My wife Lijie published a piece this week reacting to Theo Baker’s new book about Stanford. She made a point I won’t try to remake: that elite access is real, and the more important question is what people do with it. I’d encourage you to read her piece. She argues better than I could that service is the better return on the kind of access Stanford provides.

I want to add the empirical companion. Because I've been watching the press cycle around Baker's book with one question on my mind: what about the other 97%?

Baker is a real journalist. He earned a Polk Award as a freshman, broke the story that ended a Stanford president’s tenure, and has written a vivid book about what he saw during four years embedded inside the institution. The set pieces are sharp. An uncredited secret seminar taught by a Silicon Valley CEO. Freshmen courted with caviar dinners by venture capitalists. An “incubator with dorms” where talent is sniffed out at orientation. None of it is invented. The seminar exists, the dinners happen, the pattern is real.

To his credit, Baker doesn't claim this is most of Stanford. He's explicit that the freshmen flagged for unicorn potential — what he calls the Plucked — are a small subset, and part of what makes the access feel illicit is that it isn't widely distributed. The empirical question I want to add is what happens to everyone else: the rest of any given Stanford graduating class who don't get courted with caviar, who never see the secret seminar, and who are largely missing from both the press coverage and the policy imagination that runs off it.

I’ve spent fifteen years studying this. With the late William F. Miller, former Stanford Provost, I built a multi-decade dataset of Stanford alumni-founded companies. Every cohort, every venture identifiable, longitudinal data on outcomes. The aggregate numbers are familiar by now: roughly 40,000 active companies tracing back to Stanford, $2.7 trillion in annual revenue, the rough equivalent of a top-ten national economy if the alumni cohort were a country. The 40,000-company figure appears in Baker's book, where he attributes it to "a 2011 study." That study, with Miller, is the one I've been describing. The revenue and economy framings appear in roughly every press release Stanford has issued about its role in Silicon Valley over the last decade.

What the dataset also lets us see is the distribution. And the distribution doesn’t match the headline.

The average Stanford alumnus who became a founder started their first company roughly ten years after graduation. Not at orientation. Not as an undergraduate. About a decade out, typically after working somewhere else first. Most never appeared in any “secret seminar.” Most got their first investors through normal channels: a former classmate, a faculty contact, a Series A pitch against fifty other startups. Most worked for someone else first, sometimes for a decade, before founding anything.

About 8 percent of Stanford alumni founders started their first company within a year of graduation. About 28 percent started within five years. The remaining 72 percent waited longer than that. The average gap between graduation and first founding was roughly ten years.

The ‘quick founder’ archetype Baker writes about — the freshman or new graduate building a billion-dollar company — is the closest analog to his subjects. It’s also a small minority of an already-small subpopulation of Stanford alumni. Maybe two or three percent of any given graduating class. They’re the ones the press writes about. They are not most Stanford founders. They are an interesting subpopulation, not the population.

This matters because the policy stakes of getting it right are large. Every government on earth is currently trying to engineer a version of Silicon Valley. The Inflation Reduction Act, the CHIPS Act, EU innovation programs, China’s NEV subsidies, every American state’s “Silicon Valley of [X]” initiative. The mental model running through these efforts is something like Baker’s: elite institution plus ambient venture capital plus secret networks equals founders. If that’s right, the policy problem is to recreate the elite institution and the venture capital, and the founders will follow.

The data suggest something more specific.

In work with Yong Suk Lee, published in the Strategic Management Journal, we used a quasi-experimental approach to estimate what Stanford’s main entrepreneurship programs actually do. The Center for Entrepreneurial Studies at the Business School and the Stanford Technology Ventures Program at the Engineering School were introduced at different times in the mid-1990s, which let us compare cohorts who had access to each program against those who didn’t. The finding is counterintuitive in a way that almost no one expects.

These general programs do not appear to increase the rate of entrepreneurship. In some specifications, participation in the Business School program is associated with a roughly 35 percent decrease in the likelihood of starting a company. But the startups that do emerge perform better. Lower failure rates, higher revenues. The mechanism appears to be informational. Students learn enough about what entrepreneurship actually requires to figure out whether it’s a good fit for them. A meaningful share, having learned that, decide it isn’t. The remaining founders are better-matched, better-prepared, and produce better outcomes.

That’s a different story from “entrepreneurship can be taught.” It’s closer to “entrepreneurship can be evaluated,” and the institutional mechanisms that produce good evaluation are not the same ones that produce hype.

Selective, pre-venture programs look different again, and this is where the strongest causal evidence now sits. In a paper currently under review at Management Science, my co-authors Stefan Weik, Michael Fröhlich, Aaron Defort, Isabell Welpe and I study the Center for Digital Technology and Management — CDTM — a selective pre-entrepreneurship program in Munich that admits roughly 25 students per cohort from several hundred applicants. CDTM ranks applicants by composite interview scores with a sharp capacity cutoff. Candidates just above the cutoff get in; candidates just below mostly don’t. Their interview scores are nearly identical. The design lets us compare what happens to functionally equivalent people on opposite sides of an arbitrary line — the cleanest test currently available of whether selective programs cause high-quality founding or merely select the talented who would have succeeded anyway.

