America | Tech | Opinion | Culture | Charts

September 11, 2001 was 25 years ago, today. Never Forget.

The brain of a venture capitalist can be succinctly summarized as:

In venture, no one talks about the mistakes of commission: investing in companies that didn’t return as promised. There’s no scoreboard for how many of your checks get buried behind the back of the shed - it’s part of the job. In fact, our losses on investments equate to just 15 cents for every dollar invested across the funds we manage. But a single winner can have an outsized impact. Our stake in Databricks alone represents 20% of our total AUM.1

There are other kinds of mistakes that we do consider fatal. Investing in the right category, but in the wrong company. Worse yet, choosing to sit out on a company like a SpaceX, an OpenAI, or a Stripe. As my partner Marc Andreessen reminds us: if you invest in a company that unfortunately goes belly up, at least, at a certain point, you stop hearing about it. If you pass on a company and the company succeeds, you never stop hearing about it–you see it constantly in the press, on X, and in your nightmares. It haunts you every single day and every VC can recount those misses more than the successes. Our brains are wired for it.

The irony, however, is that the mistake LPs may be making today is one of commission, driven by a seemingly prudent question: Is my portfolio appropriately balanced across asset classes?

A “cottage” no more

Every week, I come across an LP who has nearly zero exposure to the three current frontier models (SpaceX, Anthropic and OpenAI). And they don’t have exposure to the leading open source ones. So how did so many LPs miss this initial wave of AI?

A few hypotheses. To start, venture allocations were historically “right-sized” to reflect the “cottage industry” nature of the business. Venture is typically a subset of a private alternatives allocation and targeted to be 5-10% of an overall portfolio. For endowment LPs, historically the most long-dated and aggressive allocators to venture, their exposure can be somewhere between 25-40%. And for some, that’s mostly through NAV appreciation vs active primary $s deployed to that percentage.

But the day SpaceX went public, historical asset allocation went out the window. Technology is the dog that caught the bus and at ~$2T market cap today, SpaceX is the largest venture-backed IPO by a factor of 10x2 and right around the corner is Anthropic (valued at $965B, rumored to be going public shortly at $2T) and OpenAI (most recently valued at $852B). That’s roughly $3.8 - $5T of combined equity value cultivated largely in the private markets and for those LPs who don’t have exposure to it, they have a lot of ‘splaining to do to their boards and investment committees.

Moreover, the biggest companies in the world are technology companies and they are getting bigger, more quickly than ever before. What was once success in venture (a unicorn) is now your average Wednesday and private companies are generating $1B+ in revenue on a timeline measurable in quarters.

While this is happening, the asset allocation frameworks LPs have relied on for decades have started to fracture. I’m reminded of something Robert Smith, CEO and founder of Vista Equity Partners, one of the great software buyout funds, famously once said: “Software contracts are better than first-lien debt”.

It turns out, the next wave of disruption can make recurring revenue less durable than expected. And we may be seeing this reflected in buyout returns, where average and top quartile returns are near 15-year lows.

For prudent institutional investors, who have historically “right-sized” their venture exposure to a small piece of their overall alternatives pie, the landscape has changed, and this is a call to resize again.

Power Law is now the shape of our economy

It’s hard to overstate how a small number of companies are now capturing a monstrous share of everything. NVIDIA chips are now even being used to broker peace deals. Take a look at the evolution of the S&P 500 below, where 8 out of the top 10 most valuable S&P 500 companies are tech companies (and turnover in the top 10 is more dynamic than ever before):

Even in credit markets, tech is the main event now:

So, tech is getting bigger, and more central to the economy, and as everyone knows by now, companies are staying private longer.

~1/3rd of technology companies valued $150B+ are privately-held and for the second-largest public tech co (SpaceX), the lion’s share of enterprise value creation happened on the private side.

Further down, the companies with real momentum are moving up at an accelerating pace: median time between rounds for actively-raising unicorns compressed to just one year in Q1 2026, down from 1.5 years in 2024, while over half the private companies that once cleared unicorn status (51.2%, per PitchBook) haven’t raised in more than two years.

The fact is that tech companies are getting larger, especially at the tail, and that growth is happening in private markets:

In the good ol’ days (2004), Google went public with a ~$1.67B offering and closed its first day at a “bubbly” $27B market cap. Adjusted for inflation, that’s roughly $3B and $48B, respectively. Today, that offering wouldn’t crack the 20 largest completed VC financings recorded by PitchBook since January 2025, and its first-day market cap would amount to just a third of those companies’ $75B median reported post-money valuation. In other words, this would be a rounding error in the news cycle.

When venture eats the world

PE has historically captured a much larger share of alts allocation, but venture outcomes are simply changing the math:

Private equity and venture capital have maintained roughly equivalent exit counts for a long time, but exit value for VC-backed companies just tripled PE’s high watermark.

The largest, majority-PE-owned company to go public was Medline in December 2025.3 $54 billion based on first-day closing market cap. SpaceX, the first of the new AI generation, private juggernauts stepping onto the public markets, closed its debut at approximately $2.1 trillion—roughly 39 times as much.

What about M&A? In July 2026, Aligned Data Centers became the largest, majority-owned, private equity backed M&A exit by enterprise value at $40B (according to Pitchbook). The biggest venture-backed M&A exit? Cursor to SpaceX at $60B.

A single venture IPO (SpaceX) can be roughly ~39x the largest PE outcome on record. Multiply that by 2 more $1T+ potential IPOs in the pipeline, and you can see where I’m going with this.

Case in point: consider CalPERS. In late 2022, the largest public pension fund in the U.S., reworked its private equity book. Buyouts fell from 91% of new commitments in FY2020-21 to 58% in FY2023-24, while growth equity and venture together rose from 9% to 43%. Its PE program went from ranked 30th to ranked 1st among the 30 largest US pension PE programs in just three years, and their growth equity and venture program is now its single strongest-performing segment.4

Other public pension plans are starting to follow suit, including MassPRIM and North Carolina Treasurer, just to name a few.

Machine intelligence compounds the trend

If nothing else, we can all agree that the train has left the station on machine intelligence. We have already seen exponential improvements in capability, autonomy and cost—and this is all before robotics really arrives. Innovation will likely continue to expand faster than any of us, investors or LPs, can map in advance.

Fortunately, this is the kind of ambiguity venture is designed to thrive in. Unlike private equity, the power law in venture absolves all sins. We don’t need to save every portfolio company to generate great returns and in fact, while we don’t invest with this expectation at the outset, the majority of the portfolio can go to zero and we can still generate outperformance from a small handful of power law winners.

The legitimate case against venture

So if the case is so clear for venture, why isn’t everyone making a drastic change to their portfolio?

Show me the money

The critique of venture that hits hardest is, of course, liquidity (even if that’s not unique to venture). As one of our partners once said to me, “I’m knee deep in TVPI, but where is DPI?”

For some managers, secondaries may be the only answer to post up liquidity and stay in the game between fundraises. It’s a fair strategy, but remember, VC firms typically generate their returns from a small batch of winners, so you might be pruning your roses to water your weeds. Even still, secondary capital is still a rounding error in the scheme of things and the vast majority is targeting a very small subset of companies. And there is absolutely no discount for the assets you actually want:

In my experience, most LPs would prefer to hold compounding assets. An example I’ve shared publicly with Packy McCormick before: a few years ago, we offered our LPs liquidity in our older vintages. Fund I had a seed position in Stripe and Fund III invested in the first round of Databricks.5 We offered various structures to give LPs liquidity on those names, and do you know what happened? Not a single LP chose to sell. Ironically, receiving money back is sometimes a burden for LPs because then they have to reinvest the capital. Can they find a comparable asset that can continue to compound at that expected rate? Even if they can, why pay interim taxes instead of letting it grow?

Returns over marks

Here’s another one we hear all the time: venture valuations are too opaque. And how do LPs calibrate to a legacy portfolio of companies from the last technology cycle that may be worth less than what their last round valuation was?

Somewhat fair, but let’s unpack. Venture marks are kept close to the vest because, rightly or wrongly, they’re often viewed as a vote of confidence (or not) in one company over another. The mark is a signal that affects both relationships with founders, and founders’ relationships with their employees and prospective hires. The reality is that the valuation process for all private assets is complex. Almost every single venture manager holds the same company at different valuation, which makes it even more confusing. As this oldie, but goodie from Scott Kupor posits, when is a “mark” not a mark? At the end of the day, the only true mark is cash in the bank account after an exit. DPI speaks for itself.

Plus, this phenomenon isn’t unique to venture. Private equity and private credit aren’t “mark-to-market” either, and now that AI is re-writing the leaderboard for enterprise value, there may be a whole lot of pain in PE.

Similarly, private credit was the coupon-like antidote to PE’s illiquidity. ‘Equity-like returns, with debt-like downside,’ they said. There’s nothing wrong with that, and credit where it’s due (pun intended), the fundamentals have largely held, and the asset class is evolving in real time, moving up the stack to underwrite the AI infrastructure buildout itself.

But for those earlier cohorts of loans, especially to supposedly uncorrelated PE-backed software companies, where AI has compressed multiples, growth has slowed, rates have moved higher, and refinancing is around the corner? There likely will be (and likely already have been) write-offs, and likely no power law of outlier upside to make up the difference, either.

Platform shifts may keep the other guys up at night, but for venture, it’s where the opportunity is.

Proximity is not participation

My last point is crucial because it’s often a trap for allocators who are feeling the weight of the argument so far. Too often, the folks I speak to feel like they are already overexposed in venture, with decades of legacy portfolios sitting with managers who have zero incentive to sell or re-value their book.

But, in continuing the theme of power law, it’s not just a smaller number of companies generating the returns. It’s a smaller number of funds with exposure to those companies:

Once again, the fat-tail of outcomes is happening within a small group of companies held by a small group of funds.

Top-quartile 5-year IRR is now double the median for the most recent vintage:

Even more recently, the top decile funds by IRR have never been farther away from the top quartile:

Manager dispersion is becoming even more pronounced.

And it’s not just paper returns–top decile’s got your DPI, too:

Top decile DPI for nearly 10 year-old funds is 2-3X larger than the top quartile (and 4-10X larger than the median). Median just isn’t good enough. Top-quartile ain’t it either.

It shouldn’t be any surprise that a small handful of funds are capturing an outsized share. After all, winning begets winning: a fund that lands one behemoth gets the follow-on rights, the founder referrals, and the information edge that make it more likely to land the next one. Access compounds the same way outcomes do.

In the bigger scheme of things, diversification is something we’ve all been taught to be prudent. But this is perhaps one of the few times where that logic doesn’t land. In a recent podcast, Aram Verdiyan from Accolade Partners revealed that based on Accolade data, 3,000 venture capital firms in the US, and only 20 have achieved consistent 3X net returns over the last two decades. Furthermore, a September 2026 PitchBook screen of 2,143 global VC funds with 2000–2018 vintages and reported DPI revealed that, only 365—or 17.0%—had returned at least 2x invested capital. Just 143 funds, 6.7%, reached 3x, and only 51, 2.4%, reached 5x. Put differently, 83% of funds failed to return 2x and nearly 98% fell short of 5x. So it’s worth saying out-loud that only the best funds have the access, the power, or the muscle to repeatedly find, nurture and exit a behemoth.

Ripping it to studs and starting over again

Institutional money isn’t missing the argument for venture. Many LPs are making this exact argument internally to their boards and investment committees themselves. Tech will drive an outsized share of value-creation, before maturing into some of the largest businesses the world has ever known. As Marc says, “technology is the dog that caught the bus”. But if left unchanged, “prudent” asset allocation to a “cottage” industry that has outgrown the cottage, may miss the bus altogether.

Subscribe now

The above reflects the subjective views and beliefs of a16z Capital Management, L.L.C. (“a16z”) or its respective affiliates, and is not, and is not intended to be, an offer to sell -- or a solicitation of an offer to buy -- any securities, and may not be used or relied upon in evaluating the merits of any investment. It should not be construed as or relied upon in any manner as investment, legal, tax, or other advice. Any projections, estimates, forecasts, targets, prospects and/or opinions expressed in these materials are subject to change without notice and may differ or be contrary to opinions expressed by others. Any charts provided here or on a16z content distribution outlets are for informational purposes only, and should not be relied upon when making any investment decision. Certain information contained in here has been obtained from third-party sources. While taken from sources believed to be reliable, a16z has not independently verified such information and makes no representations about the enduring accuracy of the information or its appropriateness for a given situation. All content speaks only as of the date indicated. There can be no assurances that a16z’s investment objectives will be achieved or investment strategies will be successful. Any investment in a vehicle managed by a16z involves a high degree of risk including the risk that the entire amount invested is lost. 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.

America | Tech | Opinion | Culture | Charts

Your company chooses your health insurance, and you have to live with it. For the first time in decades, one of healthcare’s stickiest markets is opening up as premiums keep rising, consumers expect more, and AI makes a different kind of health plan possible.

Employer-sponsored (AKA “commercial”) healthcare is peculiar - the buyer and user are different people. The company you work for chooses one or two health insurance options and subsidizes most or all of the premiums. Healthcare is typically the company’s second largest expense after payroll, while you cover co-pays, deductibles, co-insurance, and other out-of-pocket costs. And because you change insurance plans when you change jobs every few years, the incentives aren’t long-term aligned. The plan paying for your care today may not benefit much from keeping you healthy years from now. It’s one reason employer-sponsored insurance has been called healthcare’s original sin.

But it’s a massive market, with $1T worth of spend across 150M+ Americans and highly sticky products. Employers have kept renewing contracts with legacy insurance providers, even as premiums rise by 10% or more every year. Much of that price inflation has come from the rising costs of healthcare itself, including specialty drugs and administrative bloat, and a market with historically high barriers to entry. At the same time, the consumer experience continues to be expensive, opaque, and full of surprise bills.

That has been the status quo for decades. But now, three things are happening at the same time:

• First, employer healthcare economics have risen to the point that the majority of employers are shopping for alternatives, either to switch to lower cost health plans or drop traditional health insurance entirely.

• Second, consumers have an increasingly high standard for care when it comes to access, cost, experience, personalization, and quality. This is partly the result of abundant DTC and now AI-native healthcare services that have normalized consumers seeking care directly and paying out of pocket, then asking insurers and employers if they’ll cover or reimburse it.

• Third, AI is lowering the fixed-cost barrier to building and operating a health plan. Supporting plan members, helping with care navigation, underwriting risk, reviewing claims, remitting payments - all of these functions historically required large teams to execute, but AI is now the leverage to do all of these more efficiently.

These shifts are spurring a generational replacement cycle in commercial health insurance, in which employers are increasingly looking to move to Alternative Health Plans, or AHPs.

A new wave of challenger health plans, challenger PBMs, and modern infrastructure platforms have emerged to meet the demand for better employer-sponsored healthcare. Their challenge and opportunity is to serve their two audiences at once - improving employer economics and risk while meeting the higher standard employees now have for their healthcare experience. Many of these emerging companies have the opportunity to be both AI-native and AI-proof.

And the plans that replace the traditional options might look pretty different. If we look at the various ways challenger health plans are trying to compete, a common theme is that plans are becoming more upstream, personalized, and proactive. They’re becoming more involved in how consumers actually navigate seeking and receiving care instead of just financing and administering it. AI helps enable the evolution from passive to active across these dimensions.

So this is a generational moment to rebuild the commercial health plan. But in an increasingly crowded market of startups, the question is not whether there’s a compelling “why now” - it’s which of these levers, or combination of them, can become a real, durable advantage.

