In 2026, most AI diligence is still being done on the wrong unit of analysis. Investors track companies. The value is in the people. And the people move.
Compute budgets, revenue multiples, benchmark scores, model release cadences—these are the metrics that fill investment memos. They are all downstream of something more fundamental. And unlike GPUs and capital, that something cannot be manufactured on demand, cannot be spun up in a Gulf-state data centre, and cannot be replaced by the next generation of hardware.
It is people. A very small number of them.
The firms that learn to treat frontier talent flows as a primary signal—not an HR concern, not a diligence checkbox—will have a decisive informational edge in the decade ahead. The firms that don’t will keep being surprised.
The problem: scarcity is real, and it’s pricing into every deal
The numbers are now in the public record. Consider what happened in 2025:
- Meta paid $14.3 billion for a stake in Scale AI that industry observers described as primarily a vehicle to acquire its CEO.
- Google paid $2.4 billion to acquire key staff from Windsurf.
- Sam Altman said publicly that Meta attempted to tempt top OpenAI talent with $100 million signing bonuses.
- Cursor reportedly offered a single intern a $20 million package.
- Safe Superintelligence reached a $32 billion valuation with roughly 20 employees and no product.
- Thinking Machines Lab closed the largest seed round in history at $2 billion—on the same basis.
None of this is compensation in the conventional sense. It is price discovery for a resource that has run headlong into a hard supply ceiling.
The logic is simple. As one former Meta executive put it to CNBC: “If I’m going to spend a billion dollars to build a model, $10 million for an engineer is a relatively low investment.” When the cost of a training run is nine figures, firms will rationally bid individual humans up to eight- and nine-figure packages. The constraint is not money. The constraint is that there aren’t enough of the right people—and there cannot be.
How scarce is supply? A recent NBER study finds that high-IQ individuals are 20 times more likely to become R&D workers. Economist Alex Tabarrok estimates frontier chipmaking work requires one-in-1,000 cognitive ability—roughly 164,000 Americans, with perhaps 100,000 realistically available for the most demanding technical roles. Frontier AI research is an even thinner slice of that pool.
This is not a labour market. It’s a few thousand people worldwide whose presence or absence materially alters the outcome of a $50 billion valuation.
The core error: most diligence treats talent as a stock, not a flow
Most investor diligence treats people as a stock—headcount of PhDs, percentage of team with graduate degrees, which universities they came from. These are static snapshots of a dynamic system. They are also the easiest things in the market to fake, because LinkedIn headcounts and press releases are written for an audience.
The signal that actually predicts outcomes is flows: who is arriving, who is leaving, where they are coming from, and where they are going next.
Flows reveal four things that stocks cannot.
1. Competitive vulnerability before it shows up in product
When two senior researchers quietly leave a well-funded lab for a competitor over a single quarter, that is a leading indicator of product slippage that will not appear in any earnings release for another 12 to 18 months.
In 2025, this pattern played out in public. Thinking Machines, months after its record seed round, saw co-founders Barret Zoph and Luke Metz return to OpenAI, with The Information reporting that other employees were considering the same move. For the backers who led its $2 billion round, that news wasn’t a surprise to anyone watching individual-level flows. It was a surprise to everyone watching valuations.
2. The real boundaries of a company’s moat
The Newcomer newsletter put the central question for AI investors plainly earlier this year: is the unprecedented compensation for top AI engineers evidence that a much higher portion of enterprise value genuinely rests on a handful of people, or is it an artifact of extraordinary valuations in a bubble moment?
That distinction is the entire investment thesis. If enterprise value walks out the door with five people, the cap table is a function of retention, not revenue. Neither world can be distinguished from the other without watching the flows.
3. Where the next company is being built
Every major AI startup of the current cycle was visible as a cluster of departures before it was visible as a fundraise. Anthropic was an OpenAI exodus. Thinking Machines was an OpenAI exodus. Safe Superintelligence was an OpenAI exodus. Investors tracking individuals—not companies—saw these formations months before the pitch decks circulated.
4. The geographic arbitrage quietly reshaping the map
The Stanford HAI 2026 AI Index, published this week, makes the point concrete. Drawing on Zeki’s dataset of 658,000 frontier AI researchers and inventors, the Index shows that the United States’ net inflow of frontier talent has collapsed from a peak of 324.6 in 2022 to just 26.0 in 2025. That’s a fall of roughly 92% in three years. India is now the largest net exporter at -16.9. Switzerland and Singapore lead the world on a per-capita basis at around 110 frontier researchers per 100,000 inhabitants—far ahead of any large economy.