Three findings matter for the present debate. First, program participation more than doubles the founding rate, and the entire effect is concentrated in high-growth ventures. The probability of raising $10 million or more in venture capital rises from 0.7% in the control group to roughly 9% among participants — an order-of-magnitude shift. There is essentially zero effect on low-growth or lifestyle ventures. Second, the mechanism is not what most people guess. Program grades do not strongly predict whose ventures succeed. What predicts success is the thinness of a participant’s pre-existing network: engineering and computer science students, who arrive with fewer entrepreneurial connections, benefit substantially; business students, who arrive better connected to the relevant capital and talent pools, show essentially no treatment effect. Third, roughly 73% of participant co-founding relationships form across cohorts rather than within them, and 23% of participant founders receive early funding from program alumni acting as angel investors. The program is functioning as a multi-year matching market, not a curriculum.

Two pieces of context matter. CDTM operates in Munich — outside the Bay Area, outside the established VC ecosystem. The mechanism transfers. And the mechanism itself isn’t ambient Silicon Valley magic. It is a specific, designed institutional structure: competitive admission, cross-disciplinary cohorts, sustained multi-cohort alumni networks that act as both co-founder pools and informal capital. Stanford’s Mayfield Fellows Program shares these features. The Instagram co-founding story — Kevin Systrom and Mike Krieger from different Mayfield cohorts, connected through the program’s network — is the same cross-cohort matching pattern the CDTM data identifies more rigorously. Two independent settings, with very different identification quality, point in the same direction.

The Stanford effect, in other words, is not produced by ambient ecosystem magic. The Stanford effect, to the extent we can measure it causally, appears to be produced by specific, identifiable, replicable institutional mechanisms.

There’s also a timing problem with Baker’s account that I think gets undersold. Baker was a freshman in fall 2022. His four years at Stanford coincided exactly with the most extreme AI funding cycle in technology history. The pattern of VCs paying freshmen to drop out, courting eighteen-year-olds with model dinners, treating Stanford as a unicorn-spotting frontier intensified dramatically during the GPT-3-to-GPT-5 window. It’s real, but it’s also a peak-moment phenomenon, not a steady-state condition. Cohorts from 2008, or 1998, or 1988 had different experiences because the environment around them was different. The Stanford the press is currently scrutinizing is partly Stanford and partly the AI boom that happened to coincide with Baker’s enrollment.

This is the kind of distinction the longitudinal data and design-based evidence make legible and journalism mostly can't. A journalist describes the Stanford he saw. A researcher with forty years of cohorts and a regression discontinuity can say which features of that Stanford are durable institutional patterns and which are products of the specific moment in which the journalist observed.

None of this is a defense of the seminar, or the dinners, or the broader pattern Baker is right to find unsettling. It’s not a defense of Stanford either. It’s the longer-horizon version of the same observation. The reason most Stanford alumni who became founders don’t look like Baker’s subjects isn’t that Stanford lacks the elite-access world he describes. It’s that the elite-access world is much smaller than the headline implies, and most of the institutional work that actually produces founders happens elsewhere. In less photogenic places, on longer timescales, through mechanisms that don’t make for vivid scene-setting.

Baker has written the book about the part of Stanford that's easiest to see. With the data and the causal evidence, we can also describe the part that's harder to see but does more of the work. Both are true. Both are worth understanding.

If you’re interested in which institutional mechanisms, at Stanford and elsewhere, actually move the needle on who becomes a founder, I’ll be writing more about that here over the coming months. Lijie and I are also working on this through our Foundation, applying what we’ve learned to settings where the resources are scarce and the leverage is high. If you want to follow that work, her piece is the place to start.

Stanford is real, the access is real, and the question of what it's for is the right question to ask. The evidence suggests that most of what Stanford produces is built through more ordinary institutional machinery than the press cycle implies. And that's actually the more useful finding. Ordinary machinery is something other institutions — in Munich, in Hsinchu, in places without caviar dinners — can build.

General-program findings drawn from Eesley & Lee, "Do University Entrepreneurship Programs Promote Entrepreneurship?" Strategic Management Journal, 2021. CDTM findings from Weik, Fröhlich, Defort, Welpe & Eesley, "Pre-Entrepreneurship Programs and the Quality of Entrepreneurship," working paper currently under review.

In the months that followed, those teachers — without further intervention from us — brought what they’d learned back into their classrooms, with the new curriculum integrated into their teaching. From the six teachers who responded to the post-training survey, those six alone directly reached 611 of their students. Extrapolating across the full cohort, the curriculum likely reached on the order of two thousand five hundred students.

That’s roughly 100 students reached per teacher trained. By every reasonable measure, this was the most measurably effective program our small foundation has ever run.

It is also the program that finally rebalanced how Lijie and I think about what philanthropic education work is for.

What we already knew, and weren’t acting on.

Lijie has been around teacher training for far longer than the Foundation has existed. Before her engineering career — before Silicon Valley, before us — she ran teacher-training programs in rural China and worked as a program manager at China’s Ministry of Education. She had spent years watching what good teacher PD could do at scale, and what poor teacher PD couldn’t do at any scale.

So when we co-founded the Foundation in 2021, she’d already been making the multiplier argument for a decade. I (Chuck) was the holdout — my instinct, coming from Stanford research, was to value direct contact: be in the room, see the students, capture the texture of the work in real time. We compromised by doing both, often inside the same program: our 2023 LOHADA visit in Tanzania trained roughly twenty teachers alongside the fifty-plus students we taught directly, and our work at Fulbright University Vietnam has consistently paired teacher-development sessions with the student-facing programming. But the bias tilted toward students in the room — that’s what we showed up to do, and what we celebrated. Penang is where the math finally became the headline.

