, up from 15 the year prior. Based on an MGMA benchmark of $392 on average for a dermatology visit, the additional 3,315 appointments booked would translate to an estimated $1.3 Million in revenue for that month alone.

RECAPTURE DEMAND: Re-engage existing patients

One of the least visible sources of lost revenue is patients who simply fall out of the care cycle. A patient may be told to return in six months, schedule an annual skin check, come back after surgery, or return for ongoing management of a chronic condition. Unless the next appointment is booked before the patient leaves or the practice has an effective recall system, some of those visits never happen. Unlike a no-show, which creates a visible opening in the schedule, an unscheduled follow-up may never register as lost revenue at all.

Missed appointments that are never rescheduled create a similar problem. Across medical specialties, the average no-show rate reaches 18.8 percent with gastroenterology being the highest, at 25.7 percent.

A Massachusetts General Hospital study found that nearly 19 percent of dermatology appointments were missed, and only 41 percent of those patients returned to the practice within the following year.

Cost concerns are a key factor contributing to missed care. The problem is also growing, likely due to patients’ sensitivity to rising cost: 32 percent of medical practices reported more no-shows in 2026 than in 2025, while just 10 percent reported fewer, amid mounting affordability pressures on patients.

AI-enabled recall and 2-way SMS outreach systems can systematically identify patients who are due for follow-up and engage them to schedule, reducing the number who fall out of the care cycle.

PROTECT REIMBURSEMENTS: Prevent avoidable denials

Ensuring full reimbursement starts long before claims are submitted. For practices operating with increasingly tight margins, financial performance depends on accurate, pre-visit insurance eligibility and benefit verification. Prosper AI’s research reveals that practices see an average of 10 percent of their first-submission claims denied. Denial rates for more complex care, such as orthopedics and gastroenterology, are higher at 22 and 16 percent respectively. Every denial creates additional work for staff while delaying, reducing, or potentially eliminating the probability of successful reimbursement.

The sheer volume of denials and appeals has become a major operational burden. Eighty-four percent of healthcare finance professionals describe managing them as a significant challenge, including 35 percent who call it their “biggest challenge” and another 49 percent who describe it as “very challenging.” For practices already struggling with staffing, the problem can become self-reinforcing: Fewer employees are available to ensure claims are accurate before submission or to pursue reimbursement after a denial.

A meaningful share of this leakage is preventable. Missing or inaccurate claim information, incomplete patient registration data, and coding errors are among the leading causes of denials. In ACA Marketplace plans, for example, administrative errors or omissions account for roughly one-quarter of in-network claim denials.

AI-enabled, pre-visit benefit checks can automatically verify coverage details, confirm authorization status, and flag missing patient details before the visit occurs, eliminating simple administrative errors and stopping preventable denials at the source. At a time when practices have limited control over reimbursement rates, capturing more of the revenue they have already earned is one of the financial levers they can directly influence.

RECOVER REVENUE: Recover the patient responsibility

As patients shoulder a larger share of healthcare costs, recovering the amount that patients are responsible for has become increasingly important to ensuring practices can continue to provide quality care.

Patient balances include deductibles, coinsurance, copays, and self-pay bills. According to our internal data from completing over five million calls across providers and payers, patients pay an average of 36 percent of the cost of their care across specialty practices. At dermatology practices, patients are responsible for a majority of the costs, at 55 percent.

Yet most patient responsibility is not taken when care is delivered. Medical practices recover an average of just 39 percent of patient-due balances at the time of service, leaving the majority to be recovered afterward. This creates additional administrative work and increases the likelihood that balances will go unpaid.

The challenge is growing. In a 2025 survey, 29 percent of medical practices said recovering outstanding patient balances had become more difficult over the previous year, compared with 26 percent that reported improvement.

At one multi-location dermatology practice, AI reminder calls helped recover 48 percent more outstanding patient balances than standard follow-up. In a randomized comparison, households whose balances had reached a third billing statement received either AI calls or the practice’s usual statements and follow-up.

The AI calls explained balances, answered questions about charges and insurance, and offered payment links, payment plans, or a transfer to the billing team. Over approximately 50 days after the third statement, the practice recovered 31.2 percent of outstanding balances in the AI group, compared with 21.1 percent in the standard-follow-up group.

Calculate Your Practice’s Revenue Opportunity

Practice leaders can determine how much revenue they can add through a 2-step process:

1. Identify your specialty’s annual impact per physician. Locate your medical specialty to see the estimated per-physician revenue opportunity across the four levers (convert calls to appointments, re-engage existing patients, avoid preventable denials, recover the patient responsibility).

2. Calculate your total annual opportunity. Take the total per physician for your specialty, and multiply that by the number of full-time physicians in your practice for your total annual revenue opportunity.

The New Frontier: Putting Revenue-Guaranteed Appointments into Practice

Macroeconomics are out of practice leaders’ hands. These four levers are not.

By using AI to convert calls to appointments, re-engage existing patients, avoid preventable denials, and recover the patient responsibility, specialty practices can capture an estimated $125,000 in additional annual revenue per physician. The bigger opportunity is what happens when the four levers work together.

That starts with a benefit-verified appointment: an AI-enabled system confirms the patient’s coverage, calculates what the patient will owe, and tells them before the visit. The result isn’t just a fuller schedule. It’s a schedule filled with revenue the practice can count on, because every booked visit has already been checked against how it will be paid.

Each step also strengthens another lever. Accurately verified benefits prevent denials. Patients who know their costs upfront are more likely to show up and then pay for the care provided after the visit. And as patients shoulder more of their care costs, a practice that is easy to book and clear about costs stands out, which brings in new patients and starts the cycle again.

Methodology

The CallIQ Study

To measure the current state of patient access, Prosper AI created the CallIQ phone access study. Prosper AI attempted to reach 2,000 practices across 29 states and 6 specialty types: dermatology, ophthalmology, gastroenterology, OB-GYN, primary care, and orthopedics. No appointments were booked. For every call, we tracked two things: the time it took for a live person to answer, and how long it took to begin scheduling.

Public and industry benchmarks

• MGMA reports: revenue, visits and revenue per visit by specialty, which give each practice’s baseline. MGMA DataDive Financials and Operations. Used with permission from MGMA.

• Federal health data (CMS, AHRQ, CDC): visit patterns, how often patients return, time since last check-up, and payer mix

• Peer-reviewed studies: follow-up adherence, claim denials and patient responsibility. These fill gaps where no national dataset exists

About Prosper AI

Prosper AI is the patient access platform that helps exceptional specialty practices grow. Its AI workforce connects patient access and reimbursement across the front and back office—from scheduling and patient outreach to benefits verification, prior authorization, claims follow-up and patient billing. By handling patient and payer conversations directly within practice workflows, Prosper AI helps practice leaders grow appointment volume, protect reimbursements and recover outstanding revenue—giving their teams more time to focus on care. Prosper AI is live in 100+ outpatient groups and covers 150,000+ providers across more than 27 medical specialties, with deep integrations across athenahealth, ModMed, Veradigm, ECW, ImagineSoftware and other leading EHRs. Founded in 2023 by Xavier de Gracia and Josep Mingot, Prosper AI is backed by a16z, Emergence Capital, and Y Combinator and headquartered in New York. Learn more at getprosper.ai.

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’s future depends on whether we can modernize the institutions responsible for our security, resilience, and prosperity. American Dynamism backs ambitious founders building companies to meet those needs at scale.

But capital is only part of what new companies need to win. They also need leaders who understand how these institutions operate, where change gets stuck, and how to turn new technology into lasting national capability.