How Challenger Health Plans Can Compete

A challenger health plan might get attention with a superior customer experience or lower admin fees, but the ultimate test is whether it can keep its members healthier and control the total cost of care. Across the health insurance plan ‘stack,’ there are a few places where an upstart can build a meaningful advantage:

And of course, a company can integrate several of these components into a full-stack health plan offering. Given the massive market opportunity, any one of these components could support a standalone company, or become a superpower within a full-stack health plan. And each segment of the employer market, from SMB to jumbo, is a big enough prospect base for a company to focus there for years and still grow significantly before expanding beyond.

Rising costs are forcing employers to shop, just as AI is making health plans smarter and cheaper to build, and consumers are expecting more from them. Health plans have to compete again, and a previously impenetrable market category is now addressable for startups.

Commercial health insurance is one of the largest markets in healthtech. It’s a control vector for how healthcare is priced, accessed, and experienced, and it’s one of the hardest layers to rebuild. For the first time in decades, employers are ready to buy something fundamentally different - and builders have better tools to deliver it.

If you are building a challenger health plan, we’d love to hear from you. Let us know the one or two levers where your approach is structurally different, and why that difference compounds as you scale.

Special thanks to Colin Morelli, Seth Cohen, and Eva Steinman for their contributions to 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.

America | Tech | Opinion | Culture | Charts

Every consumer brand shares a quiet, expensive problem: they make more product than they sell. This is built into the model - brands forecast optimistically to hedge against stockouts, and end up sitting on trillions of dollars of perfectly good inventory. Even the most in-demand brand inevitably ends up with a color or SKU it can’t fully move.

What happens to that inventory is one of the least glamorous stories in consumer goods. It gets stranded in warehouses, quietly written off, or dumped into off-price and liquidation channels. Aside from the hit to profit, this erodes the very brand equity these companies spend years and fortunes building. And, it forfeits opportunities to reach new customers, on channels that could actually drive net new sales and loyalty.

For decades, the only options were bad ones. Sell product through discounters and jobbers and watch it resurface somewhere that cheapens the brand, or destroy it outright. Many of the best brands chose destruction rather than see their inventory end up in a discount bin. No one had built a brand-safe way to move excess inventory - one that treats it as an asset to recover value from, rather than a problem to bury.

There’s a reason for this. Excess inventory is the opposite of a clean commodity: it’s messy, heterogeneous, one-off lots with no comps, no standard pricing, and a different right buyer for every batch. Pricing and matching that supply at scale was a problem you could only throw humans at. This is why the category defaulted to relationship-driven liquidators and race-to-the-bottom auctions. That’s exactly the kind of unstructured problem AI is now good at.

We believe the winner in this market will be the platform that is AI-native and brand-safe from day one. Highstock is building exactly that, and today we are thrilled to announce that a16z is leading Highstock’s $30 million Series A. Olivia will be joining the Board, and a16z marketplace legend Jeff Jordan is also participating as an advisor.

Highstock is an AI-powered B2B marketplace that connects brands sitting on excess inventory with a vetted network of global wholesale buyers. Brands move overstock discreetly, without the fire-sale optics, while Highstock handles the hard parts: matching, logistics, compliance, payments, and even managing order paperwork. Its models do the heavy lifting so both sides can browse and transact with confidence.

Highstock is hard to replicate because of its team. CEO Camille van Horne and CTO Sean Cashin both came from Instacart, where Camille spent six years as a Director of Product Management leading the end-to-end fulfillment experience. Running fulfillment at Instacart, especially during COVID, was marketplace operations at its most unforgiving - matching supply to demand under messy, real-world constraints, at scale, where every failure is visible to the customer. Camille and Sean bring this level of precision and excellence to Highstock.

In a little over a year, that approach has produced real scale. Highstock now has more than $1 billion in inventory listed on the platform and works with over 100 major brands. Roughly 60% of transaction volume comes from brands and conglomerates that do more than $500 million in annual revenue. On the other side of the marketplace, more than 100 business buyers reach that supply through channels most brands could never access on their own, from live sellers to international e-commerce players. Along the way, Highstock has kept more than 10 million pounds of product out of landfills.

Beauty and personal care is the beachhead, but the opportunity is much larger. The same excess-inventory problem exists in apparel (which Highstock is launching now) and across consumer goods. As Highstock expands category by category, the data and the relationships compound with it.

We could not be more excited to partner with Camille, Sean, and the entire Highstock team as they give excess inventory a second life. If you are a brand sitting on excess inventory, or a buyer looking for supply, check out Highstock at highstock.com - and if you want to help build it, they’re hiring.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

Fifteen years ago, I wrote that software is eating the world - at the time, a controversial claim. The main prediction of the essay has come true. As a simple measure, technology’s share of the market capitalization among the ten largest businesses in the world has gone from 31.5% to 94.4% over that period. Amazingly, this happened even though only 1 in 300 people on earth wrote software, and every line of it had to be written by hand. Software has been eating the world at the speed of human hands. It is about to eat the world at the speed of compute.

Three years ago, when Scott Wu and his team began building Devin, the term “coding agent” barely existed. Today, if you talk to any software engineer at a leading company, they’ll tell you that agents write almost all of their code. In the past year, Devin has gone from writing 13% of Cognition’s production code to more than 90%.

When agents write 90% of the code, engineers can literally do ten times as much. The 10x engineer becomes the 100x engineer, as they shift from writing artisanal code to operating as the CTO of a fleet of agents. The 1000x engineer isn’t too far behind. As Scott puts it: “Within our lifetime, engineers will go from bricklayers to architects, focusing on the creativity of designing systems rather than the manual labor of putting them together.”

At Mercedes Benz, engineers turned what would have been an eight month long COBOL migration into 8 days of work with Devin. Rivian teams increased their test generation velocity by 10x. Devin triages and patches vulnerabilities across thousands of repos at some of the world’s largest financial institutions like Itau, where 70% of security vulnerabilities are automatically remediated by Devin.

Every time programmers get more leverage, doomers predict the end of software engineering. Compilers were supposed to shrink the profession. So was open source. So was the cloud. Instead, each leap made software cheaper to build, and demand for software, and engineers, exploded. It keeps happening because, as Milton Friedman observed, human wants and needs are infinite, so economic demand is infinite, and job growth can continue forever (see The Techno-Optimist Manifesto). Technology is how we unlock that demand. Technology gives us leverage, lowers the cost to produce, and makes previously uneconomical use cases economical. For software, Devin provides that leverage.

It is tempting to think of this as just another application of AI - one vertical among many. But if software continues to eat the world, the automation of software itself accelerates every industry.

This future is core to how we at a16z see the world. We have always believed that software expands markets, that technical founders with seemingly unreasonable ambition create the future, and that small teams armed with powerful tools can change entire industries. Cognition is a powerful expression of those beliefs: a deeply technical company taking on an enormous problem, moving at extraordinary speed, and building for a market whose ultimate size is difficult to comprehend.

Our conviction in Cognition comes from personally knowing Scott for over seven years. We first backed him at Lunchclub, where he and his team built a system for bringing people together and created relationships that have lasted through the years since. Scott is one of the most accomplished competitive programmers of his generation: a three-time gold medalist at the International Olympiad in Informatics and a world champion at 17. But after working with him for many years, we know that his technical brilliance is only part of what makes him exceptional. Scott has an unusual combination of speed, judgment, competitiveness, and leadership. He sees important technical shifts early. He holds himself and everyone around him to an extraordinarily high standard. He cares deeply about his customers, recruits exceptional people, and has built a culture that moves with the urgency of the opportunity. One of Cognition’s values is “every second counts.” They mean it.

An infinite market demands infinite ambition. Scott’s ambition and vision match the scale of the market and we are proud to back him for the second time. Scott, Steven, Walden, Russell and the entire Cognition team are working toward a world in which anyone can turn an idea into software.

I ended my 2011 essay with one line: “I know where I’m putting my money.” Fifteen years later, I still do.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

Every company runs on a CRM, and almost no one likes the one they have. It’s the system of record for how you find, win, and keep customers—the most important software a company owns, and somehow the most dreaded. CRMs are supposed to help drive productivity and efficiency for sales and support teams, while helping deliver insights on how to improve. It turns out they often do the opposite, and the promised “insights” are a stack of dashboards you have to build yourself.

Enter Lightfield, the CRM for the agentic era. That means a fundamentally different data architecture: a flexible data model that needs no configuration, builds itself from your emails and calls in minutes, and fits your business rather than the other way around. The feat underneath is a temporal context graph that captures the why behind every deal. The advantages compound from there: agents and humans work from the same platform through the same API, and the data model scales to track complex, high-touch, multi-threaded enterprise deals that play out over months without a single schema change. The classic CRM headaches, the stale fields, the weeks of configuration, the reports nobody trusts, go away because the data model is finally right.

More than that, Lightfield comes with agents that do the work that piles up around the record, the updates, the follow-ups, the handoffs, the reporting, and pricing follows the work done and value delivered instead of static seats you buy. And because it holds every customer conversation, Lightfield becomes the business world model, shaping what you build next and for whom. In a world where context is the rate limiter, Lightfield delivers a compounding context moat.

Customers have noticed. Since launching late last year, thousands of companies have adopted Lightfield, and teams are ripping out Salesforce, HubSpot, and Attio to switch. As one operator put it after moving off the largest incumbent: “If you’ve only ever ridden a horse, the only thing you ever want is a faster horse. Well, this is a car.”

Special products are built by special people. We’ve known Keith Peiris and Henri Liriani for a long time. They previously built Tome, an AI presentation tool that grew to 25 million users, and when they decided that wasn’t the right strategy to build a massive company, they did the hardest thing founders can do: they pivoted. They shrank the company to a core group of engineers and rebuilt around a problem they had experienced firsthand scaling Tome—there was no AI-native CRM for growing companies.

History says this is how generational systems of record are born. Siebel was born on client-server, Salesforce on cloud, HubSpot on inbound SMB. Each won a greenfield generation of companies the incumbent couldn’t see, and the AI-native generation is being founded right now, choosing its system of record for the first time.

We could not be more thrilled to lead Lightfield’s $47M Series A. If your company is growing fast and the tooling that got you here is starting to break, check them out. And if you want to help build the system of record that humans and agents actually love, they’re hiring!

Subscribe now

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

America | Tech | Opinion | Culture | Charts

The U.S. needs more long-range missiles.

In the first five weeks of the war against Iran, the United States fired more than a thousand Tomahawk cruise missiles. It did this against Iran: a mid-sized regional power.

The West needs to be prepared for wars against countries such as Russia and China: vast territories with thousands of hardened targets, layered air defenses, and the will and capability to fight for years.

The real challenge is not replacing the weapons that have already been consumed. It is building a new arsenal capable of overwhelming scale.

Europe is in no better position.

Europe’s long-range strike gap is even more pressing. In May 2025, Germany and the UK committed to develop a deep precision strike weapon under the Trinity House Agreement. France, Germany, Italy, Poland, Sweden, and the UK have banded together in ELSA, the European Long-Range Strike Approach, to field a long-range ground-launched missile. And in July of this year, a host of European countries pledged more than $50 billion to build deep precision strike capabilities.

These initiatives are excellent, but they are years away from results.

Enormous distances separate European soil from war-winning targets in Russia, and today, Europe has almost no sovereign way to reach them. If Europe wants to deter, it needs the ability to strike.

Cheap mass transformed the tactical fight but left the strategic one untouched.

When Deputy Secretary of Defense Kathleen Hicks launched the Replicator initiative in 2023, she made affordable mass a mission: thousands of small, cheap, autonomous systems to offset China’s greatest advantage – its sheer numbers. Arming the West with low-cost drones and munitions is one of the most important defense shifts in a generation.

But affordable mass has mostly reached the small end, and while small, tactical drones are deadly, they are not war-winning: Russia and Ukraine have traded them across the front for years, and the line has barely moved.

Even the new class of affordable American missiles is not enough. A hundred-pound warhead can destroy a truck, but it cannot close a port, crater a runway, or bring down a hardened weapons plant. Our adversaries have thousands of these targets, often tucked hundreds of miles behind enemy lines, and each of them needs heavy payloads to be destroyed.

For the West to reach these targets, strategic weapons need the same revolution as tactical ones. We cannot deter with exquisite, expensive weapons built in the dozens. We need greater reach, heavier payloads, and tremendous mass.

Introducing Anthem.

Today, Covenant is announcing Anthem: a heavy-payload deep precision strike weapon system meant to fill the biggest gap in the arsenal of the West.

Anthem delivers multiples of the range and payload of anything in its price class.

Crucially, Anthem is more than just its specs; it was designed to be affordably stored and fielded at scale. The reality of many weapons programs is that the sticker price is often a fraction of the lifecycle cost: bunkers and launchers can cost more than the effectors themselves and are often available in quantities too small for mass fielding. We’ve put as much thought into the storage and handling of Anthem as we have into its performance.

Beyond delivering end-to-end hardware solutions, Covenant is also delivering the mission-planning and command-and-control software to make mass strike missions a reality.

The result is a deep precision strike capability which can be affordably stored, launched, and controlled in the thousands.

Since Anthem’s maiden flight in August 2025, Covenant has flown Anthem in over 200 tests, including multiple flight tests with the U.S. military both on U.S. soil and abroad.

Local Production, Local Supply Chains

In just under two years since our founding, we have raised $250 million, hired hundreds of the most talented engineers and operators, and opened factories to produce Anthem in 3 countries – the US, Germany, and Israel. Our priority is getting thousands of Anthems into the field, fast.

Anthem was designed to be produced quickly and at scale. Most of Anthem’s parts are commercial off the shelf, and where they are not, we did not wait for a supplier to scale up on their own: we either brought the part in house or partnered with the supplier to stand up an entirely new production line to meet our demand.

Covenant is producing Anthems locally for our customers so that they will be ready to launch when each country needs them.

Today, we cut the ribbon on our new 105,000 square foot production facility in Dallas, Texas, home to our American production line and supply chain. Later this year, we will unveil two more production facilities: 110,000 square feet in Germany, which will feature an almost entirely European supply chain, and 30,000 square feet in Israel.

Beginning in 2027, we will deliver thousands of Anthems per year.

We believe Covenant is building the best missile engineering team in the world.

Building an excellent affordable strategic weapon is hard. At our core are some of the most experienced missile engineers in the world, led by my co-founder Ofir Dayan. Ofir understands our customers’ problems as well as they do and has the tenacity to fight requirements that stand in the way of the best solutions. He leads our band of renegades – the engineers who ran as if with ankle weights inside the American, Israeli, and German primes, and still managed to deliver the most advanced offensive and defensive weapons in service. At Covenant they run free.

No less important are the Covenant operations and flight test teams that support them. Flight-testing large missiles is intense and dangerous work, but it’s the way to get a missile into service quickly. Covenant has staged flight tests nearly every week since our founding.

Covenant’s relationship with our customers is led by Abby Denburg, our President and Chief Growth Officer. Abby’s job is not just to sell, though she does that as well as anyone. She is the force converting our customers’ urgency into Covenant’s speed.

This is Covenant.

We’re proud of what we’ve built, but the hardest work — producing at scale, delivering reliably, and servicing our customers — is still ahead.

We open with Anthem.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

We are running out of watts.

AI inference is one of the fastest-growing markets in the history of capitalism, and we are running out of nearly every physical input required to serve it: powered land, turbines, transformers, data centers, GPUs, advanced-node wafers, and high-bandwidth memory. GPU capacity is being leased at record prices, while OpenAI’s and Anthropic’s growth is constrained by how quickly they can expand their compute fleets. You feel that scarcity firsthand: rate limits tighten and responses slow as providers trade latency for throughput to produce more tokens from the same watts.

