AI Labs Took $225 Trillion of Human Knowledge and Paid $1.5 Billion for It

Five approaches to valuing what went into the machines, and what came back out

The transformer architecture is free. It was published in 2017 in a paper you can read in twenty minutes, and every meaningful improvement since has been described in the open literature. The compute is a commodity; you can rent it by the hour from three vendors. The engineering is hard but it is not scarce, and the people who do it move between labs every eighteen months, carrying the methods with them.

Take a frontier lab, keep the architecture, keep the GPUs, keep the researchers, and remove the training corpus. What you have is a random-initialized weights file that produces noise. Not bad output. Noise.

Run it on a corpus of raw sensory data with no accumulated human knowledge in it and you get, at absolute best, something with the general reasoning capacity of a clever animal. Everything a language model appears to know every proof, every diagnosis, every legal doctrine, every line of working code, every fact about how the world is put together — was put there by people who wrote it down first.

So when we ask what the AI industry is built on, the honest answer is not the algorithms. It’s the corpus. The corpus is the product. Everything else is packaging and distribution.

That reframing has a consequence people mostly avoid, because it involves arithmetic. If the corpus is the product, then the corpus has a value. And we can estimate it.

What follows is five independent attempts. They use different methods and arrive at different numbers, which is the point a valuation you can only reach one way is a valuation you made up. Every assumption is on the page so you can attack it. At the end I’ll show what has actually flowed back to the people who produced the input, and the gap is not a rounding error. It is roughly three orders of magnitude.

First, the scale of what was ingested

Public figures give us a floor. Llama 3.1 was trained on roughly 15 trillion tokens. FineWeb, the open English corpus released in 2024, is 15 trillion tokens on its own. Qwen 3 was trained on 36 trillion. Kimi K2 used 15.5 trillion in pre-training. GPT-5 and the Claude Opus 4 series are reported to use comparable or larger corpora.

Call it 15 trillion tokens for a single frontier pre-training run. At the standard conversion of roughly three tokens per four words, that is about 11.25 trillion words.

Some context for that number. The Library of Congress holds around 25 million catalogued books. At 90,000 words each, that is 2.25 trillion words — about a fifth of one training run. The entire English Wikipedia, all seven million articles and twenty-four years of volunteer labor, is roughly 4.5 billion words. It is 0.04% of the corpus. Four hundredths of one percent.

Whatever intuition you have about the scale of this, it is probably too small.

Model 1: Replacement cost

The cleanest question in valuation. What would it cost to commission this corpus from scratch, today, at market rates?

11.25 trillion words. Freelance writing rates run from about $0.02 a word at the content-mill floor to $1.00 and up for expert technical or professional work. Take the absolute bottom of the market:

11.25 trillion words × $0.02 = $225 trillion

Global GDP is roughly $110 trillion a year. So reproducing one frontier training corpus, commissioned at the worst rates anyone accepts, would consume the entire economic output of the planet for two years.

At $0.10 a word still low for anything requiring expertise the figure is $1.1 quadrillion, roughly ten years of world GDP.

These numbers are not meant to be paid. They are meant to establish something else: the corpus is not purchasable at market rates by anyone, at any price, under any circumstances. No cap table can buy it. That is not a side observation about the economics. It is the explanation for the entire strategy. When an input is essential and unpurchasable, the only paths available are to license fragments of it, to do without it, or to take it and litigate afterwards.

One of those three is what happened.

Model 2: Embodied labor

Replacement cost is a market fiction. Labor is not. How many human hours are compressed in that weights file?

Sustained composition of publishable prose, including the reading and thinking that make it publishable, runs at perhaps 250 words an hour for a professional. That is a generous rate; most working writers would call it optimistic.

11.25 trillion words ÷ 250 = 45 billion hours 45 billion hours ÷ 2,000 hours per work-year = 22.5 million person-years 22.5 million ÷ 45 years of working life = 500,000 working lifetimes

Half a million human beings, each working from graduation to retirement, writing continuously, producing nothing else with their lives. That is the labor content of one pre-training run.