Today, we continue to deliver best in class access to our founders by welcoming Colonel Jeffrey Glover and Greg Barbaccia as Advisors to a16z’s American Dynamism and Growth practices.

Col. Jeffrey Glover

Col. Glover brings more than two decades of public-safety leadership. As Director of the Arizona Department of Public Safety, he leads statewide emergency response, criminal investigations, intelligence, aviation, and highway patrol. He previously led the Tempe Police Department and is also a former president of the National Organization of Black Law Enforcement Executives (NOBLE), where he has remained an active leader.

He has been a forward-looking leader on technology in public safety, helping officers respond faster, make better decisions, and return home safely while earning the trust of the communities they serve.

Col. Glover will advise companies building for public safety and law enforcement, helping founders understand operational needs, build responsibly, and get critical new capabilities into the field.

Greg Barbaccia

Greg has spent his career at the intersection of technology, national security, and government. Most recently, he served as U.S. Federal Chief Information Officer, overseeing technology strategy, cybersecurity, and digital transformation across the federal government. He also led the General Services Administration’s Technology Transformation Services as Acting Director.

Before that, Greg spent more than a decade at Palantir in operations, security, intelligence, and investigations. Earlier, he served as a U.S. Army intelligence analyst and intelligence sergeant, including deployments to Iraq and Afghanistan.

Greg understands government technology as an operator, executive, and customer. He will advise founders across government technology, cybersecurity, defense, intelligence, and digital services.

A coalition of incredible leaders

They join an extraordinary group of leaders supporting a16z founders at the intersection of technology and the national interest, including General Bryan Fenton, Special Advisor to American Dynamism, and Anne Neuberger, our General Partner and Head of Global Affairs. And our group of advisors will continue to grow.

We are building a top tier team of experienced operators to give our founders and their companies the best possible support as they build the companies that will shape the future. Greg and Col. Glover make that team even stronger. Please join us in welcoming them as advisors to Andreessen Horowitz.

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.

We’re excited to announce that a16z is leading an investment in TypeSafe AI, the creators of Jev. Read more about why we invested in our blog post and hear from the team in our a16z Show episode.

Earlier this week, Flávio Bolsonaro secured a surprise first-round victory in the Brazilian presidential election. Or maybe it wasn’t a surprise. Charts knows next to nothing about Brazilian politics, so we really couldn’t say.

If you take a gander at the prediction markets, however, it appears as though Bolsonaro’s win was at least somewhat surprising, but not too surprising:

After pacing as a long shot for most of the year, odds began to turn in Bolsonaro’s favor in August, and by the second week of September, he was already the favorite. Unsurprisingly, Bolsonaro’s odds for winning the presidency have since skyrocketed, after the recent first-round win.

Whatever the surprise, prediction markets were early and right on this one by ~3 weeks. The outcome of the election is apparently consequential for financial markets too, judging by MSCI’s Brazil ETF, which jumped ~15% after the news of Bolsonaro’s first-round win. To be fair, the ETF also started climbing back in August (when Bolsonaro’s Kalshi odds began to turn), but not nearly as steeply or as uniformly. In other words, it seems fair to conclude that prediction markets had this one first.

All of this was a somewhat longwinded segue to some recent research that purports to show that prediction markets are actually pretty good at elections, or at least, they don’t demonstrate much partisan bias.

Prof. Zitzewitz published some preliminary findings in an NBER working paper that aggregated over 100 years’ worth of prediction markets (albeit with most of the sample coming recently) looking for evidence of over-pricing along various dimensions, including demographics and political affiliation. The premise of the study is that any bias would show up as outsized returns (either positive or negative) for whichever attribute(s) prediction markets systematically over- or underrated. That makes sense: if, e.g., prediction markets consistently underrate female candidates, then taking the “over” on female candidates would be a winning strategy (reflected by a positive coefficient).

As it turns out, the evidence of bias is pretty much non-existent:

This chart summarizes a subset of the results, and among political affiliation, gender, color, or age, there is not a single dimension where the return coefficient was above or below zero by an amount that exceeded the standard error—the only exception is for non-US elections, where prediction markets appear to reflect some bias in favor of right-leaning candidates (i.e. the coefficient is negative because markets overrate their chances), albeit over a much smaller sample. Even there, however, the 95% confidence interval indicates that the bias is not statistically significant.

Elsewhere, the results suggest a small (statistically insignificant) bias in favor of women and non-white candidates, and against older candidates, but when it comes to partisanship, the coefficient is a bare 0.03.

Take that for what you will, but the results imply that when it comes to elections, prediction markets are pretty good at leaving their biases at the door.

Rule of 40 Ain’t What It Used To Be

For years, the Rule of 40 was considered a lodestar for techco performance. The idea was to create a single metric that would enable comparisons across fast-growing companies at various stages of their lifecycle, where very rapid, and very unprofitable, growth was expected early, and less rapid, but more profitable, growth was expected as companies mature.

The Rule of [X] part is determined by summing revenue growth with profit margins, and if the number was 40 or higher, you were in good shape. So a company with 100% revenue growth, and -60% margins, was still winning because 100 + (-60) = 40, i.e. Rule of 40 on the nose . . . whereas 150% growth with -150% margins, clocking in at Rule of 0, would be less impressive (despite the faster growth).

Given that Rule of 40 is supposed to be a good thing, it’s odd to see these two things happening at once: public software cos are raising their Rule of X, while multiples are in steady decline:

With some fits and starts, the median Rule of X for public software has risen from ~17% to ~23% over the past ~3 years, with the steepest upward inflection beginning early 2025 . . . just when revenue multiples began their precipitous decline.

Here’s another way of looking at it: the correlation between Rule of X and multiples (on a rolling basis) turned sharply south and eventually negative:

Again, it’s a weird thing. If Rule of X is a good thing, then an improving Rule of X should correspond to better multiples, and not the other way around.

So, what’s going on? The most likely explanation for why Rule of X is weakening as a signal has to do with the component parts: growth and profitability. As laid out in State of Markets, public (and private) techcos responded to post-ZIRP regime change by trading growth for profitability. What that means is that Rule of X is rising, but increased margins are doing most of the work, while growth makes a relatively smaller contribution.

But when it comes to multiples, investors put more of a premium on the latter than the former, so even if the Rule of X number is high, if growth has slowed substantially, multiples compress accordingly.

If you break techcos into their Rule of X buckets, you can see that revenue growth is just a much stronger factor than the combined Rule of X score:

Mind the different scaling on both axes, but the general idea is pretty clear: higher growth companies tend to have higher multiples than lower growth companies with an equivalent Rule of X.

In fact, for Rule of 20 all the way to Rule of 40, most companies are clustering in the same range between 5-10x EV/NTM sales—and it’s only the fastest growers (and really the ones with perceived tailwinds at their back, i.e. cyber and observability) that fetch a premium multiple.

The point here is pretty straightforward: in this brave new world for techcos, Rule of 40 just isn’t as useful anymore as a performance benchmark. The SaaS-Prove-It regime is okay with profitability, but it really likes growth, and if you’ve got a lot of the former, but not much of the latter, then the market isn’t very impressed, Rule of X be damned. If you’ve got both, then that’s all the better, but there aren’t too many of those.

Atoms keep winning, but what about everyone else?

Speaking of performance metrics where the headline picture may be obscuring some parts of what lies below, consider the recent downturn in the breadth of positive earnings revisions:

When breadth is positive (as it’s been since July of ‘25), that means earnings revisions are relatively broad-based . . . and when it’s negative, it means the good news is limited to a narrower subset of companies.