This is why the largest capital investment cycle in contemporary history is underway. Seven of the ten most valuable companies in the world are hyperscalers or supply them. Five U.S. hyperscalers alone are expected to spend $1 trillion in capex next year. NVIDIA has gone a step further, working with capital partners to mobilize more than $500 billion for AI factories.

But expansion alone will not be enough. Inference demand compounds at software speed; power plants, data centers, and semiconductor fabs do not. Serving the most consequential workload of our lifetime requires squeezing more intelligence from every watt. That starts with recognizing that inference is not one workload.

The End of One-Size-Fits-All Compute

Without the GPU, today’s AI would not exist. Its flexibility enabled the early experiments with transformers. Then enormous clusters built around a single accelerator architecture powered the training runs that proved scaling laws and produced today’s frontier models.

But the diversity of AI applications is exploding. A voice assistant lives or dies on latency; batch processing maximizes throughput; a research agent may prioritize cost per token; a coding agent must balance all three. Even a single model call splits into compute-intensive prefill and memory-bound decode. Agents multiply the variation by routing among small specialized and larger reasoning models, retrieving files, executing code on CPUs, and calling external tools. Inference is no longer a single workload.

At the same time, hardware choices are proliferating. Some architectures optimize arithmetic throughput; others optimize memory capacity, bandwidth, or low-latency connectivity. No single processor can deliver the best combination of cost, throughput, and latency across every application and every step. GPUs will remain foundational, but they will increasingly operate alongside CPUs, memory-optimized systems, and other purpose-built silicon.

The optimal system must match each piece of work to the architecture that handles it best, producing more useful work from the same resources. It must be heterogeneous.

The challenge is making all of this hardware operate as one system. Large-scale infrastructure is designed around homogeneity with standardized servers, networking, power, cooling, and operating models. Introducing multiple architectures creates complexity across every layer. The software must coordinate processors that were never designed to work together, while the physical facility must accommodate different networking topologies, rack densities, power profiles, and cooling requirements. Cerebras and NVIDIA systems, for example, require different inlet-water temperatures. Sometimes heterogeneous compute is a compiler problem; sometimes it is a plumbing problem.

The Multi-Silicon Inference Cloud

Making heterogeneous compute work at scale is key to unlocking what we believe will be the world’s largest market. We believe Gimlet Labs has built the solution: the first multi-silicon inference cloud, designed to produce more intelligence from every watt.

For each workload, Gimlet determines an execution plan that balances latency, throughput, and cost. It can route different models and tools onto different processors, separate prefill from decode, or divide a model at the layer or operation level. Its compiler optimizes each piece for the target hardware, while its runtime coordinates execution across the system. To developers, all of that complexity sits behind a single inference API.

Gimlet extends that orchestration into the physical data center. It integrates GPUs, CPUs, and purpose-built accelerators into one pool of capacity, managing the differences in networking, power, and cooling that make heterogeneous infrastructure difficult to operate. This system is already delivering up to 10X gains in throughput and interactivity on frontier models within the same power envelope.

In a world where scale matters, Gimlet has done this at the largest scale possible, counting both a frontier lab and a hyperscaler as customers. In a market starved for compute, efficiency is net-new capacity and latency is product differentiation.

We believe there is no better team than Zain, Michelle, Natalie, Omid, and James to tackle this problem. They’ve seen this movie before and are one of the few teams that have the range to follow a problem through every layer of abstraction. When kernels and compilers were not enough, they pushed into high-speed networking. When software was not enough, they moved into power, cooling, and data center construction. Wherever the existing stack ends, the Gimlet team starts.

The future of inference is heterogeneous, and Gimlet is building the infrastructure layer that makes it possible. We could not be more excited to partner with the entire Gimlet 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.

America | Tech | Opinion | Culture | Charts

Experience > Skills

Charts has already observed that one of the challenges for prospective entry-level hires (other than the fact that entry-level hires have always been a relatively small share of tech hiring) is that existing tech workers appear to be getting older. Why that’s the case is some combination of aging-in-place, plus remote work unlocking a broader pool of talent, plus perhaps some premium for more experienced workers—and while AI may have something to do with it, as well, the evidence for that is pretty weak.

On the experience point, there’s some recent data from Revelio Labs that suggests experience is, indeed, becoming a higher priority for tech hiring:

Since January 2025, tech job postings have seen a marked rise in preferred years of experience, with a concomitant drop in preferred number of skills.

The change is small—~5-10% in both cases–but noticeable, and it’s consistent with the notion that tech is looking for more experienced workers, rather than workers with any particular set of skills, let alone a broad range of skills. That could very well be an AI-driven shift: grappling with a new technology means that extant “skills” are less applicable, and the new set of skills is still a bit of a whitespace. When you’re operating on a new frontier, you don’t yet know what you’ll need, other than the need to be able to figure it out—and that’s where experience goes a long way.

If you look at the types of jobs that require the largest premia to experience, it does seem like AI is the likely driver:

It’s data and systems work (and Java) that dominate the leaderboard for the recent hiring premium to experience.

AI has created new urgency around data wrangling, but the shape of wrangled data (and the process of wrangling) has changed dramatically, with optimal solutions still very much a moving target. If you don’t know what needs to be done, then the skillset hasn’t been defined yet, and someone who can figure it out as they go is the better choice. Eventually, though, as implementation matures, one would expect new skills to be defined, and it’s likely the skills/experience mix will shift again.

Netflix Having a Blockbuster Moment?

A funny thing about social media consumption is that aggregate time spent in the US is still rising (albeit barely), but not for everyone:

According to Apptopia, aggregate time spent on social media apps has risen (after a small drop), but it’s really the 17-25 yos doing all of the work—the folks 36+ appear to have tapered off their social media time since October ‘25.

But that’s not actually the funny thing.

The funny thing is that while social media usage is rising, social media’s share of screentime has actually dropped pretty substantially of late:

Since April of this year, social media’s share of mobile has dropped ~3pp. OK, perhaps “substantially” is a bit of a stretch, but after years of steady growth (including double-digit growth as of last year), the dropoff is meaningful.

So people are spending more time on social media, but even more screen time on other things. What other things? Apps perhaps. AI could also be drawing more attention. But a more likely culprit for social media’s diminished share of mobile is none other than short-form video, particularly via YouTube.

OK, great, so more social media, more YouTube, and more screentime overall–a win-win-win for staying glued to our phones. That’s true so far as it goes, but there does, in fact, appear to be at least one victim of YouTube’s assault on our attention-spans, and that’s Netflix:

Again, according to Apptopia, the number of YouTube users who also use Netflix has declined ~20% since the beginning of the year, while the number of Netflix users who also use YouTube has remained pretty steady.

YouTube viewers seem more inclined to live without Netflix than the other way around.

In general, Netflix’s active users and viewing time in the US have both been in mostly serial decline for the past 3+ years:

Weekly users and time spent for Netflix have dropped ~15% and 30% respectively.

Why that’s the case is anyone’s guess. Among other things, it could be the rise of short-form video, more streaming competition, or streamflation more generally. As a possibly relevant aside, Netflix stands out for having the highest customer LTV by far, at least according to this estimate from Owl & Co.

A Netflix premium customer has an LTV of nearly $1,500 per customer, which is nearly 2x the next best offering.

LTV reflects a combination of higher subscription prices and strong retention—good for Netflix—but one could see why customers might begin to balk at the price. One could also see why Netflix is branching out into live sporting events and clips and shorts of its own.

Bigger picture, though, it’s fair to wonder if Netflix is having its Blockbuster moment, roughly two decades after Netflix gave Blockbuster its Blockbuster moment. It’s too soon to tell, of course, but in a different timeline, where AI was not such a big deal, this would be a much bigger deal.

Cyber Game Done Changed

Have you heard of the ‘Hugging Face Incident’? I mean, you probably have, but if you haven’t, you probably shouldn’t, and then you’ll sleep easier at night. (Old heads might be reminded of the Noodle Incident, but sadly, these are not the same.)

Anyways, HFI aside, there’s plenty of recent data to suggest that the threat landscape for cyber events has changed dramatically. Consider this nightmare fuel from Epoch AI:

The number of reported critical and high severity vulnerabilities from major firms has gone vertical this year, since the beginning of the spring.

Whether there are actually more vulnerabilities, or whether there’s some reporting or definition change, it’s hard to say just by looking at the chart, but it’s likely that at least some (and perhaps a lot) of this parabolic growth is attributable to a major substantive change in either how vulnerabilities are discovered and/or what is still considered secure.

Here’s another bit of nightmare fuel for the cyber-minded:

According to the ZeroDayClock, the Zero-Day Rate (i.e. the percent of vulnerabilities exploited on the day of disclosure (or before)) has basically quadrupled since 2020—including a ~60% increase from last year alone. The Zero-Day Rate is now just under 87%, with a median time to exploit at just one day . . . but projected to reach 1 minute sometime next year.

To repeat, the ZeroDay estimate is that it will take only a minute between discovery and exploitation of a vulnerability (at the median) by sometime next year.

Even now, the “Exploit Survival Curve” falls to 0% at just 1.5 months:

Back in 2022, a full 50% of exploits were still there for the taking after 1.5 months (and at 3 months, too), but now, if you haven’t plugged all the holes, by then, someone else will have already found them.

Clearly AI has changed the game, not only for identifying vulnerabilities, but also exploiting them. The good news is that AI will also change the game for the defense.

Either way, the urgency for cyber has intensified, and as Charts has observed before, investors are way out in front of this one:

Of the $IGV constituent stocks, Cyber has massively outperformed the field, and forward revenue multiples are consistently higher, as well.

Whether and to what extent the incumbents successfully step into the breach, or whether some new cyber platforms emerge to face the challenge, remains to be seen, but there are surely going to be a whole bunch of new sheriffs in town (or else).

Data Center Job Boom Continues Apace

One final aside on the economic driver that is the AI buildout.

BLS released its latest estimate of data center construction spending, and while there was no reason to think otherwise, the pace is blistering:

Data center construction spending jumped by ~$25B in just half a year, which is roughly the same amount it jumped over the past two years.

But, more to the point, alongside the jump in construction spending is a jump in construction job openings, which have otherwise been in steady decline since interest rates went up, and most other construction slowed substantially.

Goldman estimates that ~300,000 construction and specialty trade workers have been added to the workforce since 2022 that are attributable to the AI buildout, with nearly 75,000 of those coming in the past year alone:

Again, while the overall trend for hard hat workers has been directionally negative since rates went up, that trend has reversed to the upside since the middle of last year.

Like ‘em or not, data centers are major jobmakers, bringing a pretty wage premium, as well—and it’s fair to say that without data centers, the employment picture, especially for blue-collar types, would be substantially worse.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

The bullish case for incumbent systems of record is that AI makes systems of record get more important, not less. Why? Because in theory the customer can now pipe the system’s data into Claude or Codex and skip the AI-native app.

As agents do more work through these systems, the data and actions they control become more valuable. An incumbent system of record plus a general agent (or an agent built by the incumbent) therefore might be enough to get the work done. This is what Salesforce just announced in partnership with Anthropic: Claudeforce lets you work with the Salesforce CRM from inside Claude without having to open Salesforce (or so they claim!).

The argument is partly correct. The incumbents aren’t sitting idly.

The incumbents have an incentive to control and charge for access to their data, and now to push their own agents forward too. We’re already seeing incumbents market products that move from simply storing information toward doing more of the work itself. Docusign’s Iris reviews contracts, Atlassian’s Rovo aims to resolve and route requests, and Klaviyo’s Composer builds campaigns and assesses marketing performance. These products move incumbents beyond the “slap on a chatbot” strategy and into taking action.

At the same time, Claude is starting to coordinate work across applications, making it possible for general-purpose agents to act as a layer above the incumbents. Claudeforce shows what that can look like: Claude can become the front door to the job while Salesforce still controls the underlying CRM data and actions. The interface and the system of record look like they’re unbundling, creating an opening for AI-native startups to rebundle around the job.

So where can startups still compete?

The vertical AI-native company has to win through focus. It has to perform a specific cross-system job better than an incumbent’s purpose-built agent or a general-purpose agent like Claude, which will be able to reconstruct it from the underlying tools. Focus can earn the vertical startup deeper access, a deliberate data asset, more specific context, and a learning loop that builds a better understanding over time of what good work looks like for a given job.

The agent hierarchy

To see where incumbents are strongest and where vertical AI companies may have an advantage, it’s helpful to think about the four types of application-level agents, distinguished by how much autonomy and judgment each requires.

Retrieval assistants find information, summarize documents, answer questions, and draft responses. Process agents carry out rule-based work like updating records or routing approvals, using limited judgment to interpret the task, and act. Policy agents apply the rules in an organization’s playbooks, precedents, and thresholds to ambiguous cases. Principal agents make judgment calls about what the organization should do, often with open-ended tradeoffs around strategy, risk, and resources.

A year ago, most incumbents’ AI products were still at the retrieval stage, if that. They functioned mostly as chatbots and analytical tools pulling from existing records. Now, more can take action based on the workflows, permissions, and rules already embedded in their products. A few are beginning to apply narrow, human-defined policy. Judgment gets harder when the decision depends on information outside the record that the incumbent manages.

The broad shift we’re seeing is incumbents moving up the hierarchy, from retrieval and process toward more judgment through policy, like Docusign’s Iris applying a legal team’s playbook. But that judgment is still largely bounded by the record the incumbent owns.

The job is bigger than the record

An incumbent can automate more of the work around its core record, but the customer’s job to be done is bigger than the record. This isn’t a hard technical boundary because incumbents can pull in outside data and add workflows, but their work still expands outward from the record they already own.

A customer’s work gets done through a combination of its people, processes and software. The contract is not the legal matter, the ticket is not the customer’s resolution, and the opportunity is not the sale. The full job crosses applications, teams, and even companies. Different parties may hold different information, and want different outcomes. The incumbent is limited to building a “local system of work” but the larger opportunity is to manage the job across boundaries, including all of the parties, the people, the documents, decisions, revisions, and the final result.

A general purpose agent like Claude or Codex may be able to reach across all these systems, but access alone doesn’t mean it can perform the job well. Even using MCP, pulling info across applications has latency, and each application may describe the same customer, contract, or transaction differently. Even if a general purpose agent can access every app in one company, it doesn’t automatically have access to information held by external parties (e.g. between buyers and suppliers).

Even with these limitations, Claude can still sit above systems of record. Claudeforce makes Claude the front door to the job while Salesforce still owns the CRM data and actions. And Claude can keep gaining context through deeper integrations blessed by the incumbents.

How vertical AI learns

Owning more of the job matters because it lets startups see the decisions and corrections that produced the final result. A model performs better on a specific job via a vertical harness with the right context, tools, workflows and evals. The learning loop is what improves the system - the model + harness - after each job, using the agent’s work, expert feedback, and outcomes themselves to improve the next attempt.

Completed records alone don’t automatically suffice as a training curriculum for the work. A signed contract shows what the parties agreed to, but not every alternative considered or every reason for an exception. A closed ticket shows the resolution, but not every hypothesis the support team tested.

Remembering vs. learning

The labs are also building general-purpose memory, so preserving context alone is not a durable advantage. Remembering and learning are different. Memory can help a general purpose agent recall a customer’s preferences or a decision made last week. But it doesn’t tell the agent whether the work was good or why an expert changed it or what to change next time. Focus lets the company see more examples of the same job, learn faster, and get better at measuring improvement.