What 100-to-1 looks like.

Penang is where Malaysia builds its semiconductors. Intel, AMD, Lam Research, Bosch — they all have their fingerprints on the island. It is a place that knows what high-skilled technical labor looks like and is short of it.

The Penang Science Cluster is a non-profit founded by industry leaders to build the pipeline. Their model is the right one: convene teachers from rural schools across the state, train them in disciplines the schools don’t know how to teach, and let those teachers go back and teach. We were brought in to contribute the AI and design thinking module — a small input into a much larger program PSC was already running well.

The module was one day of intensive work with the cohort of teachers, built around a curriculum we’d developed specifically for the way Malaysian middle schools actually work. (You’d be surprised how much philanthropic curriculum dies on contact with a school where the kids don’t have laptops and the teacher only has 35 minutes a week with each class.)

After the training, the teachers brought what they’d learned back to their classrooms. PSC ran a post-training evaluation, with six teachers and eleven students responding so far.

The numbers from that subset: 100% of responding teachers adopted at least one new AI tool or design thinking technique — most are using ChatGPT, with Gemini and rapid prototyping in the mix. Every responding teacher reported saving time through AI workflow integration — most one to two hours per week, the rest three to five. Of the eleven students surveyed, 73% reported discovering a career path they had not seriously considered before — software engineering, chemical engineering, product management, AI research, public policy, even chef — and the remaining 27% said the workshop gave them new specifics on a path they already liked. Teachers rated the workshop 8.7 out of 10 on average for “would you recommend a colleague”; students rated it 4.6 out of 5; and on a separate question, students rated their confidence in using AI to research and plan their futures at 4.1 out of 5. Survey responses are still coming in; these are the early numbers.

The numbers miss what’s actually interesting. A teacher, in their own words:

“Since the workshop, I have started integrating ChatGPT into my teaching workflow by using it to help design lesson plans, generate discussion questions, and create differentiated learning materials for students. It has also been useful for quickly drafting emails and administrative documents, which saves time and allows me to focus more on student engagement.”

A student:

“I have discovered to become an AI Researcher and Public Policy Researcher. What draws me to this career path is that it combines both my interests and my strengths… Before joining this workshop, I felt unsure and unclear about my future career direction. This workshop gave me the opportunity to explore my interests and skills more deeply, and helped me see how they can be applied in ways that are beneficial not only to myself, but also to society.”

Those are the easy numbers and their human counterparts. What I want to talk about is what they imply, and — equally — what they don’t.

The thing about the math.

If you spend $X to teach 50 kids directly for a week, the marginal cost per student-hour is high, the long-tail effect is approximately zero, and the work doesn’t accrete. The fiftieth kid is not measurably better off because you taught the first forty-nine.

If you spend roughly the same $X to train twenty-five teachers — who then teach roughly a hundred kids each per cohort, year over year — the cost-per-student-reached collapses, the long-tail is the dominant term, and the work compounds.

This is not a hard observation. It is something the larger foundations figured out decades ago. (The Gates Foundation’s K-12 work, when they were doing it, was almost entirely teacher-training oriented for exactly this reason.) But it is something small foundations are slow to internalize, because direct teaching has a kind of moral romance that teacher training does not.

When I tell people the Foundation reached thousands of students in Penang last year, the natural follow-up is: “Wow, were you there?” And the honest answer is: no. We trained roughly twenty-five teachers in December. Those teachers reached their students in February, in March, in April. We never met them.

That is not a failure mode. That is the point.

The honest limits.

I want to be specific about what this analysis is not claiming, because someone with serious training in education research will land most of these critiques the moment they read the numbers.

It is not claiming the 100-to-1 multiplier is generalizable. PSC is an unusually well-resourced partner: industry-funded by Intel, AMD, Lam Research, and Bosch; staffed by people who recruit teachers from rural schools across the state; equipped with rigorous evaluation infrastructure we did not have to build. The teachers we worked with were carefully selected — motivated, experienced, ready to adopt. Most teacher-training programs run with self-recruited teachers in contexts without that wraparound. The multiplier collapses.

It is not claiming teacher PD generally works. There is a long, sobering literature on teacher professional development — most one-shot training has near-zero effect on student outcomes. The mechanism that distinguishes the rare PD programs that do work, and Penang among them, is the wraparound: rigorous teacher selection, multi-day immersion, curriculum specific enough to use Monday morning, embedded measurement, and follow-up that turns one-shot training into multi-year relationship. Teacher training without that wraparound is direct teaching in a different costume — fewer kids reached, same lack of compounding.

It is not claiming we did all the work. PSC did the teacher recruitment, the venue, the school relationships, the follow-up coaching, and the measurement that produced every number I’m citing. We provided the AI and design thinking module. The multiplier belongs as much to them as to us. Probably more.

And it is not claiming we measured what actually matters most. A “would you recommend this” rating measures satisfaction; “discovered new career paths” measures self-reported insight; “adopted at least one new technique” measures observable behavior. None are what a serious education researcher would call a learning outcome. A six-month follow-up showing students retained AI concepts they didn’t have before, or that the trained teachers’ classes outperform peers on independent assessment — that would be rigorous measurement. We don’t have that yet. We’re building toward it for the next cohort.