While revisions breadth is still positive, things have narrowed recently, with the implication that analyst optimism is training its fire, more so than before.

And where, might you ask, does even-more-optimism still abound? It’s really only in two sectors: energy and tech, but more specifically hardware, not software:

For the year, earnings revisions have been positive across the board (outside of utilities), but for Q3, the story is different: Energy improved, Tech improved—driven almost entirely by semis and hardware, with only the slightest uptick for software—and Financials improved modestly, as well, but every other sector experienced a downward revision from analysts. 2027 is a similar story: Energy, Hardware and Semis are the clear analyst winners, while the rest of the market, not so much.

In other words, expectations are that profits will continue to flow to Atoms, substantially more so than Bits.

AI capex shows no signs of slowing—if anything, the industrial backlog is increasingly broad-based:

As per the ISM, the share of industrial industries reporting a higher order backlog is now ~60%, which is the highest it’s been since pandemania.

At the same time as spending on the infra layer keeps piling up, there is some evidence that spending on everything else—while still high—is beginning to taper:

According to Carlyle’s private portfolio company data, shipments of AI hardware are still climbing at their exponential pace, while the “other side of the AI spend coin,” i.e. growth of the broader category of tech-spend, has leveled off at ~30%.

30% growth is pretty high—and much higher than before—but growth does not appear to be accelerating any longer, which is again, not the case when it comes to hardware. The implication is that after a year of better than expected results for pretty much everyone, that upside surprise may be narrowing.

Put another way, while the AI buildout keeps embiggering, that hasn’t changed expectations for how those benefits may flow through to the rest of the ecosystem—the epic run on atoms is picking up speed, even as it runs ahead of the field. Those unchanged-to-downwardly-revised forward estimates for everyone outside of AI infra are backed up by some hard data, as well, insofar as broader IT spending, while still growing, does not appear to be accelerating.

Of course, the fact that analysts have made upward revisions to their estimates for most of the year means that they’ve gotten this “wrong” before, but for now, consensus is signaling that the AI buildout will stay in the driver’s seat, even more so than before.

AI Has Come For Call Centers, After All?

While the broader evidence of the AI Jobs Apocalypse remains weak to non-existent—and indeed, there’s evidence that AI has increased demand for technical and hard-hat roles—there is some data that shows at least one “AI exposed” category is taking it on the chin: call centers.

According to data from Revelio, global call-center employment has gone sharply negative, after more than a decade of fairly consistent growth:

To be fair, growth began slowing when interest rates went up, which is slightly before GPT was released into the wild . . . but the downtrend has steepened substantially since then, and has accelerated of late.

Indeed, while lower-middle-income countries were relatively insulated from the downturn through most of 2024, even there growth has gone negative:

While higher-income countries went negative in 2022, lower-middle income countries stayed positive for another ~2 years. But, as of Q4’25, call center employment growth has gone negative everywhere, even in the poorest countries.

And lest one think that the downturn isn’t AI-driven (or specific to call centers), that too does not appear to be the case:

Outside of China, interestingly enough, call center employment has lagged the broader category of white-collar service work in every major non-US market for call centers.

Could there be other reasons for below-trend call center employment, besides AI? Probably, but call centers were always expected to be a likely target of substitution, so it seems reasonable to conclude that AI is the driver here. In the US, at least, the good news is that unemployment has been relatively stable, suggesting that the people who used to work in call centers appear to have found employment elsewhere . . . but changes are definitely underway, and one would expect more to come.

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.

The Jev launch is the biggest narrative violation we’ve seen this year, in the best possible way.

The reason behind the instant take-off is obvious in hindsight. While AI is incredibly useful for creating software like a human, at superhuman speed, it’s been far more difficult to integrate into code. “AI is eating software” has ironically meant that it’s writing a lot more traditional software a lot faster, but not necessarily making the software more intelligent.

The real economic unlock should be that all software is imbued naturally and natively with the intelligence of the latest models, at performance and cost that makes sense with traditional software economics. This is exactly what TypeSafe AI accomplished with Jev, and is why it’s that fastest growing model we’ve ever seen - reaching 1 trillion tokens generated in merely 3 days post launch.

Two important properties set Jev apart from every other large language model. Rather than encoding a model decision output into text, and leaving the software to parse and normalize it, Jev hands the decision directly to your code as a typed value. The output is so small that they don’t even bother charging for it. The modality innovation is accompanied by an economical one. Jev is roughly 1/100 to 1/500 the cost of frontier models while being 100x faster for classification tasks at the comparable accuracy.

However, the significance of the Jev model is not that it’s faster, cheaper and more accurate (though it is all three), but that it opened a new path of how we think about AI and its relationship to software. Jev is the first of a new class of System One models, which are built to be deeply integrated with code (rather than talked to by people) that returns instinctive decisions in the native language of software. While the big labs chase more parameters, longer context, deeper chains of reasoning that’s “human-like”, this is an entirely new direction to truly make AI ubiquitous.

From the first conversation with Diogo, we found ourselves philosophically aligned with the mission behind this movement. Jev and TypeSafe as a company are much more aligned with developers ethos: pragmatic, composable, and deeply respectful of people who build. It’s a refreshing take in contrast to the rhetoric around the big labs about replacing software engineers and killing SaaS. TypeSafe believes AI is a tool that can make software more intelligent, and returns the power to the developers hands.

The developer community responded with enthusiasm and creativity at a level we’ve never seen before. Within the first week, thousands of use cases appeared, from generative UI, to interactive gaming, to data analysis. Each one is proof that System One models belong everywhere in software. Enterprises followed just as fast - 25% of the fortune 500 enterprises have integrated Jev. Call it the Jev-ons paradox: make intelligence cheap enough to call anywhere, and it gets called everywhere.

We’re so excited to announce that we’re leading the investment in TypeSafe to partner along this journey. Like TypeSafe, we believe the future of SaaS and software is bright. And AI should make software more powerful, and not just automate the recreation of existing software. This spiritual alignment goes back many years. We’ve long been fans of Diogo, the founder and CEO of TypeSafe, and have resonated strongly with his vision that the true bitter lesson is not necessarily about building a new being, but returning the power to builders by giving them the primitives of intelligence then letting them unleash their imagination.

The entire a16z team is excited to partner with TypeSafe. Software continues to eat the world.

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.

In 2020, CFOs needed new tools. Now the tools are changing the role.

The role was becoming more strategic, but the software was still stale. CFOs were part data analyst and part architect, stitching together exports from a constellation of point solutions and Excel orbiting thirty-year-old ERPs. The data was the bottleneck. It lived in too many places and was hard to wrangle, and the tools on top were brittle and limited in what they could do.

Now AI is removing those constraints. Assembling the data has gotten easier, and the software has started doing more of the work itself. And when that changes, both how the work gets done and who does it change too. That means the finance function itself changes, along with the team and the archetype of the person leading it.

From accountant to banker to builder

You can see this in who companies have hired for the job over time:

The dimensions of a great CFO have held up remarkably well across eras. The best CFOs don’t just deliver numbers or board decks. They also use data to guide an opinion on the direction the business should take. What the job asks of them is what keeps changing with the scarce skill of the moment.

Each era’s archetype has tracked this shift. When the job was to clean books, establish controls, and report results, companies hired CFOs out of the Big 4 accounting firms. When the job was raising capital and getting the deal done, they hired out of investment banking. When the job became planning, metrics, and business partnering, the planning (“FP&A”) skillset rose in importance, with companies often still hiring from banking or private equity backgrounds.