Profession and Institution

There are two kinds of learning. Learning how a good professional does the work. A company can begin teaching that with assignments created by experts, realistic examples, and standards for a good result.

Learning how a particular firm or customer does the work, including its templates and precedents, risk thresholds and escalation rules. That can begin with the customer’s existing materials and improve through corrections and exceptions in the training process. Some lessons about the profession can improve the product for every customer, but lessons about one firm may just improve the product only for that firm.

To learn both the profession and the institution, the startup doesn’t need to start with the largest stock of historical data. It can manufacture a curriculum for the profession, then use production to learn the institution. As open-weight models and post-training increase, assembling the right data and evals becomes even more important.

Long running agents especially benefit from learning loops. When an agent does work over hours or days, through a complex logic chain with many intermediate decisions and handoffs, the quality of the learning loop matters even more. The longer horizon task can be unbundled and evaluated through a set of checkpoints.

To be clear, incumbents will also learn from the work performed inside their products. But that learning will cover only their part of the job: the contract review inside Docusign, the ticket resolution inside Jira, or the pipeline update inside Salesforce.

How Harvey manufactured the curriculum

Harvey’s work on Tenet shows what the path for vertical AI can look like. Harvey post-trained an open-weight model using synthetic data, public legal data, and human-expert created data, without using customer data to start. It created roughly 1,750 legal-task environments, each simulating a partner-assigned matter, complete with documents and tools and an expert rubric for a high-quality result. The average assignment had about 50 specific criteria. Instead of waiting years to accumulate enough customer history, Harvey manufactured a curriculum it could use to train and evaluate its model.

Harvey’s broader work also shows how a vertical company can train a model for different parts of the job, including industry-specific capabilities like M&A diligence and understanding a firm’s total knowledge.

The startup’s opportunity is to put it all together. Vertical specificity helps the startup to break the work into specific capabilities, tackle the practical frictions around data rights, systems, and permissioning, and fits the product into the customer’s daily work. Eventually, it can feel comfortable taking responsibility for the result. Deployment and responsibility are key here!

What makes a good vertical AI market?

Using Harvey as one example, here are a few axes for evaluating potential vertical AI markets.

• Can an expert quickly tell what the AI got right or wrong, and explain how to improve it?

• Is the work hard enough that judgment matters (vs. simple rules)?

• Does the work happen often enough for the product to learn?

• Can the startup begin with one assignment and grow into doing an entire job?

The best vertical AI markets tend to meet all four, giving them a strong learning loop.

Many workflows in legal, tax, and accounting meet this test because the work repeats, requires judgment, and experts already review the results. This pattern also shows up in less obvious markets like industrial work. When a manufacturing defect appears, a quality engineer has to gather test results, supplier documents, equipment logs, and the plant rules, then decide what’s acceptable and what needs to change. That information may sit across several systems. One incumbent may store the defect report, but no single product owns the full investigation. A startup could begin with one piece of this work (e.g. drafting the investigation report) and over time own the end-to-end resolution. To do this, it would need to learn both how a good quality engineer investigates a problem and how a particular plant or customer or auditor does it well.

The job is still up for grabs

There are jobs where a system of record plus a general-purpose agent will be enough, but there’s still meaningful opportunity for vertical AI-native startups! The strongest vertical opportunities exist where the work happens often, expert judgment matters, and the learning loop is strong. The incumbent may own the record and the lab may control the front door, but the vertical AI company can still win by becoming the best at doing the job itself.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

Everyday life on Earth depends on space, whether you know it or not.

In the last 24 hours, satellites helped you check the weather, navigate to work, communicate with others, fly across the country, or support those defending our country. Our lives do not just touch space — they rely on it. If today’s satellites stopped working, the world as we know it would grind to a halt.

At Apex, we believe space is already foundational infrastructure for Earth — and this is just the beginning. Space already supports the core systems that modern life depends on. The next era goes further: innovation, economic growth, and defense will move faster because of space, unlocking capabilities beyond what Earth-bound systems alone can achieve.

For most of space history, progress was measured in firsts: first satellite, first human in space, first Moon landing. Those milestones mattered. But the next era will be measured in rate: how fast countries and companies can deploy new capabilities, replenish what is lost, refresh technology, and expand the infrastructure they depend on.

That is the new Space Race: not a race for singular achievements, but a race for resilient infrastructure, industrial capacity, and continuous technology refresh.

To make that future real, the world needs constellations: many satellites working together as systems. That requires two industrial capabilities: launch and spacecraft platforms.

Launch is ready enough. Spacecraft platforms are not.

Apex exists to remove the spacecraft platform bottleneck.

Our mission is to expand humanity’s reach by delivering advanced spacecraft platforms that scale. In practice, that means making constellations rapidly deployable for the markets that need them most.

For most of the space age, launch was rare, expensive, and unforgiving. That scarcity forced missions to be slow, bespoke, expensive, and optimized around one big shot at success. The old model was not broken. It was built for a different constraint.

Then launch became more reliable, more frequent, and more available. Not necessarily cheaper from a customer-price perspective — demand for launch has remained enormous, and prices have risen. But launch capacity, cadence, and reliability changed dramatically.

A ride to orbit every year forces scarcity. A ride to orbit every day creates new models — but only if you have satellites to deploy. As launch became more available, the constraint shifted from getting to orbit to building enough capable spacecraft platforms to use orbit.

That is where Apex plays.

Constellations Are the New Architecture

Constellations are not just more satellites in orbit. They are a fundamentally different way to deliver capability from space.

With more satellites, you cover more of Earth more often. With systems closer to Earth and distributed across orbits, communications move faster. If one satellite is disabled or degraded, the rest of the constellation can help fill the gap. And when satellites are continuously replenished, new technology gets to orbit faster instead of waiting years for the next generation.

This is how space becomes rapidly deployable operational infrastructure: not one-off missions, not isolated assets, but systems that can be deployed, expanded, refreshed, and relied upon.

In defense, that means architectures that are harder to blind, difficult to disable, and faster to replenish. In commercial markets, it means new services can be deployed and improved continuously. For allied nations and emerging space powers, it means access to capabilities that once required decades of bespoke development.

The advantage is no longer only who has the best single satellite.

The advantage goes to whoever can deploy, sustain, and improve their constellations the fastest.

The Platform Gap

Starlink and Amazon Leo are prominent examples of the shift toward constellations: many satellites working together as a system rather than relying on a single asset. More broadly, constellation architectures can offer important advantages over single-satellite architectures, including greater coverage, lower latency, increased resilience, and faster technology refresh.

But these mega-constellations were built for narrow use cases.

Their spacecraft, payloads, network architecture, ground systems, software, and customer experience were designed together around one mission. They are not configurable platforms for the rest of the market.

Most critical space missions need a configurable platform approach. Governments, defense primes, allied nations, and commercial operators often bring their own payloads, ground systems, C2, software, customer relationships, or mission architectures. They need the benefits of constellation-scale deployment without building an entire vertically integrated space company around one use case.

That is the platform gap.

Apex is doing for the rest of the market what vertically integrated constellations did for their own missions: turning repeatability, integration, and scale into deployable capability.

That is the category Apex was built to define.

Build Once. Configure. Deploy Many.

A true platform creates a standard that many missions can build on.

For Apex, that means configurable spacecraft platforms, standard interfaces, clear specifications, integration processes, mission services, and the manufacturing system required to move customers from payload to deployed capability faster.

Apex does not build one-off platforms for every mission. We build an end-to-end, configurable product suite for markets that need constellations.

Apex says yes to missions where scale matter: missions that require many of the same spacecraft platforms, where speed, affordability, quality, and production capacity determine whether the architecture works at all.

Defense, intelligence, communications, sensing, science, commercial infrastructure — the specific mission can vary. The core question is the same: does this market need many satellites, deployed quickly and reliably? If yes, that is where Apex belongs.

Productization eliminates platform-level NRE, improves reliability, accelerates schedules, and creates the Apex flywheel: every platform we build makes the next one faster, more affordable, and higher quality. Volume compounds into learning; learning compounds into better platforms; better platforms win more missions; and more missions spin the loop faster. Every one-off change that breaks the platform resets the learning curve, reinforcing how repeatability becomes a structural advantage, not just an operating philosophy.

The flywheel compounds across product generations. Aries gave Apex flight heritage, production learning, and the foundation for our software and AI orchestrated company operating system, Octopus. Nova builds on that foundation with more capability, stronger price-performance, and a system designed for larger, more demanding customers. Each new vehicle and block upgrade inherits common interfaces, subsystems, software, production processes, and lessons from the last — then feeds improvements back into the platform through block upgrades.

The factory is not separate from the platform: it is the core of the platform. A productized spacecraft, Octopus, and a repeatable production model make it possible to scale beyond a single facility. Over time, Apex extends this model through a global network of Apex-enabled factories — all building from the same standards, software, interfaces, and quality system.

Customers should differentiate on payload, mission, CONOPS, software, and capability — not on custom platform development that slows them down before they ever get to orbit.

We standardize everything that should not be unique, so customers can focus on what must be.

That is how a product becomes a platform. And that is how a platform becomes a standard.

The Apex Industrial System

Moving at constellation speed requires more than a spacecraft platform. It requires an industrial system.

Apex vertically integrates because the logic of the mission demands it: speed, reliability, affordability, quality, and time to market require control over the systems that determine whether customers can deploy on schedule.

Octopus, our proprietary company operating system, connects business development, product, finance, engineering, supply chain, manufacturing, test, and delivery. It gives Apex the control, traceability, and operating rhythm required to build spacecraft platforms repeatedly, reliably, and at speed.

This is the Apex Industrial System: vertically integrated hardware, software-orchestrated operations, and a factory built from day one to scale. Vertical integration controls the parts. Closed-loop industrialization controls the system — and lets the flywheel spin faster.

The Apex Way: Own the Risk

Speed is meaningless if customers pay the price for it. That’s how the satellite industry works. That’s not how Apex works.

Every new spacecraft platform begins as an Apex Testbed vehicle. We move aggressively — designing, building, flying, learning, and iterating as quickly as possible. We “test like you fly.” We also believe in something equally important: learn in flight. Ground testing can only take you so far. Space is the final test environment.

We intentionally take that risk ourselves and launch Apex-owned spacecraft early in every product family to expose issues, learn quickly, and improve the platform. Those are our risks to take — not our customers’. When we encounter anomalies, we solve them, fly again, and keep learning until the platform is ready.

Customers don’t wait to buy the platform. They know that by the time their mission launches, Apex will already have proven it in orbit.

This philosophy is why Apex moves quickly without asking customers to accept unnecessary risk. We move at an unprecedented pace. We absorb the learning. We fly again when necessary. Only after the platform has proven itself do our customers become the first to fly it on their missions.

Design, produce, test, learn, iterate… Apex absorbs the risk, and then passes the proven spacecraft platforms to our customers.

We absorb the risk so our customers inherit proven spacecraft platforms. That’s the Apex Way.

From Platform to Mission

The platform lays the foundation. Mission Integration Engineering takes the foundational platform and turns it into the backbone of a customer mission.

Productization does not mean every mission is the same. Customers bring unique payloads, requirements, interfaces, CONOPS, and constraints. Apex’s Mission Integration Engineers keep the platform sacred, while helping customers connect payloads, resolve interfaces, reduce NRE, and accelerate qualification. By keeping the platform constant and engineering the payload-side compatibility, we help customers get from payload to deployed constellation, fast.

For some customers, that means a spacecraft platform and integration support. For others, it means software, communications, ground, operations, and more of the end-to-end mission stack. Apex meets customers where they are while keeping the platform productized.

As the platform becomes the core of a constellation, there will be missions where Apex should selectively own the entire stack — including prime responsibility — because doing so makes the mission faster, more integrated, and more deployable. We will be selective. The goal is not to chase every mission; it is to own the missions where the scale of the platform is the key to making the constellation real.

Standardize the platform. Integrate the mission. Deploy at constellation speed.

Proof in Orbit

This is not theoretical.

Apex has already proven that a productized spacecraft platform company can move faster than the traditional model without sacrificing quality. The first Aries went from clean-sheet design to orbit in under 12 months and has operated on orbit for over two years. Since then, Apex has achieved 100% customer mission success and delivered over a dozen platforms to customers, while continuing to ramp Factory One production to over 200 platforms per year.

Speed is not the antithesis of quality when speed is achieved through repeatability.

We are not just designing satellites. We are building the platforms, factory, software infrastructure, supplier base, mission integration capability, and operating model required to make constellations rapidly deployable.

Apex’s Strategy

The strategy is simple, but difficult: own the world’s spacecraft platform category.

We do that by empowering exceptional talent to build an end-to-end product suite for markets that need constellations; scale ahead of demand through vertical integration; and win on speed, reliability, and affordability.

We stay focused by refusing to chase every custom opportunity.

By enabling our customers, humanity’s reach expands. Space infrastructure becomes deployable, not theoretical. Constellations become accessible to customers who could not — or should not — build the entire stack themselves.

Space is becoming foundational infrastructure for humanity.

The next era depends on making constellations rapidly deployable.

The world does not need another satellite manufacturer.

It needs the spacecraft platform company that makes constellations possible.

That is Apex.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

Today’s guest post is by Ruby Thelot, a professor at NYU and astute tech observer. You can watch his recent appearance on MTS here.

Every year, more people use and spend more time on Big Tech platforms. It is fashionable to hate on social media and to pontificate about how it’s worsened, because the intellectual accoutrement du jour is being a conscientious (™) tech critic. This stated preference does not match the general public’s behaviours.

“Enshittification” describes the progressive deterioration of platforms in service of business goals over user needs. In the writer Cory Doctorow’s original definition from his 2022 blog post, he uses the term to describe how platforms die: “First, they are good to their users; then they abuse their users to make things better for their business customers; finally, they abuse those business customers to claw back all the value for themselves. Then, they die.”

I do not believe, no matter how compelling it is as a theory, that it holds up to scrutiny. Simply put, more people use Instagram every year, they use it on average longer every day, year over year and, finally, their reported satisfaction has also grown over time, indicating no decrease in overall satisfaction, stated or revealed.

For our purposes, let’s focus on Instagram, whose logo was just redesigned to many online commentators’ chagrin. Certain people just hate when platforms change.

People en masse are getting on social media, by choice. From 2016 until now, the number of social media users has grown from 2.5B to close to 6B people. Enshittification isn’t real. It’s three TikToks in a trenchcoat, and, maybe, a book deal. It’s good for discourse but does not describe actual reality and user patterns.

Returning to Instagram. The average daily usage showcases an increase in time spent on Instagram every day as well. Users are coming back for more, in a highly competitive attention arena, year after year.

To be fair, certain people do feel like platforms have gotten worse, this perceived feeling describes the progressive dislocation between the platform’s main features and the user’s needs. Let’s take Facebook for example. Doctorow was probably in his thirties when Facebook started, he probably lived in an urban area, he had a specific set of needs, and that set of needs (or user goals) was met by Facebook, namely adding friends, at some point sharing images, and communicating through the wall.

Instagram has indeed changed. A look at its interface over the years showcases severe and notable differences as it evolved.

In its first five years, icons were altered, their positions were switched.