What we are claiming, more modestly: with the right partner and the right wraparound, in one program in one country in one year, we produced an order-of-magnitude better cost-per-student-reached than direct teaching, and the early indicators are unusually strong. Enough to shift the bias of our 2026 programming. Not enough to recommend every small foundation drop direct teaching tomorrow.

Why we still teach directly.

We have not stopped teaching directly. We will not stop. Three reasons matter.

The first is signal acquisition. You learn things in a classroom you cannot learn from a teacher’s after-action report or a survey instrument. When Lijie taught at LOHADA in Tanzania, what she came back with was not a number — it was an intuition for which parts of our entrepreneurship curriculum survived contact with East African secondary students and which parts collapsed. That intuition then went into the teacher-training material we now run elsewhere, including the teacher cohort at LOHADA itself. Direct teaching is the R&D function of a teacher-training operation.

The second is relationship. The Foundation’s most durable partnerships — Tanzania through Chuck’s former student James Juma, Vietnam through his former student Bao Phan, MIT MEET through their leadership — came from being physically present early on. The people who later open doors for us, who invite us back, who introduce us to other partners, did so because they met us in a room with students. We will keep showing up in rooms with students.

The third is that we love it. This may sound trivial in a piece otherwise devoted to math. It isn’t. There is something about teaching a class of kids in Molokai or Kampala that we don’t get from any other part of our lives. Optimizing the Foundation entirely for measurable multiplier effects would optimize out the part of the work that gives us the energy to do the rest of it. That is a bad trade.

The rebalancing is exactly that — a rebalancing, not a replacement. Most new programming dollars now bias toward teacher-anchored models. Some fraction continues to fund direct teaching, treated explicitly as research, relationship-building, and the part of the work that keeps us human.

Stepping back.

I’ve been lucky to spend much of my career teaching: Stanford undergraduates, doctoral students, corporate executives in dozens of countries, engineers at companies that don’t usually let outsiders in. The K–12 and refugee work the Foundation does is one slice of that. It’s worth saying out loud why teaching in these settings matters — to the people in the room, and to us.

For the people in the room: most of what’s worth knowing about how innovation, entrepreneurship, and AI actually play out doesn’t make it into textbooks. The frameworks that work, the failure modes you’d never anticipate, the way institutional context changes everything — these get transmitted person-to-person, classroom by classroom. Students who never get into a room with someone who’s spent twenty years thinking about how this stuff works are not less talented. They are under-resourced. Teaching is one of the few interventions that closes that gap directly.

For us: teaching is where the research becomes legible. I leave a classroom of Korean executives, or Vietnamese engineers, or a Penang teacher cohort, with sharper questions than I came in with. The doctoral student who pushes back on a framework, the executive who tells you why your model breaks in their industry, the teacher who explains which part of the curriculum students actually struggle with — these are inputs no peer-review process produces. From a research perspective, the Foundation’s K–12 teaching, our work with refugee entrepreneurs in Uganda, and the corporate teaching and the Stanford classroom are the same activity: they are the source of the next paper, the next program, the next correction to what I thought I knew.

Teaching is not what you do once you’ve stopped learning. It is one of the better ways to keep learning. The Foundation exists, in part, because we want our work to keep being shaped by what we hear from the people we teach.

Why we work in partnership.

The honest version of “we trained twenty-five teachers” is “we contributed an AI and design thinking module to a teacher-training program built and run by the Penang Science Cluster.” We didn’t invent the multiplier model. We didn’t recruit the teachers. We didn’t build the relationships with the rural schools, or the evaluation infrastructure, or the years-deep credibility with industry partners that makes PSC the convener it is. We were the new collaborator at a table other people had set.

This is not a complaint. It is the model.

A small foundation cannot — and should not try to — replicate what a serious operator like PSC has built. Their staff have been at this far longer than we have. They have institutional relationships, regional credibility, and an evaluation discipline we are still building toward. We were inspired by what they’ve put together, learned from how they run it, and contributed the specific thing we could contribute well — AI and entrepreneurship curriculum built on Stanford research, plus the Stanford brand and connections that opened doors PSC could then walk through. Partner brings the structural thing. We bring the specific input. That is the only sensible model for a foundation our size.

This is not unique to Penang. Every Foundation program is built around a partner who knows their context better than we ever will: MIT MEET in Jerusalem, LOHADA in Tanzania (anchored by James Juma, where the work pairs teacher training with direct student teaching), Fulbright University Vietnam (the same combined model), Makerere University Business School and Challenges Uganda for the refugee entrepreneurship work in Kampala, Kaunakakai Elementaryon Molokai through teacher Kawika Gonzales, ITRI in Hsinchu, the SFSU CS department before its alumni network took over the work entirely. We bring what we have. They bring what they have. The work compounds because of the partnership, not despite it.

Without the partner, the multiplier is a hypothesis. With them, it’s the thing.

What this means if you run a small foundation.

