3.4.3 Resource Competition (Data, Compute, Talent)

In the summer of 2025, Sam Altman went on his brother's podcast and said something that would have sounded unhinged in almost any other industry. Meta, he claimed, had been dangling signing bonuses of $100 million to lure away OpenAI's researchers — and, he added with evident satisfaction, "none of our best people have decided to take them up on that." Mark Zuckerberg had spent the preceding weeks personally texting a list of a few dozen names, hosting them at his homes in Palo Alto and Lake Tahoe, and assembling what Meta branded its "Superintelligence Labs." The reported packages that leaked out over the following months were the kind of numbers usually reserved for acquiring whole companies: a rumored $200-million-plus offer to Ruoming Pang, the executive who had run Apple's foundation-models team, and multi-year deals for individual researchers reportedly stretching past $300 million.

Strip away the specific figures — some of which the companies disputed — and the striking thing is the underlying logic. Meta was not buying a factory, a patent portfolio, or a customer base. It was buying people. A handful of them. And it was willing to pay per-person sums that exceed the lifetime earnings of a thousand ordinary engineers because, in frontier AI, a very small number of individuals appear to make an outsized difference to whether your models work.

That episode is the sharpest illustration of a broader condition. Building AI at the frontier requires three scarce inputs — talent, compute, and data — and all three are finite, concentrated, and fiercely contested. The competition for them is reshaping who gets to build powerful AI, who is priced out, and what kinds of AI get built at all. Some of this competition is genuinely zero-sum, in the precise sense that a resource captured by one player is denied to every other. Some of it is not. Understanding exactly where that line falls turns out to matter enormously — for academic science, for the Global South, and for whether the field stays open enough to be governed at all.

How small is the pool, really?

Start with the primary puzzle: the mismatch between the number of people who can do world-class AI research and the number of organizations bidding for them.

The demand side is easy to describe because it is essentially everyone. Every hyperscaler, every well-funded lab, every enterprise software company adding an "AI" layer, every government building sovereign capability, and thousands of venture-backed startups all want the same profile of person. The supply side is where the scarcity bites. Universities graduate hundreds of thousands of computer-science students a year, and a growing share now study machine learning — but the number who can meaningfully advance a frontier training run is far smaller. Zuckerberg's recruitment list reportedly ran to roughly fifty names. Industry insiders often put the population capable of leading frontier research in the low thousands, and the truly elite tier — the people whose fingerprints are on the architectures everyone else builds on — in the hundreds.

This is why the compensation numbers detach from anything resembling a normal labor market. It helps to see the tiers side by side.

The top row is a genuine market. The bottom row is closer to a bidding war over a nearly fixed asset, and the two behave differently. When compensation for an ordinary skill rises, more people acquire the skill and supply eventually responds. But you cannot manufacture a decade of frontier research experience in a hiring cycle, which is precisely why the money at the top keeps climbing — the supply curve there is close to vertical.

The clearest register distinction to hold onto: that a supply-demand imbalance exists is well established; the specific dollar figures for individual superstar deals are reported, sometimes leaked, and occasionally disputed by the companies involved. The order of magnitude is not in doubt. The exact decimal places are.

The self-reinforcing drain on academia

Follow the money's second-order effects and you arrive at the most consequential feature of the talent market: it hollows out the institution that produces the talent in the first place.

A university cannot pay a professor $2 million, let alone $200 million. So the professors leave — and they take their graduate students, their research agendas, and their ability to train the next cohort with them. Stanford's AI Index has tracked the result for years: the overwhelming majority of notable machine-learning models now come from industry, not academia — a share that has climbed to roughly nine in ten, reversing a world in which universities once led. This is not a one-time transfer. It compounds.

graph LR
  A[Industry outbids universities] --> B[Top faculty leave for labs]
  B --> C[Weaker graduate programs]
  C --> D[Fewer researchers trained]
  D --> E[Scarcity worsens, prices rise]
  E --> A
  B --> F[Frontier work happens behind NDAs]
  F --> G[Less published, less teachable]
  G --> C

Each turn of that loop tightens the next. Fewer senior faculty means fewer trained students; a smaller trained cohort means the scarcity that drove the salaries in the first place gets worse, not better; and because the departing researchers now work behind non-disclosure agreements on unpublished systems, the knowledge they generate is less available to teach the students who remain. This is why the drain is so hard to reverse: it damages the very machinery that would produce a replacement supply. You cannot simply fund your way out of it in a single budget cycle, because the bottleneck is people who take a decade to form and who are actively being removed from the formation pipeline.

The compute crunch, and the divide it draws

If talent is the scarcest input, compute is the most expensive one — and the clearest illustration of how quickly the goalposts move.