Its main functionality also changed. Notably, the feed went from chronological to algorithmic, to many’s chagrins, but algorithmic feeds show users the things they are predicted to engage with. The UX has been altered because the function of social media has changed. People now use Instagram as a means to get “entertainment”, it is no longer purely “social” media, as it used to be in yesteryears. Instagram has now taken on the role of social coordination as well. Events, book clubs, concerts all use the platform to broadcast logistical updates and keep community members informed. The “channel” feature is a good demonstration of this need. Nonetheless, the amount of time people spend on social media is increasing every year. The revealed preference is clear.

Facebook is a good example of that shift. If you pull the timeline of features that were dropped on Facebook over time—the Wall (2004-2011), the introduction of News Feed in September 2006, the Like button (2009), the shift to an engagement-based algorithm in 2009, Timeline itself in 2011, Story Bumping in 2013, and the prioritization of friends and family over brand pages in 2015—you can see that at some point, Facebook changes from being a tool that’s used specifically by college students and young urbanites to something much bigger in order to accommodate a larger and more global public. The platform evolves as its user base diversifies.

In other words, platforms start with a pretty constrained set of personas, i.e. a narrow set of user goals, but as the user base grows, as the number of monthly active users increases, there is a multiplication of personas, and therefore, a multiplication of goals. The job of the product managers becomes more difficult as the user goals become more varied, more heteroclite.

There is a third and more pragmatic argument against the enshittification hypothesis. At a company like Meta, platforms are shaped by the emergent outcome of thousands of small, motivated, data-driven product decisions. A large platform is, in practice, an enormous behavioural machine. Growth PMs, designers, researchers, and executives are constantly watching what people actually do. An interesting recurring instance of platform narcissism is that with each rollout, tech critics and certain people on social media do complain. Undeterred, the company rarely rolls things back and lets the numbers play out. Oftentimes, they expose the gap between stated preference (what people say they like and dislike) and revealed preference (how they behave numerically on the platform).

People complained about Stories, Reels, algorithmic feeds, recommendations, the intrusion of shopping and ads, suggested posts, and then, through the blunt and vulgar clarity of revealed preference, they still use them. They do not have to. The platform changes because the quantified flows of user behaviour pull it in that direction. Of course, some of these decisions are balanced with business goals, this tightrope act is quite literally the job of the PM. The point being that “I hate this” is not the same thing as “users hate this,” and “everyone I know hates this” is often just another way of saying “my cohort is no longer the platform’s central desired persona.”

This last segment is the crux of platform narcissism and the dislocation felt by early adopters in the West. This is especially true on large platforms, where you have global growth, and, thus, geographic diversification. Simply put, users in the Philippines don’t want the same things as the users in an urban centre in America want. Even at a smaller scale, users in Missouri similarly don’t want the same thing as a user who lives in Bushwick.

What people call enshittification is the feeling of feature-need dislocation. In this case, it is felt by mostly geriatric millennials. The platforms are more popular than ever. You can see it in the Monthly Active Users (MAU) growth mentioned above, 140% in ten years. The revealed preference is that people are using them more and more. Moreover, satisfaction surveys such as ACSI indicate that satisfaction with social media has remained consistent in the last 8 years, with a slight increase in the last three years, from 2020 to 2025, satisfaction grew from 70 to 75 (out of 100).

Contrary to the enshittification narrative, when broken down by platform, all platforms show increases in satisfaction. Usually, with contentious issues, one expects a divergence between stated and revealed preference, when something is perceived as more virtuous, they may claim that they do it, but in reality, their revealed behaviour showcases that they do not follow their claim. This is notably not the case here, both usage and claimed satisfaction have increased.

A misaligned platform could in theory trap its existing user base, but it would fail to attract new ones. Doctorow claims that platforms favor advertisers (or whomever they get their money from) and often price gouge their users, but, with social media, the big social media platforms are still free. The broad behavioural and satisfaction data do not support the sweeping claim that platforms are simply getting worse for users as a whole. Moreover, we would see decreases in overall use (time spent) and DAU numbers. The claim that the overall quality has decreased is subjective and specific to a narrow demographic of tech critical users such as Doctorow.

Platform Narcissism

Platform narcissism: the belief that the platforms one uses should continue to respond to one’s idiosyncratic user needs.

Perhaps the reason why the platform seems worse is because your life has changed. Over the same period of time, your goals have changed, and they no longer fit with the platform’s new features. Why the hell would a fifty-five year-old be watching Reels? Go take care of your kids.

The platform needs to respond to the needs of new users, who, from a demographic standpoint, often outnumber the original early adopter cohort and are now usually more engaged. Because they have different desires, different wants, executives make a conscious choice to expand in the direction which they think best suits their business goals and aligns with newer cohorts. This is the dislocation which causes the feeling of enshittification: the evolution of the platform away from a group or an initial user base.

Many will claim that platforms have trapped users, that the cost of departure or logging off is too onerous. If that were the case, this would be reflected in the satisfaction scores. Services with treacherous lock-ins tend not to score well in consumer satisfaction. My theory is that the dislocation is most felt by the chronically online, which includes the media class. Due to the nature of their work, they feel trapped. However, for the average Instagram user, who has ~675 followers, this isn’t a real problem. Additionally, as one can see from the per-platform chart, many options exist. People have the power, the autonomy, and the agency to pick the one that best serves them. Anecdotally, when it stopped serving them, Western millennials and Gen Z left Facebook.

The more interesting question, then, is not why platforms change, but why the language of platform decay has become so emotionally satisfying and so virally successful. A person can believe, all at once, that Instagram is worse, that using Instagram is embarrassing, that the platform is culturally corrosive, and that they personally are still using it more than ever. At the heart of the belief in platform decay is a form of cognitive dissonance. Many detractors of platforms are still very active users of the platforms they supposedly hate.

Hypotheses like enshittification provide an absolution for the platform narcissists. It is safe to wager that the platform narcissist is still online, they’re still using the platforms, but they need terms like enshittification to encapsulate their feelings and explain why they are still on the platform. The enshittification narrative resolves the contradiction by relocating agency towards the platform. Often, highlighting that users are trapped on these awful platforms against their will. Crying-worse allows the user to use the platform performatively half-heartedly and outsource the guilt. Critics such as Doctorow offer then a psychic service, a sort of societal exutory for all the amassed angst felt by platform narcissists. They are high priests of the terminally online platform narcissists and their indulgences are views on short form content railing about platforms and book sales.

In sum, the platforms are not worse. There is no such thing as enshittification. It is merely platform narcissism. What is platform narcissism? It is the belief that the platforms you used for the last few years should continue to respond to your specific needs. And frankly, they don’t need to.

Your favourite restaurant can stop serving your favourite dish. Your favourite store can stop selling your favourite brand. It’s just the way it is. These things change. And, for some reason, we think we’re beholden to some kind of static nature of these platforms because we were on them when we were sixteen or twenty-five. No. They’re going to evolve. And if you don’t like it, uninstall them, and move on. Vote with your thumbs, as they say.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

There has never been an investing environment like what we’re living through right now. Normally, investors hope and pray to catch the beginning of one generational S-curve of technology adoption. Today we’re watching at least six mega-trends emerge:

1. Enterprises are putting AI tools to work across every part of their organizations. We believe there is no upper bound on the demand to turn compute into business outcomes.

2. Consumer AI is barely even a thing yet, aside from ChatGPT’s form factor beginning to replace the search engine. But we expect it will be transformational when the time is right.

3. American Dynamism has entered a new era. National security, industrial policy, and technological progress are converging to rebuild the physical systems that underpin the country, from defense and manufacturing to energy, infrastructure, and space.

4. Robotics and autonomy are poised to be the biggest distributed infrastructure deployment in our lifetime, transforming how people and products move and interact in every setting, in our factories, on our roads, and everywhere in between.

5. Healthcare is 18% of GDP and represents one of the biggest areas of opportunity, as our biology becomes programmable and we use AI to improve systems.

6. To power the above, the entire compute stack is being rebuilt for the AI-era.

In moments of opportunity like these, we keep coming back to the founder as the key to building generational businesses. Technology has become so powerful for company-builders, and yet so unevenly diffused in the economy, killer founders who understand it will always be the premier destination for investment capital. Our job as investors is to identify these founders and help them however they need.

Today, a16z Growth has closed additional capital, bringing our fifth Growth fund to a total of $8.5B. With so much opportunity in the market right now, this became the obvious thing to do for both founders and our Limited Partners.

Helping founders accelerate through the curve

Every founder who’s been through the scaling journey knows that there are inflection points where their early instincts aren’t enough anymore. Companies must become multi-product, multi-channel, multi-geography, often all at the same time. Databricks had to bank through the corner and evolve “Lakehouse” into a dominant platform. SpaceX evolved from a launch provider into a global communications, data infrastructure and AI company. Greatness isn’t easy.

The founder’s opportunity, at the growth stage of the business, is to keep growing really aggressively through each corner. In our 7+ years running a16z Growth, we’ve helped over 100 companies through these transitions.

The a16z Growth Platform: built by operators, for operators

The a16z Growth Platform includes access to a16z’s comprehensive offering of expertise, network, and infrastructure tailored to growth-stage companies. Our GTM team connects portfolio companies with new customers; Talent has built one of the world’s best executive and technical talent networks; Global efforts unlock international relationships for capital, partnerships, and customers; and New Media helps our founders and companies brand build.

And now, we’re excited to offer even more ways to help our founders grow at scale, with increased offerings across sales, marketing, and pricing. Specifically, the team will be helping with:

• Sales & Marketing Leadership: Defining what “great” looks like; helping build the right team and structure for the next stage of growth.

• AI-Native GTM: Helping companies compete with new playbooks for agentic operations, AI-native RevOps, demand generation, and consumption-based pricing.

• GTM Strategy & Positioning: Defining ideal customer profiles, category narrative, and core messaging; aligning sales and marketing on go-to-market strategy, and ensuring budgets back the right priorities.

• Scaling Sales & Marketing Motions: Moving from founder-led sales to revenue engines, with the playbooks, campaigns, and metrics that drive predictable growth.

• Pricing & Packaging: Navigating pricing model transitions and packaging decisions as companies move from single-product to platform, including AI margin governance.

• Revenue Operations: Designing comp plans, managing usage-based pricing transitions, and building the AI-native operations that keep pace with growth.

The a16z Growth Platform is made up of elite operators who understand the challenges of growth-stage companies and how to execute in their respective domains. They bring actual game footage to every interaction with founders and CXOs. The team upleveling our sales, marketing, and pricing efforts is no different, having spent time in-house or closely advising companies such as Lovable, Atlassian, Samsara, 1Password, Miro, PagerDuty, Segment, Workday, and others during their hypergrowth and pre-IPO years.

It’s time to grow

Top operators today are in the biggest job of their lives, in the fastest-moving, most contested market we’ve ever seen. We know it takes both capital and operational know-how to win, and we are building to meet these demands.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

To hear more about Machine Age Fund, watch the companion podcast on the A16Z Show.

We’ve raised $1.1B for a16z’s newest fund: the Machine Age Fund. It’s time to open the throttle and accelerate the physical buildout of AI: the strongest tool ever developed for solving problems and bestowing abundance. It is our social and national imperative.

For anyone paying attention, this shouldn’t be a surprise. We’re seeing an upward inflection point in utility and variety of work that AI can do, just over the past year. And as we evolve from chat to reasoning to coding and other forms of knowledge work, both the demand for work and the token intensity of the work increase by orders of magnitude. Machine intelligence is going vertical.

This fund will invest into all of the computer infrastructure on which AI runs, including chips, memory, networking, and storage. It also includes full systems for running AI: from data centers to robotics to home AI appliances. The common thread among all of these layers of the AI stack is that they are all hitting the wall of today’s supply chain capability, and the limits of physics and computer science. So there is an urgent need for innovation and investment, and a once-in-a-generation opportunity to rearchitect them as platforms, all the way down to the electricity.

Reinventions like this happen at each epoch of compute, from the shift from mainframe to client server, to the Internet, cloud and mobile. However, what is different this time is the magnitude and breadth of change needed to keep up with demand, and the speed required to do so. Everything needs an upgrade, now:

• Compute density per rack increased by 28X from an H100 rack to a Rubin rack.

• Networking within a rack has grown similarly, hitting the limits of copper cabling

• Rack power moved from roughly 5-10 kW to 100-250 kW to support today’s systems and will increase to 1MW over the next 3 years.

• Data center scale is moving from tens to hundreds of MW and, in some cases, GW-scale campuses.

• Power is increasingly supplied not only from grid-only to grid plus behind-the-meter or captive sources.

We need faster, more efficient systems. We need cheaper and higher-bandwidth memory across the memory hierarchy. We need faster and more scalable interconnects between nodes and systems. We need power efficient edge devices for AI to explore and interact with the world. And of course we need all the cooling, materials, electrical, and real estate build out to support them. The hardware industry supply side is used to growing 20% to 30% per year at most; not the triple-digit growth that’s needed to catch up with demand. This will change, quickly.

Founders understand the opportunity. Over the last couple of years we’ve seen hardware startups grow from a small amount of deal flow to now over 20%. While we at a16z are generally known as software investors, we have always paid attention to hardware. Recently we’ve backed a number of hardware companies including Unconventional AI, Nexthop, Volta, Atoms, Heron Power and Mind Robotics. And we have selectively invested in hardware companies since our earliest days. In 2016, we led the series A in Skydio; we invested in SpaceX; we wrote our first check into Anduril in 2019, and were among the first venture investors in Waymo’s 2020 raise.

Hardware is in our team DNA, and it’s great to get back to our roots. Guido Appenzeller was the CTO for Intel’s Data Center Group (DPG). Raghu Raghuram and Martin Casado spent multiple decades in the data center space with system software, much of which required deep hardware partnership. Shangda Xu and David George have led investments across the AI infrastructure stack, from silicon and networking to large-scale systems and compute platforms. David Ulevitch and Erin Price-Wright lead many of the firm’s hardware and U.S. manufacturing investments through our American Dynamism practice.

We’re excited to make hardware an official motion for a16z and announce the Machine Age Fund to support it. Our GTM, talent, and marketing machine that we’ve built at a16z is now ready to serve hardware founders. And the a16z network has exceptional access to customers, suppliers, and everybody else that’s relevant, for a new age of hardware reinvention.

Ambitious founders who are reinventing AI hardware and founding the machine age, please reach out. We’re ready for 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.

America | Tech | Opinion | Culture | Charts

Token pricing began in the right place: the model layer. When OpenAI launched its API in 2020, charging for the computation a model consumed was a sensible way to meter raw inference. However, ChatGPT’s debut two years later helped spark a wave of applications built on that infrastructure that do much more. These new products combine proprietary data, tools, orchestration, integrations, and workflow logic to complete work on a customer’s behalf.

When an application prices that work in tokens, it imports the model provider’s cost structure into the relationship with the customer and anchors the product’s value to a unit whose cost keeps falling. Based on our work, it’s often a mistake to carry the model layer’s pricing logic into the application layer.

Instead, we believe companies should price at the highest layer of value that they can reliably measure, attribute, and defend.

• If you sell model access, price tokens.

• If you turn models into useful work, price the recognizable value unit, often through credits.

• If you deliver a clear and attributable business result, price the outcome.

Getting this wrong is difficult to undo. A token-based price trains customers to compare an application with raw compute. It gives away the value of data, workflow, and orchestration; exposes customers to technical complexity they cannot forecast; and can lock the vendor into weak margins. Credits do not solve that problem on their own if they are merely cost-plus tokens; they still meter infrastructure as opposed to value.

Pricing around the recognizable work does the opposite. It makes value legible, spend forecastable, and product improvements economically valuable to both sides. In our survey of 50 technical AI buyers, 27 preferred credits tied to recognizable work, while only 14 preferred tokens.