A few practical observations from the last year:

1. Teacher training without curriculum is a waste. The teachers we worked with in Penang were motivated and ready to adopt. What they needed wasn’t inspiration. It was material — a curriculum specific enough to use on Monday morning and flexible enough to fit their actual classroom constraints. Most of the philanthropic AI-literacy material out there does not survive that test. Building good curriculum is the most underrated thing a small foundation can fund.

2. Multipliers compound only with continuity. A one-shot teacher-training program with no follow-up is direct teaching in a different costume. The foundations getting real multiplier effects are the ones building multi-year relationships with the same partners — annual cohorts, refreshed materials, alumni teachers becoming mentors for the next cohort.

3. Track the kids you don’t meet, and measure what they actually learned. If your evaluation framework only counts students you taught directly, you’ll inadvertently optimize away from teacher training. But also: if your framework only counts recommendation ratings and “adopted at least one tool,” you’ll mistake satisfaction for learning. Build both behavior and outcome metrics into the partnership upfront.

4. Direct teaching is your R&D budget; account for it that way. If you cut direct teaching to zero, you’ll be running a teacher-training operation with stale material and no relationship pipeline within three years. Treat your direct-teaching work explicitly as research and relationship-building, fund it accordingly, and don’t apologize for the lower headline numbers.

5. Find your operating partners and let them lead. A small foundation that tries to be the operator will underperform a small foundation that finds excellent operators and contributes the specific input those operators don’t already have. We did not build PSC. We brought a curriculum module to a model they had spent years developing. If you are sizing up a new program and you’d be the operator, be honest about whether you should be — and if there’s a serious operator already in that geography, ask whether you’d do more good as their collaborator than as their parallel.

Where this leaves us.

The Penang program will run again this year, with more teachers, stronger curriculum, deeper follow-on, and outcome-based measurement we should have built in the first time. We are looking at similar teacher-anchored, partner-led models for our work in rural Hsinchu and at Kaunakakai Elementary on Molokai. We continue to do direct teaching — usually paired with teacher-training cohorts in the same program — at LOHADA in Tanzania, at Fulbright University Vietnam, with refugee entrepreneurs in Uganda, and with university students wherever the cohort itself is the multiplier. Both because it works for those audiences, and because it keeps the rest of the operation honest.

There is a version of this story that ends with: “and so we figured it out, and now we know how to do philanthropy at scale.” That is not the story I am telling. We figured out one thing about one program in one place, in close collaboration with a partner who’d been figuring it out for years before we showed up. We have eight programs in eight places, and the model that worked in Penang may or may not transfer to a refugee settlement in Uganda or a rural school on Molokai. We are betting that a version of it will.

Small foundations like ours often spend our first few years confused about whether we are a teaching organization, a funding organization, or an operating organization. We are not, mostly, any of those alone. We are an organization that finds excellent partners, equips them with what they need that they don’t already have, contributes what we can do well, and keeps a small, deliberate, joyful slice of the direct work for ourselves — for reasons that have less to do with measurable impact than with knowing what the work actually feels like.

Twenty-five teachers reached on the order of two thousand five hundred kids without us. They reached them through a program PSC built. With curriculum we contributed. That partnership is not despite our smallness. That partnership is what makes our smallness work.

— Chuck

I was quoted in that Bloomberg piece, and the framing I gave Annie — that the startup world is “a bit more of a Wild West,” with no audit requirements and no cop on the beat — has been the part most readers shared. I want to use this post to say what I didn’t have room to say in 200 words of quoted speech: this is not a story about one founder, and it is not, in any deep sense, a story about ARR. It is a story about what happens when an ecosystem builds an investment thesis around a metric with no agreed-upon definition, no enforcement mechanism, and no countervailing institution incentivized to police it.

That’s a story economists and organizational scholars actually have tools for. And the policy implications are not the ones most commentators have been reaching for.

Three structural reasons ARR is decoupling from real revenue

The naive ARR calculation is one month of subscription revenue × 12. It works when three conditions hold: subscription pricing is the dominant model, customer retention is high enough that next month resembles this month, and contract structure is reasonably uniform across customers. SaaS in roughly 2010–2020 met all three. AI in 2024–2026 meets none of them.

First, AI customers experiment. Enterprise budgets right now have unusually large discretionary lines for “AI exploration” — every CIO has been told by their board to have an AI strategy. That money flows into trials. A trial signed in March counts as ARR in March. The customer’s actual decision — does this tool earn its seat at renewal? — happens in September. By then, ARR has already been booked, reported to investors, and used to justify a markup at the next round. Net revenue retention numbers, if they were available, would tell a different story; they generally aren’t, because most AI startups are too young to have meaningful 12-month cohorts yet.

Second, pricing has shifted. A growing share of AI revenue is usage-based — tokens consumed, calls made, seats actively engaged. Darren Yee at NYU made the point well in the Bloomberg piece: you cannot take one month of subscription and multiply by twelve when most of the bill is usage. The lumpiness is structural, not transient. Companies layer nominal subscriptions on top of usage-based billing and report the combined number as ARR, but the usage portion behaves nothing like a recurring annuity.

Third, front-loading. A 12-month prepaid contract signed today can be reported as $X of ARR on day one, even though the customer has 11 months left to decide whether to renew. The accounting is technically defensible. The economic substance — the real-world claim about revenue stability — is materially weaker than the number suggests.