Today’s scarce skill is something new: designing the operating system the company runs on. This means the data, workflows, agents, skills and controls that turn context into intelligence and decisions. You can see this in the rise of the “finance engineer” inside finance teams, and in CFOs increasingly playing that role themselves.

The finance leaders operating this way describe the function differently than their predecessors did. Sarah Friar, OpenAI’s CFO, frames finance as a real-time function. Finance is building toward a zero-day close and a continuously updated forecast, a big shift from the traditional monthly close rhythm. Whether for allocating compute or capital, the best finance leaders now sound less like scorekeepers and more like builders.

More data, better tools, slimmer team

The shift in the archetype of the CFO, and the broader finance organization, is being driven by the rapid progress of AI-native tools. Easier, faster data access unlocked better tools, changing how the work gets done and by whom, and in turn how finance teams are structured and how they operate.

1. Data stopped being the bottleneck. The central complaint of our 2020 piece was that the CFO’s raw material - the data - was out of date, fragmented, and painful to pull together. Finance teams were spending most of their energy at the bottom of the pyramid stitching together inputs: pulling exports and reconciling spreadsheets and chasing down the explanation for a variance buried in an email thread, or worse, in Slack DMs and channels. AI makes it faster and easier to extract, ingest and interpret data scattered across different systems and formats. Teams can spend less time acquiring information and more time actioning it.

2. The finance stack is becoming part of the finance team. In 2020, we mapped the finance stack layer by layer and argued that a wave of modern tools was coming for each one. With AI helping solve the data issue, that wave came. The new generation stands up in days rather than quarters, automates the data collection itself rather than waiting to be fed, and delivers insight in real time rather than at month-end. Every area of that original market map now has an AI-native contender for everything from an ERP to tax to procurement.

A few examples from our portfolio:

• Rillet’s ERP cuts implementation from months to days or weeks, automatically pulls in data, and turns a monthly close into a daily one.

• Concourse deploys agents for financial planning and forecasting, pulling in real-time pipeline data to build live scenarios.

• Lio uses a multi-agent model for procurement and whose agents conduct vendor diligence, negotiate contracts, oversee internal sign-offs, and monitor fulfillment.

• Stuut manages the end-to-end accounts receivable cycle from outreach to follow-up to collection and dispute resolution, automating the cash application process.

• Sphere turns complex indirect tax compliance into an end-to-end automated process.

• Petual automates SOX testing and enterprise internal audit workflows.

• CFO.ai is a personal CFO that builds and maintains real-time financial models while answering questions before they’re asked.

These are all workflows that previously needed dedicated teams or external service providers.

The last generation of software helped finance professionals manage the work. This generation does the work itself. The best AI-native finance software now does finance.

3. The rise of the finance engineer. With the data problem receding, something we didn’t predict in 2020 has emerged. The most advanced finance teams are not only using AI software tools but building their own automations, dashboards, and more internal tools. Many of these people had never written code, were skeptical of AI, and didn’t have access to engineering resources. Over the last 9-12 months, they’ve become AI-pilled.

Anthropic’s finance team has built and maintains a library of 70+ finance-specific AI skills they treat like production code. One of them produces a 90-95% complete monthly financial review, compressing hours of human work into about thirty minutes. At OpenAI, Friar’s team builds custom GPTs for investor relations and procurement, as well as live dashboards to replace static board books.

One growth-stage CFO in our portfolio asked their team to create agents and within two weeks they had 50. They weren’t all useful but they gave everyone a chance to experience the magic moment when an agent actually produces real work. Another said their finance and accounting teams are becoming product managers, working with engineers to create their own tools.

Just as we’ve seen the rise of the GTM engineer, the growth engineer, the talent engineer, many companies are starting to have a “finance engineer.” Similar to the other roles that combine domain expertise and the ability to build, this is commonly a finance person, frequently with an engineering background, who has become proficient in Codex or Claude Code and leads the charge on building tools for the team. They are often connecting systems and fixing workflow bottlenecks. For example, this could mean building a tool that combines customer revenue with model inference costs, cloud usage, and support costs, then flags accounts where margins are deteriorating or usage has made a particular contract unprofitable.

Historically, this required unlikely support from core engineering teams or attempting to hire engineers on to the finance team. Many CFOs themselves now spend more time building with AI or using AI-native tools, and less time in meetings. “These new products are fun to use” is feedback we hear often.

The operating discipline behind the finance engineer is simple: if you do something twice, write a workflow; if you write the workflow twice, write an agent.

4. The finance team gets smaller and higher leverage. Most finance teams are still organized the old way: a controller layer, an FP&A layer, an ops layer, and a CFO. All of this is built for a monthly cadence and one-person-per-task ownership. That structure made sense when every layer existed to move and reconcile data.

The old heuristic of finance at ~5% of headcount is changing too. The leanest teams we see are targeting 2% and prioritize talent density over headcount. The new finance hire can wire multiple systems together with an API call, write SQL, and prototype an automation. AI fluency is becoming a baseline requirement. “Walk me through your model” is becoming “walk me through your prompt sequence.”

Senior domain experts can get the most leverage out of AI. At Anthropic, the heaviest user on the finance team is the head of tax. As one of our growth stage CFOs put it, it’s tempting to assume the most junior people will be the strongest adopters, but they lack the experience and context to know what “great” looks like, or what to build.

5. Planning becomes continuous. Usage-based revenue (e.g. tokens, minutes, API calls) has always been complex to forecast. Moreover, standard processes and tools that operate on a monthly cadence aren’t fast enough, especially since agents have become actors in the P&L. Many finance leaders are now leaning into continuous forecasting, which provides a live view using statistical models, account-level evidence, and finance judgment. FP&A is evolving from annual budgeting to continuous operating decisions. You can now run complex scenarios live in the meeting.

6. Controls become a critical feature of the product. With software taking on more work in finance, an auditable, trusted chain from each of finance’s numbers to its source is critical. Finance needs to know where each number came from and how it was calculated, and where and which agents were involved. If an agent is drafting a journal entry, for example, it needs to keep the supporting work papers and evidence and follow approval rules before posting it. AI can still be used to reconcile records and flag and explain variances and exceptions, but finance sets the rules and owns sign-off. Having controls enables speed. The team can focus on the strategic work without having to reconstruct and check every step by hand.

7. Finance gets a bigger mandate. With real-time data and the ability to run scenarios quickly, finance is being pulled into many more strategic decisions. For many AI companies, compute is one of the biggest operating expenses and most consequential capital allocation decisions. Now finance teams can pull together a current view of usage data, infrastructure costs, and customer revenue, and AI tools can help determine how a new feature or an increase in customer usage would affect margins. That gives finance a stronger basis for weighing in on pricing, product launches, and spending commitments.

Companies are also making many more decisions where the CFO is a key arbiter. For example, how teams divide work between people and software is a key question across organizations and is a finance decision. With more granular data on the tradeoffs across time spent, token usage, software utilization, and more, finance can assess these tradeoffs. Build versus buy is part of this decision set. Building might now mean a big engineering project or just a non-technical team vibecoding a smaller tool. Buying might mean paying for usage or completed work, so costs change based on adoption. Now with AI, finance can model these more complex scenarios more quickly and accurately.

The role, rewritten

The CFO job has changed before. What feels different now is how much of the finance function and the org itself is being redesigned. The CFO historically built finance teams around layers of people needed to collect, reconcile, and review data spread across many tools, all set up to track and manage the work. Now the CFO, with a smaller team and greater leverage, can build and use the tools that do the work. Decisions that happened at weekly or monthly intervals can now happen continuously.