1. Price the Layer You Sell

The two ends of the stack are relatively straightforward. Model providers sell inference, which can be metered in tokens. Some applications deliver outcomes that are observable and attributable enough to price directly.

The middle is harder. That’s where most AI applications live.

For example, an account-research agent is not selling searches and model calls. It is selling a completed account brief. A coding agent is selling an implemented change. A data platform might sell a completed query, pipeline, or agent run. The application’s job is to abstract the complexity beneath that unit of work.

Because every category packages value differently, there is no universal AI application metric. Voice AI may begin with minutes, then move toward conversations resolved. Copilots may begin with seats, then add usage-based pricing as agentic work creates greater variation in cost and value.

The right question is not, “What is the AI pricing metric?” It is: “What unit of value does the customer already understand, and can we measure it consistently?”

2. Customers Want Legibility, Not Tokens

Customers will still ask about tokens. Usually, they are asking for one of two things: comparability or control.

First, buyers want a common benchmark. Tokens appear to let them compare a specialized application with a general-purpose model API or an internal build. But the comparison is usually false precision. Each application combines different models, data, tools, and levels of automation. A token flowing through one product does not produce the same work as a token flowing through another.

Second, buyers want to allocate spend. Finance and IT teams need to trace AI spend to a department, project, client, or invoice, often across a growing portfolio of applications. Requiring them to forecast tokens separately for every product recreates the complexity those products are meant to hide.

Both needs are reasonable. Nonetheless, token pricing often creates more friction than clarity.

A support leader can estimate how many conversations the company handles. It is much harder to predict context length, retrieval volume, retries, reasoning time, or output tokens. What should be a straightforward ROI calculation becomes a separate compute-forecasting exercise for every AI application the company deploys.

The better answer is to expose enough underlying usage detail to build trust without turning that usage into the commercial unit. Show customers what work was completed, where capacity went, and why certain tasks consumed more than others. Give finance and IT the reporting they need for budgets and chargebacks.

Transparency does not require the billing meter and the underlying cost meter to be the same.

In highly competitive or technically sophisticated markets, companies may still need to offer token pass-through, particularly for unusually expensive or volatile model usage. But this should be an explicit component of a hybrid model, not the default expression of the product’s value.

3. Credits Should Map Work to Value, Not Hide Tokens

For the broad middle of the AI stack, credits can be an effective way to package variable work. But a credit is a currency, not a value metric. What matters is what the credit buys.

A weak credit system converts token counts into an opaque internal currency. It hides the meter without improving it.

A strong credit system maps to work the customer recognizes. It might use a few intuitive effort bands: a small bug fix costs less than a multi-file feature; summarizing one contract clause costs less than reviewing the full agreement; enriching one record costs less than running a multi-step account research workflow.

The best credit systems do three things:

• Abstract infrastructure complexity. Customers buy the work, not the ingredients.

• Explain relative effort. Simple work consumes a little; standard work consumes a predictable amount; complex work consumes more.

• Create commercial flexibility. One pool can cover several workloads, agents, or automations while procurement manages one contract.

The test is comprehension. Buyers usually know their workload before they know their compute load. If customers cannot understand the credit system in a few sentences, it is too complicated.

4. Credits Can Protect Margins

Credits do more than just make usage understandable. Designed well, they also protect your gross margins.

In traditional SaaS, an additional user often adds little marginal cost. In AI applications, every user can generate inference, retrieval, search, tool calls, third-party data, media generation, retries, and failed runs. Fast growth can hide a weak business if each new dollar of revenue is quickly paid back to model, cloud, or data providers.

The pricing system needs to separate two decisions:

1. What is the work worth? Customer value, willingness to pay, and competition determine the price of the credit pool

2. What does the work cost to deliver? Relative cost and complexity determine how many credits each task consumes

This separation enables the vendor to protect margin while preserving the margin upside from model routing, caching, prompt optimization, better infrastructure, and a more diverse (and changing!) mix of proprietary and open-source models. In fact, if underlying model costs fall, the application may even be able to retain part of the benefit because customers are paying for work, not just reimbursing the company for compute.

Clay offers a helpful example and has often been ahead of the curve in how they think about pricing. In a 2026 pricing memo, the company wrote that it had mispriced credits in its Pro segment back in 2022 and operated that segment at a loss for years. Its new model separates Data Credits for third-party data from Actions for orchestration work. Clay keeps fixed pricing for models with predictable costs while passing through the actual token cost of more volatile and expensive reasoning models without a markup. In other words, Clay uses different meters for different layers: Actions for platform value and token pass-through for unpredictable model costs.1

5. Move to Outcomes When Value Is Clear

Credits are most useful when the product performs valuable work but the final business result is not cleanly attributable. Once the outcome becomes observable, attributable, and valuable enough to support a stable price, the meter should move again.

Under those conditions, price the outcome: a resolved support conversation, a qualified lead, a booked meeting, a processed claim, or a recovered dollar.

If the results are not attributable enough, price the unit of work through credits.

If the customer is buying raw model access, price tokens.

Many products will use hybrids: seats for access, credits for variable work, token pass-through for unusually expensive or unpredictable model calls, and outcome fees where attribution is clean. Multiple meters are not the problem, but avoid using the wrong meter for the wrong layer.

Make Value Visible

Token pricing pulls the customer conversation toward a cost curve that keeps falling. This is a poor anchor for a product whose usefulness, reliability, and role in the workflow should keep rising.

The better path is to price at the highest layer of value you can reliably measure, attribute, and defend. Translate variable work into understandable units. Use credits to package those units when flexibility matters. Move toward outcomes as soon as customers can recognize and trust them.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

Most forms of theft take something from you. You notice the car is gone, the account is drained, the files are locked.

Brand impersonation is a strange one. Nothing of yours ever leaves. A copy of you, your company, your logo, your login page, goes to work somewhere else on the internet, collecting passwords and payments in your name. And, to make matters worse, if you’re being copied you’re usually the last to find out.

That is what makes it so hard to fight. Impersonation is hard to find, because it runs on other people’s platforms, aimed at other people, in your name. It is harder to stop, because a single takedown rarely ends it. Like the Hydra of Greek myth, cut off one head and two grow back: a new lookalike domain, a fresh ad, another cloned account. How did Hercules beat Hydra? Only by searing each neck shut so nothing could regrow.

The modern equivalent is an always-on social engineering defense, one that watches every surface at once and cauterizes the regrowth so the heads stay gone. The machinery for stopping impersonation has never been larger. Every big platform runs a trust-and-safety team, brand protection is a mature software category, and ‘AI-powered detection’ sits on every vendor’s homepage.

The losses keep climbing anyway. Americans reported losing $3.5 billion to imposter scams in 2025, up ~19% from $2.95 billion the year before, and imposter scams were the single most-reported category of fraud, according to the FTC. We’re seeing more detection and more losses, at the same time.

Doppel hunts fake versions of real brands (all industries) across the internet. The fakes turn up wherever their customers are: lookalike domains, social accounts, mobile apps, paid ads, marketplaces, and messaging. Doppel’s Threat Graph sits across the surfaces where impersonation lives: social, paid ads, domains, mobile app stores, and messaging, and it records what happens to a fake from the moment it is spotted to the moment it’s brought down. These are confirmed fakes, reviewed by analysts, across hundreds of brands, over the four quarters from spring 2025 to spring 2026.

For once, we do not have to ask the victims what happened. We can watch. To map this, we have to look at the graph: the shared infrastructure connecting seemingly isolated attacks.

Most impersonation goes undetected

Let’s start with how many impersonators companies actually notice.

Before continuous monitoring begins, brands are nearly blind to their own impersonation: across their first 90 days with Doppel, brands and their existing partners surfaced only about 9% of the confirmed fakes targeting them. Doppel found the other 91%.

94% of brands caught fewer than half of their own impersonators, and the median brand caught none at all. Impersonation isn’t a visibility problem at the margins; for most brands, the fakes simply never cross their radar.

In general, less than 10% of all brand impersonations are self-discovered:

This is not a knock on those teams. Impersonation lives outside a company’s walls, on platforms they don’t control, aimed at people who aren’t their employees. No internal security stack can see it, no matter how good. That blind spot is the whole reason continuous external monitoring exists as a category.

Impersonators usually operate on multiple platforms

Seeing it is only half the challenge, because the attack is built to appear in many places at once. There is no single pane of glass. A campaign is many-headed by design: a head on social, another on a search ad, another on a lookalike domain, all at once, and operating 24/7 with the help of AI.

The usual human math doesn’t apply here. Spinning up a new head costs the attacker minutes; finding and cutting one off costs a defender hours, and increasingly the heads appear everywhere at the same time: near-simultaneous multi-surface campaigns have grown roughly sevenfold in two years. No team scales with that by hand.

Over 80% of impersonated brands are faked across two or more surfaces at once. A tool that only watches domain registrations, which is what most legacy takedown vendors were built to do, only catches a ~6% sliver of the picture, and misses the rest of it entirely.

Of the brands that had confirmed impersonations, 80.6% were hit on two or more of the surfaces monitored for them. That figure only counts surfaces each brand actually has monitored: a brand that’s only bought domain and social modules can never show up as more than two surfaces, however many surfaces attackers actually use against them. So the true multi-surface rate is almost certainly higher than what’s shown here; this is a floor, not a ceiling.

Where those surfaces land shifts by industry:

Software and media brands are impersonated almost entirely on social media. Healthcare and manufacturing get hit mostly through mobile apps. Hospitality lives on lookalike domains. There is no universal channel, and a defense that watches only one surface stays blind to most of the attack.

Impersonations are hard to kill

Say a brand does see all of it. Getting it down is the next fight. Once a fake is found, someone still has to get it removed, and a brand almost never owns the platform it lives on. So the main lever is the takedown request, a formal note to whoever hosts the fake asking them to pull it down. Most of those amount to little more than ‘please remove this.’

Of every threat Doppel filed a takedown request on over the past year, roughly 97% had their content removed and/or the hosting pulled. The 3% that survive cluster almost entirely on the handful of platform types that ignore takedown requests altogether.

Even the fakes that do come down are not down instantly, and the harm is not spread evenly across a fake’s life. What stretches takedown times into weeks is not slow removals; it is the platforms that never answer at all.

The good news is that half of all fakes are gone within a couple of days. A stubborn few live for months, and that long tail is where most of the total exposure sits, costing customers their money or a brand its trust. The longer a fake stays live, the more time it has to reach new victims, steal credentials or payments, spread across channels, and repeatedly erode trust in the brand.

Where a fake lands on that curve depends less on the fake than on who you have to ask to remove it, and you do not get to pick who that is. The clock belongs to the platform, each with its own timeline and willingness to act, and most of a fake’s life is the wait for the host to move.

Dirty domains (no, not those) deserve their own look.

Once removed, a fake page seldom returns. The domain behind it usually survives, because the industry grants blocking and content removal far more readily than registrar-level removal, and it re-arms fast: almost 60% of taken-down domains serve a new confirmed fake within 24 hours, and close to 70% within 90 days. One-asset-at-a-time enforcement is a treadmill no vendor escapes.

About half of malicious domains are registered and pointed at a brand within a month. But over one-in-five are aged domains that sat quietly for years before someone woke them up. Domains are the Jason Bourne-style sleeper cells of the scam world. In fact, the oldest in this data had been sitting untouched for the better part of a decade before it was aimed at a brand.

Domains are also the one surface where beating a scam doesn’t last.

Take a bad domain down and, hydra-like, a fresh head grows back on the same neck: most within 90 days, and often within a single day, on the same registrable name. Every other surface stays down once it is removed. On domains, a takedown buys you a day.

Across the board, most scams don’t go quietly into the night, and campaigns push back with an almost hydra-esque quality.

In the two weeks after a takedown, the same brand tends to see a rise in brand-new fakes.

Why? Threat actors keep trying variations and iterations to test the waters on what is and isn’t being blocked or monitored. And, with AI, this is done much faster.

Read one way, this is a losing battle: The services hosting the fakes may not answer, and the impersonations could grow back overnight. Read correctly, it’s a targeting problem. Most fakes end up dead; what survives is the operation behind it, respawning for pennies. Chasing fakes one at a time concedes that math. Mapping the operator’s infrastructure (every domain, account, and kit they run) and cutting it down at once is how the math flips.

A small number of impersonators do the lion’s share of the impersonating

So who is actually behind all this? Fewer operators than the noise suggests.

Group the fakes by who runs them and the busiest tenth accounts for nearly half of everything. A relatively small set of professional operations, rather than a planet of lone opportunists, produces most of the damage.

They also share tooling. Plenty of fakes are built from the same kit, a ready-made bundle of page code and templates that clones a brand’s login page in minutes. It would be easy if you could follow those connections to a single kingpin behind everything, but the data will not take you there.

Nine in ten shared fingerprints show up on just one brand, which is one operator reusing a tool against a single target and tying only their own fakes together. The rare fingerprints that span hundreds of brands turn out to be commoditized plumbing, a popular host or an off-the-shelf template that unrelated attackers all happen to use.

That is what makes fakes so hard to track. There is nothing central to grab hold of, no master list and no single cluster to strangle. The operators are few, but they hide behind disposable, off-the-shelf infrastructure they can rebuild overnight, so watching any one fake tells you almost nothing. Keeping pace means watching every surface at once and without pause, because a fake taken down is usually a fake already regrowing somewhere else. The heads keep coming back.

Fake detection doesn’t have to be a Herculean task

So what actually stops this? An always-on social engineering defense does three things at once: it watches every surface continuously, it acts on a fake within minutes of finding it, and it tracks the operator behind the regrowth rather than just the artifact in front of it. Recurrence and the balloon effect are why: kill a single fake and a new head grows back. And because a few operators and shared kits drive most of it, mapping them aims at the half of the problem that matters, instead of chasing artifacts one at a time.

Which brings us back to the copy of you, out there on the internet, working under your name while you are the last to know. Finding the fake was only the start. The hard part is everything that comes after.

This is where the modern internet can borrow from the ancient Greek tales, with Doppel serving as the Hercules to the scammers’ Hydra. Hercules did not win by swinging faster. He won by stopping the necks from sprouting new heads, searing each one shut before it could grow back. The takedown-by-takedown fight is the swinging. The win is an always-on social engineering defense: watching every surface at once and choking off the regrowth, so the heads stay gone.

Methodology

All figures come from Doppel’s Threat Graph and cover confirmed, analyst-validated brand impersonation across hundreds of brands over Q2 2025 through Q1 2026 (April 2025 to March 2026). The data is customer-anonymized, with raw volume counts withheld and shown as shares or indices. Coverage is impersonation only, across social media, paid ads, lookalike domains, mobile apps, marketplaces, messaging and telecom, and crypto and wallet services.

Multi-surface figures are bounded by two factors: the brand population (Doppel-monitored brands with at least one confirmed impersonation in the window) and the surfaces counted per brand (only the surfaces monitored for that brand, per purchased modules). Both factors cap the published rate from above. The true multi-surface rate is very likely higher.

Takedown outcome figures measure end-state removal (content and/or hosting no longer live) rather than formal per-request acceptance, across all threats with at least one takedown request filed in the window.

Takedown time figures measure median time from filing to removal, calculated only over requests that resulted in removal, across all phishing/impersonation takedown requests filed in the window.

Domain recurrence figures require the follow-on alert to be analyst-confirmed malicious (the same standard used elsewhere in this study), measured over a full 90-day post-removal observation window per domain.