Put these three together, and the same nominal ARR figure can describe radically different underlying businesses. That’s the ambiguity Roy Lee exploited — clumsily, with a 35% lie that was easy to falsify. The more durable problem is the founders who don’t lie at all, who pick the most flattering legitimate definition each time, and whose numbers nonetheless overstate true recurring economics by 20–40%.

Why VC due diligence doesn’t close the gap

The standard answer — and the one I gave in the Bloomberg piece — is that VC and acquirer due diligence is supposed to be the cop on the beat. In principle, that is right. In practice, the incentives don’t align as cleanly as the model assumes.

Will Gornall and Ilya Strebulaev’s Squaring Venture Capital Valuations with Reality (Journal of Financial Economics, 2020) showed that unicorn valuations are overstated by an average of about 48% once preferred share terms are properly priced. The mechanism is what matters: VCs and founders both benefit from the headline number, and the LPs who would in principle care are not at the diligence table. ARR has the same structure. A VC who marks her portfolio to ARR, raises her next fund partly on those marks, and competes for allocation in the next hot round has limited incentive to demand that founders disclose cohort-level retention. The founder doesn’t want to. The other VCs in the round don’t want to. The LP — the only party with skin in the game on the truth of the number — sees the marks and not the underlying.

This is a classic institutional-design problem. A metric is informative only if some actor in the system has both the ability and the incentive to verify it. In public markets, that role is played by auditors, the SEC, short sellers, and enforcement actions. As an independent director and Remuneration Committee chair on a Hong Kong–listed public company, I see what that machinery looks like up close — quarterly review cycles, named auditor liability, regulator inquiries that come on a predictable cadence. In private markets, the equivalent infrastructure has never been built, because for most of the venture industry’s history it didn’t need to be: funds were small, LPs were sophisticated, capital was patient. None of those conditions still hold.

This is partly an American problem

It is worth pausing to note that the convention I have been describing is largely an American one. I co-direct the Stanford Technology Ventures Program (STVP) for international entrepreneurship, and through STVP’s global programs we run field research and teaching across six continents. From that vantage point, the parochial nature of “ARR as universal yardstick” is hard to miss.

European venture markets, with more conservative LP bases and a stronger founder accounting culture, tend to push cohort-level disclosure into the diligence process earlier. Singapore family offices — which have grown into a meaningful share of the global LP pool over the past decade — increasingly include net retention reporting in fund-level terms. Chinese AI startups face the opposite pressure: their domestic disclosure regime is tightening through STAR Market and HKEX scrutiny even as Western VCs grow more permissive about ARR ambiguity. Israeli founders, who typically raise from US funds, end up triangulating between conventions, and Indian founders increasingly do the same.

None of these ecosystems has solved the problem. But the “Wild West” framing applies most squarely to American venture finance in 2026, and reform may well come from outside it. The work I have done with collaborators on how institutional environments and industrial policy shape entrepreneurial outcomes — particularly comparing the US and Chinese ecosystems — keeps returning to the same lesson: convention is local, capital is global, and when those two collide the convention usually moves first. If the largest non-US LPs continue to formalize cohort retention as a reporting term, US GPs will follow.

The case against the obvious fix

The obvious response is “audit them” — extend GAAP-style requirements down into seed and Series A. I don’t think that’s right, and I told Annie so for the piece. The cost of imposing audit machinery on a 12-person company is real. It would push out exactly the kind of high-variance experimentation that produces the small number of category-defining outcomes that matter. Work I did with Bill Miller estimating the economic impact of Stanford alumni–founded companies puts the annual revenue from that single university’s graduates on a scale comparable to the GDP of a top-ten global economy, and STVP’s global programs have reached over 200,000 students with that same entrepreneurial training across six continents. Most of the value comes from a thin tail. Choking off the experimentation at the base of the funnel to police a metric problem at the top is the wrong trade.

What would actually work is lighter and more institutional in character.

Cohort retention norms. The single highest-leverage move is for the largest LPs — public pension funds, sovereigns, university endowments — to begin asking, as a condition of allocation, that their GPs disclose net revenue retention by cohort for their portfolio companies. The mechanic is straightforward: take all customers who signed up in a given month, track what that same group is paying twelve months later, and report the ratio. Best-in-class SaaS lands at 120%+; healthy is 100–115%; below 90% means a leaky bucket regardless of what the headline ARR says. Public SaaS companies routinely disclose this on earnings calls because investors demand it. Private companies don’t, because their LPs have not yet demanded it of GPs and GPs have not yet demanded it of founders. The metric is well-defined, the data already exists in every Stripe and billing system, and it cuts directly through each of the three structural problems above. No regulation required. The change in equilibrium would happen in a quarter.

Acquirer playbook updates. The corp dev teams at the strategics doing AI acquisitions should standardize on a “true ARR” calculation that strips out trials, prorates front-loaded contracts, and discounts the usage portion. Several already do. Publishing the playbook would normalize it.

Disclosure-not-audit. Chris Sloan’s line in the Bloomberg piece — always err on the side of disclosing too much rather than too little — is the right ethical norm and is also, in expectation, the right strategic norm. Founders who disclose more get the benefit of the doubt the next time something looks off. Founders who disclose only the favorable number get re-priced harshly when the market turns, which it eventually will.