The CFO is becoming a builder and architect of the company’s operating system. That means figuring out what’s best done by the team and by software. The AI-native CFO has to be automation-first and know how to build, train, and resource a modern finance team. They also need new frameworks for what to build, what to buy, and how to calculate the ROI of those decisions. The upside is more time spent on the business itself and on being a strategic partner to the CEO. The job is still finance, but more of the job now is designing how finance gets done.

Thank you and for their help on this piece!

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.

Today’s models have absorbed more code, science, and human writing than anyone ever could, but knowing is not doing. A surgeon can read everything about an operation but still needs real, repeated time in the operating room to perform and master it. Closing that gap between knowing and doing is where much of the frontier is today, and more of it is happening through reinforcement learning: instead of learning from examples of the right answer, a model gets a goal, a sandbox, tools, and a grader, then tries and learns from what worked and what didn’t. This is an RL environment. The demand for them has skyrocketed over the last 18 months as labs are hillclimbing on real world tasks related to coding, research and computer use.

Building RL environments that actually work is harder than it looks. Models are relentless at reward hacking, finding shortcuts, and exploiting vulnerabilities: taking actions like finding and reading the answer key, rewriting the tests, crashing the grader, or quietly ignoring instructions the grader isn’t checking.

Every shortcut a model gets away with in training gets reinforced. Moreover, habits spread. Anthropic found that a model rewarded for cheating on coding tasks became more deceptive and willing to sabotage in completely unrelated situations. Smarter models find subtler loopholes, including ways out of their sandboxes and attempt to cover their tracks. Even the best environments have a shelf life, once a model masters a task, the learning stops.

Preference Model has focused on the domain that matters most to the labs right now: AI research and ML engineering itself. The industry is converging on the idea that the fastest path to more capable AI runs through models that can help build better AI: writing kernels, debugging training runs, curating data, designing experiments. If models can meaningfully accelerate the work of the ML engineers who train them, progress compounds. That makes machine learning engineering tasks uniquely valuable to labs, and uniquely demanding to build environments for. They’re long, open-ended, with countless ways to pass the test without solving the problem.

Over the past year, the Preference Model team has built RL environments for leading labs. Their focus has been on building the infrastructure to make harder, more resistant environments as models improve: tooling that finds where models are weak, generates new tasks to target those gaps, and tests environments against agents actively trying to break it.

This week they’re open-sourcing Karotte, the framework they’ve used in production to build those environments. Karotte bakes in strong defenses: killing stray processes before grading, rejecting files designed to crash the grader, and more. It has been hardened through more than a million evaluation runs and controlled red-teaming.

Jennifer Zhou knows this problem firsthand. As an early member of Anthropic’s team, she helped build the pretraining data infrastructure, the tokenizers and Claude’s pretraining datasets. Ning Cao was an early employee at DatologyAI. Both saw up close how directly data quality drives model quality, and they set out to solve it.

We’re thrilled to partner with Jennifer and Ning and the Preference Model team as they build the training grounds for capable and aligned models. The future models are being trained here, and we can’t wait to help them build the foundation!

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.

At the end of 2024, Texas’ grid operator had 63 GW of large new customers in its queue. By this June, that figure reached 474 GW, more than five times record peak demand, about 90% of it data centers.

Then Texas hit pause. What began as directions for data centers to pay for their own grid upgrades has since escalated to a freeze on all new permits. If you care about AI and American reindustrialization, this is important to understand. Why did it do this?

The upcoming election is part of it, but the underlying issues are worth digging into. The first is the interconnection queue. This pause is just as much about how projects apply for approval as it is about the projects themselves. Developers routinely spam requests across several sites, and many are speculative builds with no customer yet. Often these new developers have never even plugged in a GPU, let alone a power plant. So dealing with these low-quality submissions puts ERCOT, which runs most of the state’s grid, in a tough spot. Planners can’t tell which are real, and they don’t want ratepayers covering upgrades built for projects that never show up.

The second issue is community alignment. Noise, water, emissions, and power bills are top of mind for anyone who lives nearby. And regardless of the reality of these issues, people deserve straight answers when a data center comes to town, and they haven’t always gotten them (as of June, only 28 of 377 companies had answered a state survey on their resource use). In Hood County, commissioners were asked to support a tax waiver for “Project Patriot” without knowing who it was. To be fair, code names are common while companies shop for sites since a famous buyer can drive up prices, but that logic becomes tougher to defend once officials vote on tax breaks.

This broader political backdrop is important, but I’ll be focusing on just the energy side: why it’s so hard to power a data center, all the ways developers are trying to do it anyway, and where things are likely headed from here.

What connecting to the grid actually means

How Amazon purchases power helps illustrate what’s changed. For the last decade, they’d find a utility with network capacity, sign up as a large load, then enter long-term contracts to match their use on paper. These were often purely financial, meaning you didn’t always have to prove the power could reach you, and the grid connection was easy because utilities typically had spare network capacity.

That’s no longer true. The grid is stretched thin, and almost any new large load or generator now requires an upgrade. Power flows across every connected path, so buying from one plant doesn’t reserve a route to your building (there are even markets for transmission congestion rights).

So buyers first went after firm power (available around the clock) they could claim more directly, typically by restarting retired plants or by building next to existing ones. Neither was a true escape, and some of the loudest fights are over “colocation,” which ERCOT describes as drawing power from a neighboring plant before it reaches the grid:

• Susquehanna: Amazon bought a campus next to Talen’s nuclear plant, but FERC, the federal grid regulator, rejected Talen’s bid to send it more power directly. Utilities had argued it would let the campus dodge grid fees and push fixed transmission costs onto everyone else. Under the restructured deal, Amazon pays for delivery like any other customer.

• Freestone: CyrusOne’s 760 MW campus beside Constellation’s gas plant needed approval from the Public Utility Commission of Texas (PUCT), since it effectively takes much of the plant’s output from everyone else. The PUCT said yes in May, as long as the campus can cut its use or switch to backup power within 30 minutes of an ERCOT call.

• Armstrong County: Two Crusoe data centers, of 265 MW and 260 MW, share a wind farm that can produce at most about 265 MW. Both got the same rule as Freestone, so each must shed its entire load when ERCOT calls. Together, they'd cut nearly twice what the wind farm can actually produce, which Crusoe called excessive, but the PUCT kept it.

Interconnection studies are how grid planners identify the wires and substations that need upgrades for any addition to the network. ERCOT used to review studies individually under rules built for up to 50 large loads at a time, but 2025 brought 225 new requests by mid-November. So this June, the PUCT approved a new batch approach. Under “Batch Zero,” ERCOT studies large loads of 75 MW or more together and allocates grid capacity among them.

Primarily, the studies ask what happens when something breaks. NERC’s standard, the baseline for reliability across North America, covers a broad set of outage scenarios, with ERCOT adding its own requirements. Planners might simulate a transformer outage, then knock out a line or generator on top of that. Maintaining that reliability standard without shedding more load often means building additional infrastructure.

A newer risk is load unexpectedly dropping off the grid all at once. Many data centers switch to backup power at the first voltage dip to protect their hardware, so one bad fault can pull an enormous load off the grid in seconds. This happened earlier this year in Virginia, but was fortunately handled well.

Texas has even less room for error since ERCOT’s grid is largely isolated. Under some conditions, ERCOT can lose only about 3.2 GW of load at once before causing serious issues. A new voltage ride-through rule now requires new data centers to stay connected through routine faults, and I’ve even heard of labs running dummy jobs after a training run fails just to keep load from dropping abruptly.