Domain age is measured from a domain’s registration date to first detection. Operators are grouped by shared technical fingerprints such as page code, tracking IDs, hosting, and kits. A few summary statistics were calculated by a16z from Doppel’s data.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

Today we’re trying something new. We’re sharing a slide deck, “Intelligence is the Primitive. Applications are the Diffusion Layer” by , that he originally put together for a smaller audience. There’s a lot of great stuff that gets made internally at a16z, and we look forward to sharing more of it here on a regular basis, when we’re able. - Danco

Overview: what is the overall situation?

Apps & the Application Layer

Consumer

Thank you for reading! If people like them, we’ll publish these on a regular basis.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

Today’s post was done in collaboration with Protege. Please also see our recent podcast with Engy Ziedan, Protege’s Chief Scientific Officer.

You go to the doctor, and the doctor tells you: “The medical student will be taking care of you today. They’re very capable; they’ve scored 100% on every single test in all of med school. Not a single question wrong. You’re in good hands!”

Something seems off, right?

This is the big hairy question for AI adoption in healthcare. AI is now extraordinarily good at tasks where you can train it on “right answers,” supervise as it learns the craft, and then hand over some tasks. The more training data and compute you give it, the better it gets. The smaller the gap between “Is it good at the test?” and “Is it good at real life?” the better AI is going to do.

The AI healthcare dream paints this picture: “We’ll train AI on what ‘health’ looks like, on every piece of clinical data we can find, and it’ll be much better than us at spotting problems and making recommendations.” It’s a great dream, and we should pursue it.

But there is a big challenge with healthcare that is harder than most domains. How do you evaluate “Is this AI any good?” It’s not so simple for at least three big reasons:

First, practical evals in medicine are subjective. Clinical decisions are hard to work with as a training data set. Doctors make decisions with subtle and personal idiosyncrasies that are notoriously hard to standardize against. If you’re training an AI on clinical data, any benchmark or eval you design is going to be quietly opinionated, scored against what the physician did that particular day (and not “did they live a healthy life forever after?”). Your evals are a reflection of what you would do in the moment, not what’s absolutely correct - there is no shared, absolute definition of “ground truth” in medicine.

Second, the underlying data is hidden. Everyone who is trying to build useful models in AI health has acquired various clinical and administrative data sets for training and eval design - and, understandably, they guard these datasets tightly.

Third, evals and benchmarks unavoidably turn into marketing. Not because companies are acting duplicitously, or underhanded in any way; it’s just what happens. “The model did well on the benchmark” is as much a statement about how you wrote the test; not a statement about its clinical merit versus a different model, or a different evaluation. It’s a kind of principal-agent problem.

Health AI has an Oracle Problem

The only way to objectively measure whether the model was right in its decision is to follow the patient forward, for a long enough time - which is a subjective decision of its own. That’s a much, much harder kind of AI training than simply creating benchmarks and testing against them.

And it means it’s very hard for hospital networks and other buyers of AI products to actually evaluate who’s best, especially in a world where models get updated every few months - or less. If you ask the government to arbitrate, they might say “there should be a neutral task force that decides what’s best practice”, but that approach can’t possibly keep pace with model and product evolution.

The ultimate barrier to true AI adoption in hospitals is trust in what you’re getting, and that won’t be completely solved by more training data or more compute because of the subjectivity involved. Someone has to solve the Oracle Problem. That’s what we’re trying to do with Protege. It is not often admitted to, but it is creeping up as the problem in healthcare AI, as model capacity starts to surpass that of physicians in key applications.

Below, we illustrate some of these key issues using data from Protege’s partner network. For computational efficiency, we randomly sampled millions of records from nearly a trillion tokens of EMR text across billions of notes. The figures below are descriptive rather than causal estimates but can shed light on the key issues.

Usage Creates an Urgent Problem

OpenAI reports that more than 300 million people use ChatGPT for health-related questions each week. That number is likely an undercount, since it omits use in enterprise settings where firms hold zero data retention rights.

The data shows the same indications. Explicit references to patient AI use were essentially absent through 2023, then rose sharply following GPT launch. Part way through 2026, about 1,686 per million notes recorded interactions where the patient wanted clarification on something a model said. We see the same inflection in voice mail messages where patients call to explicitly ask about something they chatted with an AI model about.

AI is also being used more often to write medical notes; soon nearly one-third of all SOAP (Subjective, Objective, Assessment, and Plan) notes will be written by AI. These are the notes that are standard for documenting what happened during a patient visit.

We are interacting with models much more broadly when it comes to healthcare – as professionals and pedestrians alike, across age groups, and for both trivial questions and entire disease odysseys. These models can deliver more health for fewer resources. But models are ‘just’ making predictions. Predictions have a cost; they can nudge or sway people to make consequential choices. And, as with any prediction, it is only as good as we can judge and pre-judge the outcomes.

The “Right Answer” is Becoming Elusive

To measure model performance, we have models take tests known as benchmarks. A benchmark compares a model’s answer with what we call the right answer. But this is not automatically “ground truth”.

A classic example of this is knee replacements. If you hurt your knee and need surgery, you can get either a “partial” or a “total” knee replacement - and which one you get basically depends on who your surgeon is. If you look at a dataset of knee replacements, you’ll find that patient characteristics, comorbidities, facilities, and year only explain 3.4% of the variation in the choice to perform partial or full. But if you add the identity of the surgeon, it immediately becomes 14.8% – 77% of the explained variation in partial vs. full comes down to who the surgeon was.

You can think of this as “taste” and it’s sticky and idiosyncratic. And it’s why training AI on medicine is hard. It creates a scenario where a model may be “right, but off” because it did not agree with the individualistic approach of the physician it was graded against.

This is not peculiar to knees. Across 15 operations with two variations, physician identity accounts for 7% to 77% of the explained variation. In every turn and decision in a long-horizon medical record, the model is in fact going head to head against physicians’ individual styles and preferences.

A model can therefore disagree with the recorded action and still be clinically defensible. Conversely, it can agree with the action for the wrong reason. A static benchmark cannot distinguish the two.

The implication for benchmarks? If an evaluation contains only a few hundred cases, often one or two per clinical decision, whether the model is “right” can depend on which physician’s preference became the label. Adding more physicians to grade the answer may increase consensus. It does not necessarily recover the ground truth. The rarer the procedure and the more expert the task, the harder this becomes. Clinical decisions are dynamic; a doctor’s memory crosses cases.

Doctors are also hysteretic: a temporary shock can leave a lasting effect on future decisions. In medicine, one outcome can change how a physician treats the next patient.

For example, after a vaginal delivery complicated by hemorrhage, physicians become more likely to use a Cesarean in the next delivery. Given that admission for a birth within a close interval is predetermined, the move away from that modality for a few days indicates humans have recency bias – expected!

But will we account for that when we pairwise compare what happened in a clinical record and the model’s output?

We’re testing AI models on how they perform on set, pre-described tasks or problems in a vacuum. But healthcare is not performed in a vacuum; decisions are influenced by the patients the physician sees, the patients the physician doesn’t see, other clinicians, their capacity, what order care happens in over a period of time, and many other variables and context that are difficult to record.

The Answer is in the Records... if You Can Find It

By now the field has agreed to move away from synthetic vignettes and give the model full (de-identified) medical records. But there are still downstream issues.

Across a random set of real patients in the data, the median patient record contains about 8,500 tokens and the mean about 39,000. But five of six public healthcare AI benchmarks give the model less context than the median patient record. Several provide fewer than 200 tokens (less than 2.5% of the median record) per case.

So while the model is given full records, many benchmarks may still test a compressed vignette. Benchmarks do not necessarily represent the same tasks that clinicians actually perform, and they do not necessarily represent the same context either.

And to make things harder, full records aren’t foolproof either. Cleaning a record of bias is itself a science. Remove too little and the model receives the answer on a platter. Remove too much and the case becomes clinically unnatural.

Machine Leaderboards are Built from Human Decisions

Even after choosing cases and cleaning labels, a benchmark still contains many idiosyncratic scientific preferences: the harness, prompt, rubric, grader, verbosity penalty, answer format, and ordering of choices. These choices can change not only the score, but which model appears to win.

In a 19-way diagnosis task, we evaluated models across four random permutations of the same answer choices. The model saw the same case and the same 19 labels; only their positions changed. Yet models frequently changed answers!

Prompt wording can do the same. In a lab-trajectory task, we compared two models across four assessments: a base prompt, and then an experimental prompt where we added an additional sentence that mathematically defined how the model should classify a patient’s lab results, rather than let the model define it independently while holding encounters, charts, ground truth, and grader fixed. Adding this one sentence to the prompt collapsed the differences between the models, and flipped the winner outright on one of them.

This means something really interesting. It doesn’t mean “One model needed help.” It means that without the definition, the benchmark test was actually measuring “whose definition of ‘stable’ matches ours?”

Models also task-hack; they attempt to find ways to game the evaluation rather than solve the task. Ask for one diagnosis and a model may offer several, hoping one receives credit. Force one multiple-choice answer and position effects appear. Penalize verbosity and the evaluator must define verbosity. Search the chain of thought for words such as “reward,” “verifier,” or “ground truth,” and we detect only the forms of reward hacking that we already know how to name.

Medicine is likely to contain subtler forms of task-hacking that no regular expression will find.

Real Benchmarks Do Not Conflate Tests with Productivity

We no longer need AI models to pass the MCAT. What we need to answer is: “How productive is AI at getting the overall picture right, and making us healthy?”

It is pretty easy to measure the wrong thing. For example, a common mistake health economists make when measuring the value of a policy or technology is to conflate healthcare utilization with health. Similarly, hospital admissions and re-admissions are generally considered bad, but sometimes they are actually good! We still don’t know if these measures of throughput reflect health or simply the use of healthcare.

To measure the true effects of AI on health, we must begin to measure the impact it has on critical factors like life expectancy, quality-adjusted life years, nurse burnout, physician turnover, medical denials, medical debt, medical mistrust… the list goes on.

Importantly, none of today’s benchmarks measure misalignment in live care delivery settings. AI can display over-eagerness, sycophancy, and other forms of misalignment. Measuring this is increasingly important as technology becomes central to the delivery of care.

The fundamental question as we assess AI: With the same level of healthcare spending, are we able to produce better health when we adopt AI technology more quickly? These are the benchmarks we need across health systems, frontier models, vertical AI vendors, physicians, and teams within a system - not simply measures of how a model performs on an abstract case graded after the fact.

The Hard Thing About Benchmarks is the Verification Ceiling

Benchmarks have hit an issue medicine encountered decades ago. Trials measure efficacy, but the world cares about effectiveness. Today, benchmarks lack even a good measure of efficacy and are far away from effectiveness.

Training data and compute may not be the binding constraints. Instead, we are constrained by learning how to know when the model was right. And as AI is continually deployed in healthcare settings, all stakeholders - model developers, vertical AI application builders, healthcare systems, and adopters of health AI solutions - will need to better understand and agree upon how AI models perform in real-world settings.

Healthcare in the real world is messy and spans dozens of modalities. Protege exists to make sure AI creates more health, not just better scores. We use real world data from our data partner network to build better tests: tests that measure what healthcare AI must do for real clinicians and real patients to produce real health gains.

This was pretty niche for a general audience newsletter, so if you read this far then there’s a good chance you’re also working to close that gap, or maybe should be! We’d love to hear from you. You can read more at https://withprotege.ai/model-builders/healthcare, and reach out at https://withprotege.ai/contact.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

Winds of Thematic Change

ETFs have become an increasingly prominent feature of public markets, especially for retail. Among other flavors, there are thematic ETFs, active and passive ETFs, index-tracking ETFs, credit ETFs, and all of them offered with varying degrees of leverage and without. Why it’s the case that ETFs are on a tear is some combination of low fees, low barriers to entry, easier distribution, and good marketing (as well as more retail market participation, generally).

According to data from Citadel, ETF net inflows are pacing for their strongest year ever, with July setting an all-time record.

One thing that ETFs are definitely able to do is capitalize on whatever happens to be the theme du jour. “Oh, you think robots are the next big thing? We’ve got an ETF for that. You want to ride the memory rocketship? You’ll love the taste of DRAM. And for a more adventurous experience, it pairs nicely with PHOX, our recent photonics vintage.”

That’s interesting so far as it goes, but what’s perhaps more interesting is just some perspective on how much the ETF “themes” have shifted in a relatively short period of time:

Back in 2020, the top 5 themes included clean energy, emerging markets tech, and healthcare. By 2026, the thematic story is completely different: AI, Nuclear, Space, Defense and Infrastructure are dominating the leaderboard.

It’s a total ETF takeover for the ambitious, capital-intensive world of atoms over bits. Whether it all works out is not for Charts to say, but ngl, 2026 is way more exciting.

Data Centers Are A Blue Collar Bonanza

People may love data centers in their ETFs, but for whatever reason, they increasingly do not love data centers anywhere remotely nearby.

Without weighing in too strongly into that whole debate (not yet, at least), suffice it to say that, like ‘em or not, data centers are economically a very big deal, and in some cases, the biggest deal in town.

For the states actually building data centers, the buildout is a substantial chunk of all non-residential construction spending:

States like New Mexico and Wyoming aren’t actually building all that much—less than 3 gigawatts under construction—but because neither state builds much of anything, data centers represent ~60%(!) of all private non-residential construction. Even for a much larger state like Pennsylvania, just ~3 GW of data center is nearly 30% of non-resi spending. Texas, by contrast, is building far more capacity, but it amounts to just 10% of total spending, which is still quite substantial, just not as substantial as 60%.

In all events, whatever one’s views of data centers, the fact remains that they are one of the more important (if not most important) economic impulses in the game.

Wells Fargo took a stab at tabulating some of the “significant economic benefits” that run concurrently with data centers (either in operation or under construction):

Since 2024, for the counties with already operating data centers, the picture is uniformly better:

• more housing,

• higher home values,

• less unemployment; and

• more job growth.

For the counties currently building data centers, the employment picture is definitely better, but arguably the drawdown in new housing has been worse and home values haven’t appreciated quite as much.

To be fair, causation and correlation are not obvious here. A lot of extant (and new) data centers are in Loudoun County, VA, one of the wealthiest counties in the country. Likewise, a lot of the new data center builds are in Texas, which experienced an historic level of residential building (and home value appreciation) prior to 2024, so the housing retreat is coming off a much higher base.

But, to return to the point above, building data centers is almost certainly good for the employment picture. In addition to the hundreds and thousands of new hard hat jobs that are being created, data centers pay a lot better than comparable employers:

According to data from Indeed, the wage premium offered by data centers runs as high as 64% (for a facilities manager) to a still substantial 10% (for electrical engineers).

Data centers are making it an increasingly good time to be in the skilled trades. A recent comment in a Dallas Fed Report highlights the point:

A heavy industrial construction contractor said, “We have been paying what I believe to be a very competitive wage for skilled concrete workers, $28–32 per hour. The data centers are offering $45 per hour and a $150 per diem for concrete workers.”

That’s a ~50% wage premium for data center concrete workers—a life-changing kind of raise.

Data from ADP tells a similar story. If you look at the pay premium for job-switchers—usually a proxy for where demand is hot, since a pay bump is necessary to induce the switch—the premium is absolutely cooking for data center-adjacent categories:

Wage-growth for job-switchers in Construction, Manufacturing, and Nat’l Resources/Mining is 6-9.5 pp higher than for job-stayers. That job-switching premium is far higher than in any other sector.

Again, the point here isn’t to write the definitive case for data centers, one way or another. The point is, however, to note that saying “no” to data centers almost certainly involves saying “no” to much higher wages for blue collar workers (and the most substantial investment impulse around).