The trajectory of training costs is genuinely staggering. Stanford's AI Index estimates that the 2017 Transformer — the architecture underneath essentially every large language model — cost around $900 to train. By 2020, GPT-3 cost a few million dollars. GPT-4, released in 2023, is estimated at roughly $78 million in compute; Google's Gemini Ultra at around $190 million. Anthropic's Dario Amodei has said publicly that models costing on the order of $1 billion are being trained now, with $10-billion training runs plausible within a couple of years. In under a decade, the cost of a frontier run has risen by roughly six orders of magnitude, because training compute has been doubling roughly every five to six months — far faster than any hardware cost curve is falling.

The physical bottleneck is graphics processing units, overwhelmingly Nvidia's — the H100 and its successors, each priced in the tens of thousands of dollars and, for stretches of 2023–24, essentially rationed. What matters competitively is not owning a few but owning them by the hundred thousand. Meta disclosed a target of around 350,000 H100-equivalents; xAI stood up a cluster it called Colossus with 100,000 H100s and announced plans to expand it by an order of magnitude. Access at that scale is secured two ways: by being a hyperscaler that builds the data centers (Amazon, Microsoft, Google), or by signing multi-billion-dollar, multi-year supply agreements, as OpenAI and Anthropic have. Everyone else waits, rents at a markup, or does without.

This is the compute divide, and it is worth being precise about what it forecloses. Here is the tension at the heart of the temporal question — whether the divide is closing or widening — because two different cost curves are moving in opposite directions.

For a fixed capability, compute gets dramatically cheaper every year. The AI Index found that the cost of querying a model at the level of GPT-3.5 fell more than 280-fold in about eighteen months — from around $20 per million tokens in late 2022 to roughly $0.07 by late 2024. On that measure, yesterday's frontier is racing toward being nearly free. A startup in 2026 can build on a model as capable as the best system of 2023 for a rounding error.

But the frontier itself is receding faster than costs are falling. If a state-of-the-art run cost $78 million in 2023 and is heading toward $1 billion and beyond, then the absolute gap between what a hyperscaler can attempt and what a university lab can attempt is widening even as the per-unit price of any given capability collapses. Both things are true at once, and conflating them is the most common error in this debate. The divide at the frontier is growing. The floor beneath everyone else is rising. Which of those facts matters more depends entirely on what you are trying to do — and that is exactly the question that separates the winners from the priced-out.

The data game, and its approaching wall

The third input is data, and here the competition has a hard temporal edge: the internet is a finite resource, and the frontier is running out of it.

Researchers at Epoch AI put numbers on this in a widely cited 2024 analysis. The usable stock of public human-generated text is on the order of a few hundred trillion tokens, and at current growth rates in model appetite, they projected that leading labs would effectively exhaust the high-quality public supply somewhere between 2026 and 2032, with a central estimate around 2028. That is not a distant abstraction. It means the era of "just scrape more of the web" is ending inside this book's shelf life.

What replaces it is a scramble for whatever data remains behind a wall — which turns data from a free commons into a licensed asset. Reddit signed a data-licensing deal reported at around $60 million a year with Google. OpenAI struck agreements worth a reported $250 million over five years with News Corp, and similar deals with the Associated Press, Axel Springer, and others. Stack Overflow, Shutterstock, and academic publishers have all monetized their archives. The pattern is a classic enclosure: a resource that was effectively public is fenced, priced, and made available preferentially to whoever can pay — which reinforces the incumbents, because the company with more users generates more proprietary data, which trains better models, which attracts more users still.

Two forces cut against that lock-in, though, and they are why the data advantage is more contestable than the compute one. First, the relationship between raw data volume and model quality is looser than the early "more is better" scaling story implied. Curation, filtering, and training technique matter enormously; Anthropic has repeatedly fielded models competitive with far larger data holders, which suggests that how you use data can partly substitute for how much you have. Second, synthetic data — models generating training material for other models — is rapidly maturing, and where it works it uncouples capability from the finite human archive altogether. Whether these forces ultimately erode the proprietary-data moat or merely reshape it is genuinely unknown. The honest position is that the wall is coming, and the field does not yet know whether it climbs over, tunnels under, or piles up against it.

Why distribution is becoming the decisive front

Here the chapter turns to a claim that deserves careful handling, because it is doing a lot of work in the current strategic conversation: as model quality converges, distribution becomes the deciding factor.

The convergence part is well evidenced. On public leaderboards, the gap between the best model and the fifth- or tenth-best has narrowed sharply — a spread of a dozen-plus points on the Chatbot Arena scale a year or two ago has compressed to a handful. GPT-class, Claude-class, and Gemini-class systems now trade the top spot month to month, and open-weight models from Meta, Mistral, and Chinese labs like DeepSeek follow close behind. When several products are all "excellent," excellence stops being a differentiator. The buyer chooses on price, convenience, and — above all — on what is already in front of them.