Priced at the US median wage of roughly $30 an hour, those 45 billion hours come to $1.35 trillion for a single corpus.

And that figure is a floor so low it’s almost dishonest, because it prices only the typing. It assumes the knowledge already existed and someone merely transcribed it. It assigns zero to the experiment, the fieldwork, the failed hypothesis, the decade of clinical practice behind the case report.

Model 3: Civilizational capital expenditure

The corpus is not writing. It is the output of the machinery that produced the knowledge the writing records.

Global R&D spending currently runs around $2.5 trillion a year across all nations and firms. Global spending on education the mechanism by which any of it transmits to the next generation is on the order of $5 trillion annually. These are approximations from international statistical bodies and the exact figures move, but the magnitude is not in dispute.

Take R&D alone. Take only the last fifty years, and assume the average annual spend across that period was $1 trillion, well below today’s rate:

50 years × $1 trillion = $50 trillion of research investment

That buys you the modern scientific corpus. It buys none of the preceding four thousand nine hundred and fifty years: Euclid, the Islamic Golden Age, double-entry bookkeeping, common law, every grammar and pharmacopeia and navigational table compiled before someone thought to track national research budgets.

The written record is roughly five thousand years old. The corpus is the terminal output of the longest-running and most expensive project our species has undertaken, and it arrived at the labs with no invoice attached, because the people who funded it every taxpayer, every patron, every institution, every parent who paid for a education — never structured it as a saleable asset. It was structured as a commons.

A commons is not the same thing as free. It means the cost was borne collectively and the benefit was meant to return collectively. Enclosure is the word for what happens when someone fences a commons and sells tickets.

Model 4: Their own revealed price

Here is the model that requires no assumptions from me at all, because the industry set the number itself.

In Bartz v. Anthropic, approved in 2026, the settlement paid approximately $3,000 per work across roughly 500,000 books, totaling $1.5 billion — the largest copyright recovery in American history and the first judicially endorsed per-unit price for AI ingestion.

Fine. Take their number and apply it to the corpus.

Component Approximate count At $3,000/work Books ever published ~150 million $450 billion Academic papers ~200 million $600 billion Patents (grants + applications) ~21 million $63 billion Long-form subtotal ~371 million ~$1.11 trillion

That excludes the web entirely billions of pages, the forum answers, the documentation, the recipes, the reviews, the blogs. It excludes code repositories. It excludes every transcript.

Against that $1.11 trillion for the enumerable long-form corpus alone, the total returned to creators is $1.5 billion.

That is 0.135%. Roughly one work in every seven hundred and forty was paid for. And that one was paid for because somebody sued.

It’s worth being precise about what the settlement even covered. Judge Alsup held that training on lawfully acquired books was fair use, and refused to extend that holding to books pulled from shadow libraries. The $1.5 billion did not price the training. It priced the piracy the acquisition of a central library of, in the plaintiffs’ account, some seven million pirated books. The value of using the corpus was assessed, in that ruling, at zero.

Model 5: Value capture

The last approach ignores inputs entirely and looks at where the money ended up.

As of mid-2026: Anthropic raised $65 billion at a $965 billion valuation in May, on a $47 billion revenue run-rate, having grown from about $1 billion in ARR in roughly fifteen months. OpenAI closed $122 billion at $852 billion in March. The merged xAI-SpaceX entity is valued at $1.25 trillion. Alphabet, whose search product now returns a generated answer as its primary output, sits near $4.2 trillion.

Excluding Alphabet, the three frontier private entities alone carry about $3.07 trillion in enterprise value.

Against $1.5 billion returned to creators:

$1.5B ÷ $3.07T = 0.049%

For every $1,000 of enterprise value created on top of the human knowledge corpus, approximately 49 cents has flowed back to the people who produced it.