All of this can sound overly conservative, but the system was designed to put reliability ahead of cost. Put simply, the grid is built for the hot summer days and frigid winters when failure can mean life or death. That duty is what makes sizing the grid so hard; you build for a few peak hours but pay for it all year.

So who does pay for all this resilience? Texas typically splits transmission costs by each large customer’s demand during the grid’s summer peaks, so a big load that ramps down on the hottest afternoons can skip much of its share of the transmission bill. In July, regulators proposed counting all 12 monthly peaks instead and charging large loads as if they ran at full size. In other words, large buyers would pay in proportion to how big their electricity pipes need to be, not how much is flowing through them.

Sounds simple enough, but splitting up costs is often the slowest part of interconnection. An upgrade built for one campus may also improve reliability for existing customers or make room for future growth that’s hard to value up front. And if any new development runs over budget or its load never shows up, everyone else is forced to cover whatever the developer’s commitments didn’t.

And those commitments are surprisingly cheap to make. The PUCT’s new large-load rules, effective October 8, charge a flat $100,000 study fee plus a $50,000-per-MW deposit. The deposit weeds out some speculative projects, but it’s a thin filter. ERCOT can reassign a project’s capacity if it falls two years behind, but even a project that loses its capacity forfeits just 20% of the deposit (~$10 million on a 1 GW campus), plus whatever the utility has already spent. Most importantly, the deposit typically only backs the upgrades built for that project, not the more expensive regional lines whose cost everyone on the grid shares.

Fully connecting a large data center can take 5 to 10 years. So to accelerate deployment, a phased connection approach is becoming more common, which gives the utility more gradual targets to plan around, as well as a way for the developer to prove they can handle everything they’ve asked for. All of this also assumes Batch Zero is moving, which it isn’t right now. ERCOT has even paused approvals for data centers of 75 MW or more to switch on, including 17 that had finished every other ERCOT step. Until the audit’s December report settles which projects are eligible, ERCOT can’t study them together, so the timeline is “TBD” and the load forecast is on hold.

What bringing your own power solves

For a developer facing the interconnection queue, skipping the grid entirely looks appealing. Idle GPUs cost far more than the electricity to run them, so speed matters most — what folks call “time to power.” That’s why developers are planning to bring their own power on site (“behind-the-meter”), but almost always alongside a connection to the grid, or as a bridge to one:

• Abilene: The Oracle campus runs on grid power with gas backup. Next door, Crusoe announced a 900 MW Microsoft campus with its own on-site plant, but the CEO of Lancium, Crusoe’s partner on the site, has described gas generation as backup.

• Shackelford County: Vantage is building a campus with 1.4 GW of compute for Oracle and OpenAI, designed to run off-grid with on-site gas. But its site plan includes a switchyard beside a high-voltage line — curious!

• Pecos County: Pacifico describes its planned microgrid (a behind-the-meter setup) at GW Ranch as never drawing from ERCOT, yet the site’s new owner, Amazon, says it’s designed to join the grid once it can connect.

• West Texas: Under a 20-year agreement, Chevron plans to build a dedicated gas plant beside a planned Microsoft campus that won’t initially touch the ERCOT grid, though Chevron has applied for a connection. Chevron now says the permit freeze could push its final investment decision into 2027, but it still expects first power by 2028.

• Armstrong County: The Google campus Crusoe is building beside a wind farm is tied to both the farm and the grid from day one, with Crusoe calling it “across the meter,” which I’ll admit is catchy.

SemiAnalysis is already tracking 75 GW of equipment orders for behind-the-meter assets, but every site I’m aware of intends to connect to the grid as soon as it can, chiefly because on-site power almost always costs more. So, once the grid is available, say in year five, you switch.

In the meantime, pairing on-site assets with even a partial connection is smart for the same reasons we built a grid in the first place. When one plant trips, the rest of ERCOT’s 1,460-plus generating units cover for it, but an islanded load (cut off from the grid) doesn’t have that luxury. The grid also provides things we take for granted, like inertia, fault current, steady voltage, and black start. An island has to supply all of that itself — power systems folks know how hard this can be.

AI workloads make the job even harder. At xAI’s first Memphis site, swings of 10 to 20 MW several times a second were wearing out turbine shafts until xAI added 150 MW of Tesla Megapacks. A grid ERCOT’s size dilutes swings like that, but on a private plant the turbines really feel it. Expect more batteries and other energy storage as rack-level power density grows and swings become more dramatic.

Then there’s fueling a site. Winter Storm Uri’s lesson is that gas plants can fail together. And since they tend to keep little fuel on site (NERC calls gas a “just-in-time” fuel), spare turbines don’t always help. Sometimes the pipeline doesn’t even exist yet. Solar trades that fuel risk for the sun and weather, and at gigawatt scale, you need a lot of batteries. Keeping a 1 GW campus running through one 14-hour winter night takes 14 GWh from batteries, about half of all the battery storage on ERCOT’s grid as of June. For an island seeking 100% uptime, covering rare events like a cloudy week or another Uri gets expensive.

Whatever the fuel, you want an island that fails gracefully and predictably, with no single point of failure. Even nuclear, about as reliable as power plants get, runs only about 92% of the time, mostly because each reactor goes offline for weeks to refuel. Redundancy in this case means effectively an entire second power plant. Thus, behind-the-meter setups favor modularity, like the more than 500 gas engines of about 4 MW each planned for Shackelford, though hundreds of engines can be a pain to maintain.

Full reliability for an island is very, very hard, but some labs and hyperscalers have shown that, forced to choose between reliability and speed, they’ll pick speed. Meta has turned to tents with no backup generators, and SemiAnalysis finds buyers growing more willing to accept outages, with some island designs aiming for as little as 99% uptime, or about 88 hours of downtime a year. With GPU time this expensive, that still beats years of waiting, so for many buyers a temporary island makes sense even if it’s messy.

A “private grid” that ties several plants and campuses together can take back many of the benefits a lone island gives up, at least in theory. But today Texas, like most states, only allows building your own power within tight limits:

• Supplying yourself: You aren’t a utility if you supply only yourself, your employees, or your tenants, and nobody resells the power.

• Running a private use network: You can also run on colocated generation, sell the surplus into the grid, and draw from it when you fall short. This is the model for the Armstrong County campus, and for the West Texas one once it connects.

  • Selling to a neighbor: If you sell to the factory across the road, you need a retail electric provider certificate.

  • Stringing a wire: Build your own line, and you’re probably running into the local utility’s service territory.

The Cato Institute’s consumer-regulated electricity proposal would loosen those limits by allowing private utilities to serve multiple customers across their own network. This isn’t an entirely new idea; Utah’s SB 132 lets loads of 100 MW or more contract for a fully off-grid system. Texas currently doesn’t let a network like this serve multiple customers, but if the demand for power remains insatiable, I’d expect the more permissive states to win larger chunks of the buildout with this “Wild West” utility structure. (However, you may also risk a utility “death spiral,” with the grid’s fixed costs falling on fewer and fewer customers.)

The flip side is a utility building the island itself. Outside ERCOT, El Paso Electric plans to put 813 small gas generators from ERock (366 MW in all) beside Meta’s new campus and run them as an island, on Meta’s dime, for up to five years. This is an option because, unlike the transmission and distribution utilities inside ERCOT, it still owns power plants. After the island period, it would connect the plant to its grid and could seek to spread the cost across all its customers, though in September administrative judges recommended approval only if those customers are protected.