That is where distribution wins. Google can place Gemini inside Search, Android, and Chrome, reaching billions of people who never chose an AI product at all. Microsoft threads Copilot through Office, Windows, and Azure. Meta drops its assistant into WhatsApp, Instagram, and Facebook, a combined audience in the billions. Apple bakes intelligence into the device in your pocket. The specific mechanism by which this beats a superior model is worth naming precisely: incumbents own the default. A marginally better model that a user must seek out, install, and switch to loses to an adequate model that is simply already there when they open an app they already use. Defaults compound with behavioral data and switching costs — the more you use the built-in assistant, the more it learns your context, and the more painful it becomes to leave. Distribution converts a one-time quality edge, which erodes in months, into an installed habit, which does not.

The counter-example proves the rule. ChatGPT's late-2022 breakout — from launch to a hundred million users in two months, and to hundreds of millions of weekly users by 2025 — is the rare case of a challenger building distribution from nothing through sheer virality. But notice what made it stick: Microsoft's capital and enterprise channels arrived immediately behind the viral moment. Without that, OpenAI might have remained a research lab admired by insiders. Viral distribution is real, but it is lightning; incumbent distribution is a power grid.

The epistemic caveat matters, though. The whole "distribution is destiny" thesis rests on the assumption that capability gaps stay closed. If a lab achieved a genuine step-change — a model qualitatively better at reasoning or reliability, not marginally better on benchmarks — quality could snap back to being decisive, and distribution advantages would suddenly look fragile. The convergence is a present-tense observation, not a law of nature. It holds until it doesn't.

Where the competition is zero-sum, and where it isn't

All of this can be sorted by a single distinction that is easy to state and easy to forget: some AI resources are rivalrous, and some are not.

A rivalrous resource is one where my having it prevents your having it. The frontier researcher who signs with Meta cannot simultaneously work at OpenAI. The H100 allocated to a hyperscaler's cluster is not available to a university. The exclusive license to a proprietary dataset forecloses that source to everyone else. People, hardware, and locked-up data are genuinely zero-sum — captured by one player, denied to the rest.

But the intellectual core of AI is not like this at all. A published algorithm, a released architecture, a training technique described in a paper — these are non-rivalrous. One lab's use of the Transformer does not diminish anyone else's; a breakthrough published on arXiv is available to every competitor by morning. Knowledge, unlike a GPU, does not get used up.

This table is the analytical spine of the whole chapter, because the interaction between the two columns is where the real story lives. The field has run, for a decade, on the non-rivalrous column: open publication, shared benchmarks, code released with papers. That commons is precisely what made the rapid progress possible — everyone building on everyone. The danger is that intensifying competition over the rivalrous column erodes the non-rivalrous one. When knowledge sharing might strengthen a rival who is fighting you for the same scarce engineers and chips, the rational move is to stop sharing. And that is exactly what has happened: the labs that once published their most important work now describe frontier systems in "technical reports" that omit the architecture, the data, and the training details. Competition over the rivalrous inputs is quietly enclosing the non-rivalrous commons that fed the whole enterprise.

What it forecloses — and for whom

Zoom out from the firms to the people and places excluded, and the consequences of all this become concrete.

For the academic researcher, the compute divide is not an inconvenience but a redrawing of the map of what science can even ask. A university lab can fine-tune existing open models, probe them, study their failures, and do first-rate work on efficiency, safety, and evaluation. What it cannot do is train a frontier model from scratch and discover what only emerges at that scale — that experiment is simply off the menu, regardless of how brilliant the researchers are. Governments have noticed. The United States launched a National AI Research Resource pilot to give academics subsidized compute; the UK and EU have stood up public supercomputers like Isambard-AI and Dawn for the same reason. These are real, and they are also small relative to a single hyperscaler's cluster — gestures in the right direction that do not yet close the gap.

For the Global South, the exclusions stack. Compute is concentrated overwhelmingly in the United States and China; most of Africa and much of Latin America and South Asia have negligible frontier-scale capacity. Data is worse: the proprietary datasets and even the public web skew heavily toward English and a few other high-resource languages, so models underperform in the languages billions of people actually speak, and the local data that would fix this is neither collected nor owned locally. And talent flows the wrong way — the most capable researchers trained in Lagos, Nairobi, Bangalore, or São Paulo are recruited to labs in California and London, a brain drain now supercharged by the same salary numbers that empty out Western universities. The question of whether this drain is a market outcome to accept or a policy problem to address is a genuine values dispute, but the mechanism is not in doubt: the same forces that concentrate AI power inside a few firms also concentrate it inside a few countries.