Why the ethics framing is necessary but not sufficient

Founder ethics matters, and it runs as a serious thread through STVP’s programming — from the Entrepreneurial Thought Leaders (ETL) speaker series, where founders regularly walk through the hard calls they got wrong, to the Xfund Ethics Fellows Program, the student-led cohort program built specifically around developing the personal principles entrepreneurs will lean on when the pressure to overstate is greatest. The Cluely confession will almost certainly show up as a teaching case in the next iteration of MS&E 272, the global entrepreneurship course I co-teach with Vimbayi Kajese. Students need to wrestle with these moments early, before they’re sitting in the chair Roy Lee was sitting in.

But individual ethics is the wrong layer at which to expect this problem to resolve at the system level. Even fully ethical founders pick the most flattering legitimate definition each time; the question is whether the institutions around them — VCs, LPs, acquirers, journalists, faculty — reward or penalize that picking. That is an institutional question, not a character question. We can teach character all day, and we should. We will not teach our way out of a measurement convention that every party with a seat at the table is incentivized to leave ambiguous.

The right concept is earnings quality

A sharper way to put all of this — credit to Ben Hallen, who pointed this out after the first version of this essay went up — is that what private markets are missing is a concept of earnings quality for ARR.

Earnings quality is a well-established idea in financial accounting. Two companies can report identical earnings under GAAP and have those earnings mean radically different things in terms of how durable they are, how much they reflect underlying economic activity versus accounting choices, and how confidently an investor should extrapolate from them. Public-market analysts spend a lot of time asking about earnings quality. They look at accruals, at deferred revenue, at one-time items, at the relationship between reported earnings and operating cash flow. The headline number is the start of the conversation, not the end.

ARR has no such concept attached to it. Two AI startups can report the same $5 million ARR and have wildly different ARR quality. One cohort signed annual contracts after a six-month sales cycle and will retain at 95% next year. Another cohort signed three-month trials in the last quarter, with 60% likely to churn at first renewal. Same nominal number, different earnings quality, different actual business.

What is striking, as Ben pointed out in the comment that prompted this section, is that quality of earnings analysis is already common practice in another part of the deal economy: when individuals buy small businesses. The standard small-business acquisition playbook involves a “quality of earnings” review — a financial professional digs into the underlying economics, separates durable revenue from one-time effects, and tests whether the seller’s reported numbers actually describe what the buyer is buying. The buyer pays a few thousand dollars for the analysis and treats it as table stakes. That a Main Street acquirer of a $2 million HVAC business gets a more rigorous earnings-quality review than a venture investor putting $20 million into a $5 million ARR AI startup tells you something specific about the institutional design of private markets at the higher end.

The most sophisticated venture investors and acquirers do, in practice, surface ARR quality in diligence — they ask for cohort retention data, they probe the contract structure, they discount usage-based revenue. The question is why this practice has not become standard, and why the headline ARR number continues to set the terms of debate. The answer, again, is institutional. The Main Street buyer of an HVAC company has every incentive to know what they are buying because the wrong answer ruins their year. The venture investor marking a portfolio to ARR has weaker incentives — the headline number gets them the markup, the markup gets them the next fund, and the truth of the underlying earnings quality only matters if and when the position is realized, often years later.

Cohort retention is the metric that surfaces ARR quality. So is the share of revenue that is usage-based versus subscription. So is the percentage of contracts that are prepaid annually versus monthly. None of these are exotic — they are routine in public-market disclosure for SaaS companies and they are routine in small-business acquisition diligence. They are missing from venture-stage practice almost entirely. The fix is not a new metric. It is the application of an old discipline to a new asset class.

This reframing also clarifies why the lighter-touch interventions I described above are likely to work. Cohort retention disclosure is exactly the kind of additional context that allows sophisticated investors to assess earnings quality without imposing audit overhead. It is the venture-stage equivalent of asking a public company to break out recurring versus one-time revenue. The information is cheap to produce, hard to game once standardized, and dramatically improves the signal-to-noise ratio of the headline number.

A larger point about metrics and ecosystems

Step back from ARR specifically. The deeper pattern is that entrepreneurial ecosystems develop measurement conventions during a period of relative stability, those conventions get embedded in deal terms, fund marks, press coverage, and recruiting pitches, and then the underlying business changes and the convention drifts from the thing it was meant to measure. The convention persists because too many actors are now invested in it.

This is not unique to ARR. It happened with daily active users in social media, gameable through engagement-loop design. It happened with gross merchandise value in e-commerce, gameable through subsidized transactions. It happened with monthly recurring revenue in early SaaS, gameable through one-time fees disguised as subscriptions. Each cycle, the ecosystem eventually develops a sharper metric — net revenue retention, contribution margin, organic DAU — usually after a public blowup forces the issue.

ARR is in the early innings of that correction. Cluely is the public blowup. The next 18 months will show whether the ecosystem develops the disclosure norms that would let ARR remain useful, or whether the metric becomes so degraded that sophisticated investors quietly stop using it and a new one takes its place.

Either outcome is fine. The one to avoid is the middle path — everyone keeps reporting ARR, everyone privately knows it’s unreliable, and the gap between the number and reality keeps widening until the next downturn forces the reckoning all at once.

Thanks to Annie Bang at Bloomberg for the original reporting and the conversation that prompted this longer treatment, and to Marina Temkin at TechCrunch for the original Cluely reporting that started the thread. The framing I lean on here owes a great deal to Will Gornall and Ilya Strebulaev’s work on private market valuations, which remains a solid academic anchor for thinking about this class of problem.