These aren’t abstract statistics. They are the leading edges of where the next cluster of AI companies will form, and where the competitive pressure on existing hubs will come from.
Why this is specifically an AI problem
Plenty of industries have star talent. Three things make frontier AI genuinely different:
The contribution function is extraordinarily non-linear. In most industries, losing a senior engineer means a project runs slower. In frontier AI, losing the right senior researcher can mean a training run fails, a benchmark goal slips nine months, or an entire research direction is abandoned. The person who solves a specific problem in post-training alignment may literally be one of a dozen people on Earth capable of solving it.
The IP lives in the people, not the code. Weights can be copied. Research taste, agenda-setting instinct, and the tacit knowledge of what to try next cannot. When a researcher moves, genuine know-how moves with them—in ways that NDAs and non-competes have so far failed to contain. This is why the acquihire has become the dominant M&A pattern in AI. According to PitchBook, 365 M&A deals closed in the AI space in the first half of 2025, representing over $10 billion in total deal value. When buyers pay billions for teams and sunset the products, the asset is the people.
The geography is politically contested in ways that compound talent risk. When Meta launched its Superintelligence Lab, it hired 11 researchers—all immigrants, seven from China. Nearly one-third of the world’s top AI researchers are Chinese, and most work in the United States. Yet, as the American Enterprise Institute noted drawing on The Economist, Washington is drifting toward restriction: floated revocations of Chinese student visas, proposals for a $100,000 H-1B fee. The Stanford AI Index’s finding that US net inflow has collapsed is not a mystery. It is the direct consequence of policy choices being made in real time. Every time the H-1B debate reopens or UK skilled-worker rules change, the expected value of a basket of US or UK AI holdings shifts—and most investors have no way to measure by how much.
The shift to make: from company-level diligence to individual-level intelligence
Most AI investors still screen like this:
- Who is on the cap table?
- What does the team page on the website say?
- How many PhDs from top-five schools?
- Who did the lead researcher work for before?
Evidence-led AI diligence screens like this instead:
- Who, specifically, are the 5–15 people whose departure would materially damage this thesis?
- What is the base rate of departure for researchers with their profile at their stage of career?
- Which of them have been publicly collaborating, co-authoring, or acknowledging work with researchers at competitors in the last 12 months?
- What does the second-order network around each look like—and how many of them would follow if the lead went?
- In the geographies this thesis depends on, is frontier talent flowing in or flowing out?
The first set of questions can be answered by a junior analyst with a LinkedIn premium account. The second set cannot. And the second set is the one that determines returns.
The takeaway: this is a new asset class, and nobody owns it yet
Every major asset class in financial history has gone through the same progression. First it is a qualitative art practised by specialists on instinct. Then it becomes a data discipline with standards, instruments, and a shared language. Finally, it becomes table stakes. Credit went through it in the 20th century. Equities went through it in the 1970s. Climate risk is going through it now.
Frontier AI talent sits roughly where credit analysis sat before the ratings agencies: enormous capital at stake, universally agreed to matter, and almost entirely tracked through gossip, LinkedIn, and press releases.
The allocators who build genuine data infrastructure around it will be able to answer questions that nobody else in the market can yet pose coherently:
How much of this fund’s NAV is concentrated in fewer than 50 individuals? What is the half-life of the research team at the three largest positions? Which countries are net importers of the specific sub-discipline this thesis depends on—and is that trend accelerating or reversing?
These are investor questions. They have answers. The answers just happen to require a dataset that has, until recently, been impossible to build.
That is the work Zeki exists to do. Zeki identifies and tracks the 658,000 individuals worldwide—outside China—with a proven track record of producing genuinely new AI, recognised not by self-reported titles but by the research, data, and models they have actually contributed. Zeki’s data powers the Stanford HAI AI Index’s analysis of talent sovereignty, and it is increasingly used by investors who have realised that in the most expensive asset class of the decade, the input that matters most is the one almost nobody is measuring.
The AI cycle will moderate. Valuations will rationalise. Compute costs will fall. But the structural scarcity of frontier AI talent—the one-in-a-thousand cognitive profile, the tacit knowledge, the productive collaborations that take years to form and minutes to dissolve—will persist long after the current bubble narratives resolve one way or the other.
The investors who build this muscle in the current cycle will be the ones still generating returns in the next one. The rest will still be staring at benchmark scores, wondering why their best portfolio companies suddenly stopped shipping.
Sources reviewed for this insight: Stanford HAI 2026 AI Index, CNBC, Newcomer, The Information, American Enterprise Institute, The Economist, PitchBook, IMF Finance & Development, Quarterly Journal of Economics, Brookings Institution.