And for the field as a whole, talent concentration bends the research agenda. When the scarce top researchers are absorbed into a handful of labs racing for the frontier, the work that gets the best minds is the work those labs prioritize — capability, scale, product. The areas that a well-functioning research ecosystem would fund in parallel — interpretability, safety, fairness, robustness, and applications for underserved communities that promise no commercial return — are systematically under-resourced not because anyone decided they don't matter, but because the people who could advance them fastest are contractually busy elsewhere. Concentration doesn't just decide who builds AI. It decides which questions get asked.

What is owed, and what is genuinely unknown

Two questions remain that the evidence cannot fully settle, and honesty requires naming them as such.

The first is normative. Leading labs built their capability standing on a global commons — decades of publicly funded mathematics, openly published research, freely shared code, and the scraped output of billions of people who never consented to train a model. Having drawn so heavily on that commons, do they owe it continued openness in return? There is a real argument that competitive secrecy, past some point, becomes a kind of enclosure of a shared inheritance — taking from the commons while ceasing to give back. There is an equally real counter-argument that some capabilities are genuinely dangerous to publish, and that restraint is responsible rather than selfish. The book does not pretend these resolve cleanly. But the trend is clear and worth watching with concern: the direction of travel is toward less openness, and each lab's individually rational secrecy may add up to a collectively worse outcome for a field that progressed fastest when it shared.

The second is empirical, and it is the deepest uncertainty in this chapter. Is the current intensity of resource competition speeding the field up or slowing it down? The optimistic reading is that fierce, well-funded rivalry has produced the fastest technological progress in living memory — that competition works. The pessimistic reading is that secrecy and concentration are net negatives for aggregate progress: duplicated effort behind closed doors, a hollowed-out academic pipeline, and safety-critical work left underfunded, all masked in the short run by the sheer volume of capital being thrown at the problem. We do not know which reading is correct, and the honest answer is that it may be positive-sum on the timescale of the next model release and negative-sum on the timescale of the next decade. That tension — collaborative in knowledge, cutthroat in resources — is not a paradox to be resolved but a condition to be managed, and managing it well may be one of the defining challenges of the years ahead.

Summary

1. The talent gap is real and the market is bifurcated. A supply-demand imbalance in world-class AI researchers is well established: the frontier-capable pool numbers in the low thousands (the elite tier in the hundreds) against tens of thousands of competing organizations. Ordinary AI roles pay a normal-market premium (~$160,000 median, plus 25–45% for scarce skills); senior lab researchers earn $500,000–$1.5M; and a few dozen superstars have drawn reported — sometimes disputed — offers of $100M–$300M. The order of magnitude is certain; the exact figures for individuals are not.

2. The academic drain is self-reinforcing and hard to reverse. Roughly nine in ten notable models now come from industry. Because departing researchers stop training the next generation and stop publishing, the loop tightens over time — you cannot buy back a decade of research formation in one budget cycle.

3. Two compute curves move in opposite directions. For a fixed capability, cost collapses (querying a GPT-3.5-level model fell ~280× in eighteen months). But the frontier recedes faster: training costs have risen from ~$900 for the 2017 Transformer to an estimated ~$78M for GPT-4 and toward $1B and beyond. The floor rises for everyone; the ceiling rises faster for the few.

4. Data is hitting a wall. Epoch AI projects the high-quality public text supply is effectively exhausted between 2026 and 2032 (central estimate ~2028), driving an enclosure of licensed data (Reddit, News Corp, and others). Better curation and synthetic data may partly erode the proprietary-data moat — but whether they do is genuinely unknown.

5. Distribution is becoming decisive because quality is converging. With top models clustered close together, incumbents win by owning the default — Gemini in Search, Copilot in Office, Meta AI in WhatsApp. This holds only as long as capability gaps stay closed; a real step-change would make quality decisive again.

6. The rivalrous/non-rivalrous split is the key. People, compute, and proprietary data are zero-sum; published research and algorithms are not. The danger is that competition over the former is enclosing the latter — the open commons that made the field's progress possible.

7. The excluded face structural, not incidental, limits. Academics can study and fine-tune models but cannot run frontier experiments; the Global South faces compounding compute, data, and talent-drain disadvantages; and the concentration of top talent leaves safety, interpretability, fairness, and underserved applications systematically under-resourced. Whether the whole system is net positive-sum or negative-sum for aggregate progress remains an open, and important, question.

Sources

• Stanford HAI — 2025 AI Index Report
• Epoch AI — Will We Run Out of Data? Limits of LLM Scaling Based on Human-Generated Data

Note: I was unable to fetch live pages in this session, so the two sources above are the landmark analyses this chapter draws on from prior knowledge (training-cost estimates, inference-cost decline, industry-vs-academia model share, and the data-exhaustion projection). The reported 2025 talent-war figures (Meta's Superintelligence Labs offers, the Reddit/News Corp licensing deals, cluster sizes) should be re-verified against primary reporting before publication, as several were disputed by the companies involved.

Last updated: 2026-08-19