Anything that runs on fuel also needs an air permit that matches how it operates, so a diesel generator permitted only for emergencies can’t run all the time. Optimistically, permits can come fast when things work. Sometimes they don’t, though. In Texas, the freeze now blocks them for data centers until the audit is done.

We should also ask what instances of behind-the-meter “bridge” gas are actually bridging to. xAI’s first two Memphis data centers answer that in different ways. The first ran temporary turbines off an existing gas main until the grid arrived, then began removing them. Along the way, it ran dozens without air permits. For the second, xAI built its plant across the state line in Southaven, Mississippi, but a July order requires all 69 turbines to retire by mid-2027 as a permanent 1.2 GW plant goes up in their place. One bridge led to the grid, and the other to a power plant of xAI’s own.

In August, though, the federal Tennessee Valley Authority (TVA) agreed to serve that data center directly, too. Turns out it’s hard to stay away from the grid!

What flexibility can buy

A campus that can keep itself running can also be easier for the grid to accommodate, even welcome. It’s a large paying customer whose demand can “flex” when power is tight, whether by cutting its draw or exporting surplus power. This is how xAI got approved for grid power at its second Memphis site. What made its promise to flex credible was its ability to carry its entire load for four hours on its own power, and what the CEO of Memphis Light, Gas and Water called “the world’s largest grid-connected battery system.”

Batteries are only one way to flex. A campus can also shift computing to other hours or data centers, or switch to its own generators. So how much room could flexibility open up? Tyler Norris and colleagues at Duke estimated that, setting transmission limits aside, ERCOT could add about 10 GW of new load without new generation if that load gave up 0.5% of its yearly grid electricity. Since the average cutback lasts about two hours, it’s also conveniently battery-shaped. Building on this, a study of PJM (the largest US power market) by Camus, encoord, and Princeton found that for each GW of new data center load, making 20% of that load flexible would save other customers $78 million a year, while bringing its own capacity for the other 80% would keep another $326 million off their bills.

For flexibility to be valuable, it’s important that it’s always available when operators need it. In Texas, generators already work this way under connect-and-manage. Put simply, they can hook up early as long as ERCOT can cut them down when lines are congested. Generators can live with that because at worst they sell less for a while, but a data center that has promised its customers uptime is more challenging. Some loads, like Bitcoin miners, have made the trade anyway. Batch Zero gives large loads two optional paths here:

• Bring your own power: A campus can count its own power plant toward its size as long as it can cut back within one minute if the plant fails. So far, 11 Batch Zero projects have picked this path.

• Agree to cuts: A campus can draw up to the full amount it asked for, but ERCOT can automatically cut anything above its guaranteed share whenever lines are full.

A campus could also pay its neighbors to cut their load instead. In PJM, Google is funding Voltus to pool up to 100 MW of batteries, thermostats, and other flexible devices across a territory. But PJM only counts what it trusts the pool to deliver, and it’s a capacity deal that helps the whole grid at its peak, not a fix for any one congested line. That’s a harder sell in Texas, an energy-only market where transmission is the main bottleneck. Texas has also barred colocated campuses from getting paid for similar services, since under SB 6 they already have to shut off when ERCOT tells them.

Even so, Texas has been a leader in distributed resources. One Base Power fleet, run with the co-op GVEC, passed ERCOT’s pilot tests to sell directly into the wholesale market on its first attempt, and is now expanding to 50 MW. These pilots are important; planners need that kind of proof before they’ll design around these fleets. But once trusted, distributed resources can rapidly add capacity without waiting for an expensive new “peaker” power plant, lines, substations, or a lengthy interconnection process.

Other hardware can help, too. Unlike legacy steel units, a solid-state transformer uses semiconductor switches, so software can measure and steer the power flowing through it. Alongside network upgrades like reconductoring (restringing lines with higher-capacity wire) and dynamic line ratings (rating lines for actual weather instead of worst-case conditions), that greater visibility and control can squeeze more out of wires ratepayers already funded. It’s a big reason we backed Heron Power, which is set to install its solid-state transformers at a West Texas battery site with RWE.

The problem is that most of these tools help operators keep things running day to day, but planners don’t always count them when they size upgrades. Nothing in physics forces that, though. A September study by Piq Energy, using Base Power’s data on potential fleets, found that about 80 MW of home batteries, strategically sited to relieve transmission constraints, could resolve all overloads triggered by a hypothetical new 100 MW data center near Fort Worth.

Batch Zero doesn’t consider things like this yet. It still plans upgrades for a flexible campus’s full planned load, since that path is a bridge to firm service, and there’s no option to stay flexible for good in exchange for smaller upgrades. Planners could instead size the upgrades smaller by crediting flexibility and other resources that relieve the same bottlenecks, assuming they’re measured in real time and perform reliably.

All of this saves time and money by getting more out of what’s already in the ground, but no amount of flexibility gets the grid out of building more generation and wires for all the demand coming down the pipe.

Getting to hundreds of gigawatts

On-site power and flexibility will decide how the next few campuses energize, but the labs and hyperscalers I talk to worry most about scale. Their power teams tend to split in two: one picks sites and equipment for the next couple of years, and the other asks how to connect hundreds more gigawatts after 2030.

This is a lot! Run, say, 100 GW all year and it’s 876 TWh, about a fifth of what the country used in 2025. Here’s what power developers told the Energy Information Administration (EIA) they planned to add in 2026 across the entire grid.

Every source helps, and much of it is headed to Texas anyway. But at last year’s average capacity factors (how much plants actually produce versus their maximum), the planned solar, wind, and gas plants would make around 150 TWh a year, or about a sixth of that 876 TWh.

To be fair, that gas bar likely understates what’s being built, since EIA’s survey only counts plants tied to the grid. In this way, much of the 75 GW of on-site power equipment already on order could be ghost capacity that charts like this one will miss. Still, gas remains popular because it generally runs whenever you need it. And because of that, the constraint is mostly getting the equipment in the first place, so buyers are turning to alternatives that can be easier to find, like reciprocating engines and fuel cells.

That said, gas feels like an incomplete answer to me. I’m no Greenpeace warrior, but running 100 GW around the clock at gas plants’ average rate would release nearly 8% of the country’s energy-related emissions. Often the easiest equipment to get is even less efficient, too. Moreover, operating such a fleet could seriously test our gas supplies.

Solar is compelling because it already has the production scale the labs are aiming for. The problem is that it’s mostly in China. The world added more than 600 GW in 2025, but China alone makes more than 80% of the world’s solar components and battery cells. From what I can tell, Chinese suppliers don’t mind selling to us that much, at least partly because they see our scale-up as “cute.” However, China reportedly weighed curbing exports of specific solar manufacturing equipment. Washington has also put on pressure, with forced-labor shipment holds, new tariffs, and a phaseout of wind and solar credits. Despite this, Elon is aiming for 200 GW a year of US solar manufacturing on his own, obviously solar-pilled.

Scaling firm power that isn’t gas is much more complicated. Uprates and restarts can squeeze a little more from the existing nuclear fleet, but the real upside is new reactors, as we’ve argued before. Meta and Amazon have signed big deals, but much of the capacity is still options and targets. So far, the military has been a stronger buyer to build microreactors on its bases, which is how factory-built reactors can learn to get faster and cheaper (blame EPC as much as the NRC). Geothermal could also leverage drilling (something Texas knows well) for repeatable power — Google and Meta appear quite interested.

Regardless of the power source, it all still ends up waiting on other equipment like transformers and switchgear, and all the crews to install them. Large power transformers now take more than two years to arrive, and the FCC has limited new foreign-made inverters alongside an August emergency order that could further bar Chinese-made equipment from the grid.