Chuck Eesley is a Professor of Management Science & Engineering at Stanford University and co-director (for international entrepreneurship) of the Stanford Technology Ventures Program (STVP) .

Right now, entrepreneurship education teaches AI as a Generator:

❌ Generate a slide deck.

❌ Generate a business model.

❌ Generate 1,000 “bespoke” cold DMs.

The result? Total market congestion. When the marginal cost of personalized outreach drops to zero, the value of that outreach drops to zero. We’ve turned the venture ecosystem into a high-speed noise machine.

But the deeper problem isn’t spam. It’s structural.

Steve Blank’s great contribution was replacing “here’s my plan” with “get out of the building.” Lean Startup methods moved founders from storytelling to customer discovery — from assertion to evidence. That was the right shift for its era.

But hypothesis-testing frameworks have always had a foundational weak point baked in: they rely on founders to honestly convey what they found. In game theory, this is called cheap talk — assertions that are costless to make, impossible to verify, and systematically biased toward the result the speaker wants to be true. A founder does 15 customer interviews, gets ambiguous signals, and reports “strong early validation.” No fraud. Just the entirely human tendency to weight confirming evidence more heavily than disconfirming evidence.

AI doesn’t introduce that problem. It industrializes it.

The synthesis is cleaner. The narrative more coherent. The gap between what customers actually said and what the deck concludes they meant has never been easier to paper over — without any intent to deceive. Agentic AI turns motivated reasoning into a polished deliverable.

Spence’s insight from signaling theory cuts right to it: a signal is only credible if it is costly to fake. Cheap talk, by definition, fails this test. And right now, almost everything we’re teaching founders to produce — the pitch, the persona, the discovery summary, the MVP demo — has become cheap talk. Not because founders are dishonest, but because the mechanism was always under-designed, and AI has exposed the flaw at scale.

To be clear, this isn’t an argument against structured frameworks.

Bill Aulet’s Disciplined Entrepreneurship and MIT’s Orbit/JetPack tool represent exactly the right instinct — grounding AI in a rigorous, proven process rather than letting it run loose. JetPack accelerates founders through 24 steps of structured analysis in hours instead of weeks. That matters.

But there’s a warning that cuts to the heart of it: with AI, it’s never been so fast to run in the wrong direction. Acceleration is not verification. The next evolution isn’t faster generation of better outputs — it’s a different question entirely: how do we know the outputs are true?

The progression looks like this:

∙ Blank: Get out of the building (replace assertion with evidence)

∙ Aulet/JetPack: Move through the evidence-gathering faster (structured AI-accelerated generation)

∙ The next step: Make the evidence harder to manufacture (AI as verifier, not generator)

Each era inherits the previous one’s tools and exposes their blind spot. Lean Startup exposed the business plan. JetPack exposed the unstructured process. The mechanism design overhaul exposes the cheap talk embedded in both.

So what do we actually teach instead?

The answer isn’t to abandon hypothesis testing. It’s to close the loop that Lean Startup left open — the verification loop. We should be teaching founders four things:

1. Costly Signal Design.

Not every signal needs to be expensive — but the signals that matter most need to be hard to fake. This means teaching founders to design their validation process around evidence that carries real costs: a Letter of Intent that required a legal signature, a pilot that required a customer to reallocate budget, a co-development agreement that required someone to show up. These are signals that carry weight precisely because they required something from the other party, not just from the founder.

2. Separation of Synthesis from Evidence.

Founders should present raw customer data — recordings, verbatim quotes, decision logs — separately from their interpretations of it. AI can be genuinely useful here, not as a synthesizer that smooths over contradictions, but as an auditor that surfaces them: “Three of your fifteen customers said the opposite of your headline finding. Here they are.” The tool serves the verification function, not the narrative function.

3. Adversarial Simulation Before Real-World Exposure.

Before a founder runs a single customer interview, AI can stress-test their assumptions — not by generating favorable personas, but by playing the skeptic. A well-designed simulation steelmans every reason a customer wouldn’t buy, a competitor would win, or the unit economics wouldn’t hold. The founder who has survived 50 adversarial AI interviews arrives at their first real customer conversation with sharper hypotheses and a much higher signal-to-noise ratio in what they’re listening for. The output isn’t a polished narrative. It’s a set of refined, falsifiable bets.

4. Mechanism Design Thinking.

The most underrated skill we can teach founders isn’t prompting — it’s system design. Who has an incentive to tell you the truth, and under what conditions? What would a customer have to give up to signal genuine intent versus polite interest? How do you structure an interaction so that a “yes” means something? These are mechanism design questions, and they belong in every entrepreneurship curriculum alongside customer discovery and financial modeling.

Mitchell observed that technical forces eventually outpace social ones. The technical force of 2026 is Agentic AI. The social challenge is Trust. And trust, at its core, is a mechanism design problem — not a content generation problem.

We don’t need more founders who can generate a compelling narrative. We need founders who can build systems that make the truth easier to tell than to obscure.

The future of entrepreneurship isn’t about being the loudest. It’s about being the most verifiable.

#AI #Entrepreneurship #MechanismDesign #SignalingTheory #LeanStartup #DisciplinedEntrepreneurship #VentureCapital

Read more