It’s Not Your Imagination–Ridesharing Really Is More Expensive

If it feels like taking an Uber is more expensive these days, then that’s probably because it is more expensive, at least according to data from Gridwise Analytics.

Since 2024, both the average and median Uber fare have increased ~20% (and appear to be rising still):

Notably, it’s only Uber that’s gotten more expensive over that period—the median and average Lyft rides are slightly cheaper than they were at the beginning of 2024 (and run ~24% cheaper than Uber overall), although prices are rising lately.

A big part of the price increase seems to originate with rising platform fees, especially for Uber:

Uber’s platform fees have been rising for the past year+ (and the median really stepped up in October), while Lyft’s platform fee only started rising recently, after a period of substantial decline.

Good for Uber and good for Lyft.

To be fair, it’s not only good for Uber and good for Lyft—it’s also been good for Uber/Lyft drivers:

Average gross driver pay per trip has also been rising since 2024, and recently hit an all-time high.

It remains to be seen over the longer term if higher ridesharing prices put any downward pressure on demand, but for now, the rides are getting more expensive, and the benefits are flowing through to both the platforms and the drivers.

Another interesting aside is that while the overall hiring market remains pretty tepid, in addition to more people starting their own businesses, more people are pursuing “gig work,” as well—presumably, drawn by the increased pay (at least, in part).

But while ridesharing gig work is getting more popular, it’s far from the fastest growing kind of gig. The hottest new thing in gig-work is unequivocally “social commerce,” i.e. QVC for the social media age:

Among BofA’s customer accounts, all categories of gig-work have increased (except for vacation rentals), but social commerce has increased by more than 30%, which is way more than any other category, albeit off a smaller base.

Why precisely social commerce is booming is anyone’s guess. It could be shifting media consumption (from tv to social) combined with the broader ecomm boom. It could also very well be that AI has lowered the cost of all the other aspects of running a social media QVC-styled operation, making social commerce overall a more sustainable vocation for more people. Or maybe that Insta ad-targeting is just too good to resist. Or all of the above.

Agents Are Early, But They’re Already Changing Things (And More Tales From The Frontier)

Charts has observed before that while AI demand is rising overall, the growth is far from normally distributed. There is a huge divergence in usage between the median and the power users, and some recent data from OAI suggests that the divergence is getting even bigger.

Output tokens have more or less doubled across the typical firm, but the top decile of firms is pulling farther and farther away:

Across all industries, there’s an ~8-fold gap in token output between the typical and top-decile enterprise, where the latter has increased output by more than 17x since April 2025. Within select industries, the gap is even wider, rising to almost 12-fold in Information (i.e. tech), where the top decile of firms is outputting 32.5x(!) as many tokens as they were a little over a year ago.

Unsurprisingly, the power users aren’t generating more tokens because they’re chatting with their AI companions. Power users have mostly graduated beyond chat to more advanced AI tools:

Plug-in and Skills adoption for the top decile of enterprises is ~2x and 6x as high as the typical firm, respectively, although even the power users have a long way to go before they reach OAI’s own internal adoption levels.

Interestingly, though, tech isn’t the fastest growing adopter of higher powered tools, at least of late. If you want to measure usage sophistication by Codex adoption, then knowledge workers are getting more sophisticated across the board, but none more so than legal:

Legal workers have increased their Codex adoption by a staggering 108x since February of 2026. Query how much of this shift is driven by the broader Codex rollout, but either way, Legal stands out from the pack.

The other implication from this shift to more sophisticated AI adoption is that it’s changing both the growth and composition of token consumption dramatically.

While only a small portion of AI adopters are actually using fully-deployed agents, agents are nonetheless driving a massive share of tokens:

According to data from OpenRouter (and chartmaker Peter Walker), agents are using nearly 5x as many tokens as humans, and agent usage has grown ~14X since just February.

Not only do agents use far more tokens than people, they use tokens very differently. Agents do far more work with cached tokens than people do:

Again, according to OpenRouter, more than 85% of agentic token burn comes from the cached prompt—and cached tokens account for nearly all of the relative growth of token usage.

The reason is pretty straightforward: while humans tend to have a “prompt and response” styled dialogue, agents are designed to iterate (and iterate) towards a goal. The initial prompt contains the token-intensive “pre-fill” (i.e. the core context around the goal, whether it’s policies and procedures, code guidelines, etc.), but then the agent begins to read-write incrementally—progressively adding to the cache—as the agent progresses towards a final result.

Cached tokens are far cheaper than the pre-fill—which is good, and makes the economics of agents increasingly viable—but are also (by definition) memory-consumptive. You see, now, why high bandwidth memory especially is in such hot demand—all those agents busily at work (without going back to the drawing board with each cycle) need memory to cook. And agents are proliferating rapidly, despite being very very early in their overall development.

There’s another possible second-order effect of agents worth noting: agents may be decreasing demand for legacy automation and workflow tools.

According to data from Similarweb, outside of Gumloop, traffic to automation tooling sites has been in serial decline:

N8N, Zapier, and Make all predate LLMs, and while they dominate the overall automation segment (by site visits), each is experiencing double-digit declines on a trailing 12-week basis. Gumloop, by contrast, was launched in 2023 as an “AI native agent builder,” and is the only automation platform gaining traction (according to Similarweb).

It’s still far too early to stick a fork in Zapier et al—they have AI at their disposal, too, after all—but agents have barely arrived, and they’re already beginning to change the landscape in substantial ways.

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.

America | Tech | Opinion | Culture | Charts

To hear more from Gabriel and Angela on Borderless Founders, watch the companion podcast on the A16Z Show.

Silicon Valley has always run on the magic of paying it forward. Everyone here has been helped by someone, at a critical moment - for no reason other than just to be helpful.

There’s a special version of this pay-it-forward culture for international founders. When you are from another country and you make it here, you find a community of people who have done the same. They, too, left behind their friends, family and primary language to pursue an idea they couldn’t quit. So when these founders get together, believe us, they have each other’s backs.

The emergence of these international-tech-mafias has changed the playing field for international founders. There is now an advantage to having one foot in your home country, and one foot in Silicon Valley. You know the strongest engineers in your home market before the rest of the market catches on. You can get early customers from large companies at home that give you big proof points early. Your diaspora will answer your calls before your traction would normally warrant it - future hires, advisors, and design partners.

We call these people “Borderless Founders.” If that’s you, you’ve already excelled across multiple dimensions at home. And you gain a genuine, underappreciated advantage in company-building by building your Silicon Valley reputation while compounding your network abroad. But - here’s the thing - you have to make it in the door here, first. You need to find someone who’ll sponsor your visa; you need to find somewhere in SF to crash. And before you even think about those things, you need the right introductions to Silicon Valley while you’re still back home.

Borderless Founders: the a16z network is here to help you.

The Borderless Founder advantage

Startups are an exercise in preferential attachment. As Marc Andreessen says:

Borderless founders possess three critical advantages when it comes to driving preferential attachment for their companies:

Customers: Knowing leaders at home and across their diaspora gives Borderless Founders an outsized ability to land key design partners or customers early. This early traction with large organizations builds enterprise credibility much faster.

Pit, a seed-stage company building AI platforms in Stockholm, closed H&M (a publicly listed company) and Stena Metall (a multi-billion dollar private firm in the industrial sector) as its first customers. While Pit’s CEO, Adam Jaffer, was previously the founder of Voi Mobility, he was not yet a household name on the global stage. His local reputation enabled him to land local brand name customers sooner, putting Pit on a global path much faster.

Supersonik founder Daniel Carmona had won global hackathons and bootstrapped several companies in Spain, but he was not yet well-known in the U.S. He leveraged the Spanish diaspora to connect with the CRO of Salesforce—who had studied at the same university in Barcelona—and secured Salesforce as their first major contract. Daniel’s reputation among trusted Spanish peers as a top-tier expat helped accelerate the deal.

Brand: Founders can receive significant press coverage in their home markets and become representatives of their country’s technology sector. Strong recognition early in their lifecycle can accelerate their brand internationally.

For ElevenLabs, the Polish government recently joined the cap table and even opened an AI lab in Poland with the goal of inspiring more AI companies from its 45,000 technical graduates each year. This triggered a wave of international press for ElevenLabs.

Talent: This may be the most significant advantage of all. In today’s fierce market for talent, especially in the Bay Area, Borderless Founders can identify and hire the best talent in their home countries before the rest of the market catches on.

Segura, an insurance AI company in Brazil, recruits from lesser-known university robotics teams that attract some of the best and most highly motivated students. Pit knew several former founders ready to “get off the bench” and try again. The company hired 6 technical leaders who had each served as CTO at a Swedish unicorn. Tako identifies high slope Brazilians in the US who are looking for a big mission to work on back in their home country.

Just look at the data!

A16Z’s borderless playbook

A16z was founded to give founders a bridge loan of power and momentum on their journey toward inevitability. We help Borderless Founders accelerate their preferential attachment, when they first land in the U.S. and as they scale globally.

Here is our playbook (built over 1M+ airmiles!):

Map the nodes from each country and help them continue to “pay it forward.” Start with local luminaries: the founders and operators who built significant companies in the region and are trusted by the next generation. Each country has a handful of founders focused on giving back. In Latin America, these include Sebas Mejia (co-founder of Rappi), Guillermo Rauch (founder of Vercel), and Pedro Franceschi (co-founder of Brex). In Europe, these include Fredrik Hjelm (founder of Voi) and Kristo Käärmann (founder of Wise), among many others.

We begin by hosting intimate dinners where established founders share their journeys with a curated group of next-generation founders (but in the meantime, potential hires). These luminaries already enjoy giving back; we simply make it scalable by handling the curation. By bringing top talent into the room, we help established founders with recruiting and branding. As one attendee recently noted: “He was so impressive! I’ve been telling all my friends they should go work there!” The best future founders often seek to work for these luminaries, viewing the experience as the ultimate “CEO training.”

Go deep with the relationships that matter most—in person. A physical presence in home countries is vital for building trust. Ben Horowitz hosted a dinner in Stockholm for a group of Swedish builders. The government representatives in attendance better appreciated the momentum of the Swedish tech sector (including companies like Lovable and Legora) and doubled down on their support. They provided ministerial access, included the companies in trade diplomacy, and featured them in regulatory discussions.

Consistent presence matters more than one-off trips. Gabe flew to Sweden nine times last year!

Sometimes these luminaries become repeat founders themselves! Two people at the table, who had previously built Kry and Voi, later started new companies, and we invested in both at seed.

Build strong communities within each diaspora. Help the next generation of talent before it becomes obvious. We build relationships with people in major global cities who remain connected to their home countries. By engaging with a network across all experience levels, we identify the smartest people early in their careers.

We aim to be the “welcome committee” for the smartest global founders (regardless of where they choose to place their HQ).

Our Borderless Founder dinner series has hosted 600+ founders and founding hires over the past 5 years. Our most recent speaker was Victor Cardenas, CEO of Slash, which recently achieved unicorn status. He remembers attending a Borderless Founder dinner three years ago, when Slash was just getting started.

We play the long game. At a Borderless Dinner where Guillermo spoke about dropping out of high school to build NextJS, another guest mentioned Gabriel Petersson, a student in Sweden with a similar background. Gabriel Vasquez reached out to Petersson and began building a relationship long before he joined OpenAI. We supported him by making him an a16z scout and connecting him with mentors.

Getting the snowball rolling

To quote Marc again:

How does the a16z ecosystem help? It provides a faster path to preferential attachment, within your own diaspora and by bridging to others.

Customers: We help our companies cross borders. TelevisaUnivision needed to distribute its Spanish-language library globally while preserving the actors’ original intonation and humor. We connected them with ElevenLabs. Within months, their programming could travel across languages seamlessly, supporting a subsequent agreement with Netflix. No stage is too early: we introduced Factor, a German-led company, to a large bank in Latin America that became one of their first pilot customers.

Brand: The a16z New Media team offers “go-direct as a service” for our portfolio startups, providing:

• In-house creative (Video, design, editorial)

• Owned channels (Our X account, Substack, podcast feeds, and more)

• Full service from a dedicated team, from pitch to launch and beyond

• A big network of people, with structure, software, and AI to help leverage it

All of these motions work really well for Borderless Founders and their companies: they are already differentiated in a cool way, and New Media can help amplify their stories (which are often both amazing and undertold!). Listen to the CEO of Addi on building a bank from scratch in Colombia (link), CEO of Jeeves on solving cross-border payments with Stablecoins (link), CEO of Kavak on how his car marketplace is run by agents (link & link) and the CEO of hyperexponential on building an AI platform in insurance (link).

Talent: While Borderless Founders start with strong networks at home, we accelerate their U.S. transition by providing access to trusted operators, advisors, and experts. For example, we often help founders transition from founder-led sales to scalable models. One company tripled its top-of-funnel and increased ARR by 70% in just a few months through our network.

Sometimes we introduce Borderless Founders to one another. We backed Fernando Gadotti’s second company at the pre-seed stage and had been building a relationship with Sebastian Mejia while he was still at Rappi. We connected them, and they co-founded Tako, a company reducing bureaucracy in Brazil starting with payroll.

Your background is your moat

More than half of America’s billion-dollar companies were founded or co-founded by immigrants. Because Silicon Valley is now a state of mind rather than just a place—with AI and remote work collapsing physical distance—you no longer have to choose between your home network and the American market. In fact, 44% of investments made by a16z Apps in the last two years feature international founders, split evenly between HQ in the US and abroad.

The best founders leverage their multi-country advantage to win globally. Talent, customers, brand—compounding across borders, not despite them. The founders who internalize that treat their home country as an unfair advantage to capitalize on. If you’ve been told your background is something to overcome, you’ve been given bad advice. Your background is your moat. Our job is to help you widen it. If you’re a Borderless Founder, we’d love to meet you.

Thanks to Chris Schnabl for his help on this piece and with our all Borderless Founders.

Subscribe now

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

America | Tech | Opinion | Culture | Charts

We are excited to announce the fellows who will be part of the inaugural a16z Talent Engineer Fellowship, an 8-week cohort of talent leaders for the people building the next operating system for recruiting.

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 talent at companies like Jack & Jill, Ashby, OpenAI, AirBnB, Clay, Decagon, and Ramp.

Many are talent engineers who are building 0-to-1 AI-native talent systems and now lead- or are the powerhouses behind those functions at high growth AI companies. Folks like Jared Long, Scott Bianco and Samantha Moberly from Cursor, Dan McCarthy at Zapier, Cynthia Chen at Decagon, Danielle Poreh and Ryan Ellis at Clay, Umesh Khanna and Ishan Deshpande at SpaceXAI, Johnny Sanchez at Airbnb, Saya Iwasaki at Notion, Jaime Jose at Ramp, Tommy Kilmer and Brian Hickey at OpenAI, and more.

Some are technical leaders redefining talent, building agentic headhunters, career development platforms, and talent-matching systems at scale like Jeffrey Li from Paraform, Shahriar Tajbakhsh at Metaview and Fillmore, Matthew Wilson at Jack & Jill, Sahil Seth at Talent Pluto, Edmund Cuthbert at Superposition, Seb Cuadros and Ali Alobadi at Fonzi, and more.

Others are founders taking people graphs to the next level, scaling the infrastructure behind how people and networks connect like Pieter Ouwerkerk at Happenstance, and Jan van der Vegt at Exa.

Congratulations and welcome, a16z Talent Engineer Fellows!

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

Sign-up

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.