Someone also has to build the wires. In 2008, Texas regulators ordered the CREZ lines to carry West Texas wind, then spread their $7 billion cost across every ratepayer. Now they’re approving even bigger 765 kV lines, but it’s going slower than many would like. Some of that is just (unfortunately) typical construction, which is slow and expensive anywhere, but there’s also a myriad of additional regulatory hurdles on top. The federal permitting deal taking shape in Washington could help move things along if Congress can pass it.

Admittedly, I’m more confident that we’ll need a lot of power than I am about the exact shape it takes. My bet is that a handful of setups, depending on geography and flexibility, get built over and over. Maybe on-site gas and batteries carry a campus until its grid connection shows up, then stick around as backup to flex when the grid is tight. Solar gets layered on now where it fits, and geothermal and reactors come in once they prove out.

Past 2030, it’s even harder to say who ends up building and owning all that power. One answer is that the same company builds both the plant and the campus, which Google’s purchase of Intersect may signal. Another is that oil and gas companies, like Chevron or Williams, become broader grid builders, and it isn’t hard to picture them, or “neo-utilities” like NRG and NextEra, building private grids that serve several campuses, assuming the law enables it.

I don’t know which way it goes yet, but they’ll likely all be buying from the same equipment makers. That market is huge and surprisingly ill-equipped to meet inflecting demand. Given that, I see two major ways for startups to break in:

• Integration: Some take familiar hardware and win on integration or business model, the way Base Power runs home batteries as a trusted aggregated resource.

• Technology: Others bring new, superior technology that early adopters will take a chance on in a constrained market, which is what Heron Power is aiming to do with solid-state transformers.

A third pitch, crudely put as “worse but faster,” sells well in a shortage and can be immensely profitable, but I’d ask what those profits are being reinvested in, because it might get hard to compete if broader supply catches up.

Indeed, shortages like these are an opening for startups, but it helps to understand why incumbents aren’t quick to fill them. The last bet on a turbine boom ended in a $22 billion GE Power write-down and helped cost GE’s CEO his job. A startup has to survive the busts incumbents are planning around, as well as compete globally with Siemens Energy, Mitsubishi Power, and all sorts of suppliers in places like India and China. Customers may pay for speed today, but keeping them will take reliable, competitively priced equipment and a service team that knows what it’s doing.

Every step of site development is hard. Vendors like GE Vernova now take nonrefundable deposits just to reserve a manufacturing slot, leaving developers with a chicken-and-egg problem. Lenders want a long-term contract with a solid customer, who wants a credible timeline, and that timeline takes deposits the developer usually needs lenders to fund. Bring in an unproven vendor and the loop gets even harder to close (and markets notice fast when it breaks). Once it closes, everything else still has to go right.

For a new vendor, even getting into a campus’s first phase is difficult. As far as I know, hardly anyone has placed a large, firm order for data center power equipment without thousands of hours of operation. For example, FTAI’s big order rests on the CFM56, a jet engine that has logged more than a billion flight hours, while Crusoe stepped back from Boom as Boom’s first engine core was still gearing up for tests. So even credible teams building awesome technology face a cold-start problem. Thus, getting your product designed into a real project matters far more than early deposits or revenue.

After the pause

Let’s get back to Texas. I think it’s fair to check that projects in Batch Zero are what their developers swore they are, and credible projects should move ahead soon. The broad permit freeze is harder to defend, though I see the state’s logic if the goal is to approve nothing until the audit is done. That said, I don’t like that it makes a developer that has funded its first phase, even one bringing its own power, wait like a speculator. Texas should narrow it now by exempting generators permitted only for emergencies and projects that have funded their first phase. The October 19 update to the governor from the state’s environmental regulator is the obvious place to start this conversation.

Once the audit is done, the ongoing tests can be more straightforward. A developer that posts its security deposit, pays for the capacity it reserves, and hits phased milestones should get a connection date it can plan for. It should also connect sooner if it agrees to cut back when the grid is tight (Batch Zero already does some of this). Going further, the utility building upgrades should probably answer for delays much like a developer does. Perhaps Texas could open lines to competitive bids with cost caps and penalties.

Whatever Texas decides, power will likely stay tight for years while the bottleneck keeps moving. Importantly, data centers are not that unique in the equipment they need; they just hit these limits first. In this sense, they’re the perfect rehearsal, because much of what we want to (re)build in this country will run on the same stuff.

That also makes it an opportunity I’d hate to waste. AI companies will pay nearly any price for power and can build almost anywhere, so they can help fund upgrades the grid needs anyway. Analyst Hans Royal estimates inference could pay an absurd $5,600 per MWh for power and still earn a decent return, nearly 60 times what the average US industrial customer pays. That can strain local prices for gas, power, labor, and materials, but the infrastructure it funds can be worth far more. Admittedly, that argument can be a hard sell right now. In an August poll, 56% of Texas voters said more data centers would hurt local energy bills.

The worry is fair since the honest answer is that it depends on who pays. In a perfect world, developers cover all related upgrades and pay for the capacity they reserve, supporting shared infrastructure construction that improves the grid for all of us. Even just by buying a lot of power, a campus spreads a service territory’s fixed costs over more sales, shrinking everyone else’s share. Berkeley Lab found that from 2019 to 2025, the states with the most load growth generally saw average prices fall after inflation. Imagine all the money racing into AI helping the next factory connect and the next household electrify without an unaffordable bill (or a grid too tight for an EV, or a home robot). That household may never open a chatbot and still come away far better off. We’re already seeing early versions of this:

• Indiana: The utility serving Amazon’s New Carlisle campus proposed using data center revenue to cut household bills by about $100 a year.

• Alabama: A Tuscaloosa County campus will pay $270 million in community benefits over 20 years. In return, it gets about $314.5 million in tax breaks, though school taxes (about $131.5 million) aren’t abated.

• Pennsylvania: One developer offered every household in Hazle Township $10,000 if its campus gets approved, which some residents (perhaps correctly) called a bribe.

I would not be surprised to see hyperscalers covering a whole town’s power bills as part of hosting a local campus. Most states, Texas included, don’t let a utility single out one town, but at a small utility where data centers use most of the power, like Oregon’s Umatilla Electric, covering a typical household’s bill for each of its roughly 17,000 meters would only cost about $30 million a year. Such an agreement could pencil because that’s less than 0.5% of what a 1 GW campus costs to own and run.

It’s also important to read the fine print when deals are made. In Arkansas, Google agreed to pay $443 million up front toward an Entergy solar plant for its data center. The catch is that Entergy counted it as prepayment for power, so it can still seek the plant’s full cost plus a return from all its customers. And Entergy sued two newspapers to stop them from reporting on the contract, then dropped the suit after a judge refused. Even if the deal is legal, this isn’t how you build trust with the community.

A campus that levels with its neighbors and actually follows through keeps people on its side, and the next one gets easier to welcome. Get that wrong and everyone else can end up paying for it. In PJM, the market monitor says data centers account for 38% of the latest capacity bill, some $6.3 billion. Let me be clear: none of that justifies a blanket pause. But it does mean getting the rules and incentives right as we scale up development. AI companies have pledged to pay their way, and most would rather connect to the grid than avoid it. It seems ideal for everyone that we let them do so instead of driving that spending into private islands, or pushing the infrastructure upgrades that we’ll need anyway onto everyone else’s bills.

Texas created the energy fast lane first; now it has the opportunity to show us a better one, and I expect it will. Power is where America finds out whether it can still build.

If you’re working on energy problems, I’d love to learn from you. Please reach out.

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.