Not 49 dollars. Forty-nine cents.

The part that makes it a transfer rather than a windfall

A windfall makes someone rich without making anyone poorer. This is not that, and there is a measurable mechanism.

The corpus needs replenishment. New knowledge has to keep entering it, and the mechanism that funded that entry for the last twenty-five years was attention: search sends readers, readers generate subscriptions and advertising, revenue pays writers and researchers who produce the next tranche. Crude, but it functioned.

That mechanism is being severed, and we have numbers.

And the replacement channel does not exist. AI chatbot referrals grew more than 200% in the same period and still account for less than 1% of total publisher referrals. The Reuters Institute, surveying 280 media leaders across 51 countries, found publishers expect a further 43% decline over the next three years.

So the flow of value runs one way: the corpus is ingested, the answers are served without attribution or click-through, the funding mechanism for producing the next corpus collapses, and the enterprise value accrues to the party holding the weights.

The labs are aware of the endpoint. Forecasts from Epoch AI put the effective exhaustion of high-quality public web text between 2026 and 2028, and the industry has already moved to licensing deals and paid expert-generated data to cover the gap.

Sit with that last fact, because it is the whole argument in miniature. They are now paying market rates for data. The moment the free corpus ran short, a price appeared, contracts were signed, and budgets were allocated. They know exactly what the input is worth. They simply declined to pay for the fifteen trillion tokens that were already lying around.

Where this argument is weakest

Three honest problems, because a case that hides them deserves to lose.

Most of the corpus has no claimant. Euclid has no estate. The bulk of accumulated human knowledge is public domain, orphaned, or deliberately released under open licenses by people who wanted it used. Any scheme built on per-creator royalties will fail on this immediately, and critics will say so.

The labs did contribute. Compute capex in the tens of billions, genuine engineering, reinforcement learning from human feedback, safety work, distribution, and risk capital deployed years before revenue. Nobody serious claims the corpus alone produces a model. The claim is about the split, not about whether they added anything.

These models measure different things and shouldn’t be summed. Replacement cost, embodied labor, civilizational capex, revealed price and captured value are five lenses on one asset. They range from $1.35 trillion to figures exceeding global GDP. That spread is a feature it tells you the true number is enormous and unknowable but anyone quoting a single figure from this piece as the value of the corpus is misusing it.

Notice, though, that the first objection points somewhere sharper rather than weaker. If most of the corpus belongs to everyone and no one, then individual royalties were never the right instrument. The correct claimant on the surplus is the public the same public that funded the research, staffed the universities, and paid the taxes that built the commons in the first place.

What follows

Four asks, none of which require anyone to stop building models.

A corpus levy. A percentage of AI revenue into a public fund, on the principle that the primary input is public heritage. At 2% of a combined $75 billion run-rate that is $1.5 billion annually the entire Bartz settlement, every year, without litigation.

Mandatory training-data provenance. The industry deployed model-level output watermarking worldwide within weeks of a regulator asking. Apply the same standard upstream. If you can mark every token that leaves, you can log what came in.

Fund the flow, not just the stock. Payment for ongoing production, not one-time settlement of past ingestion. The corpus is not a mine. It is a crop, and something is currently eating the seed.

A statutory license. A set rate, applied automatically, so that access to five thousand years of human knowledge is not decided one lawsuit at a time by whoever can afford the lawyers.

Half a million working lifetimes. Fifty trillion in research spending across five decades and an unpriced four thousand nine hundred and fifty years before that. One work in seven hundred and forty paid for. Forty-nine cents back on every thousand dollars of value created.

Those are the numbers. Argue with the assumptions they’re all on the page, which is more than the training manifests offer. But argue with the assumptions, not the conclusion, because the conclusion survives an enormous amount of adjustment.

You could be wrong by a factor of a hundred in the industry’s favor and this would still be the largest uncompensated transfer of accumulated human value in the historical record.

About the Author

Brian Curry is a Kansas City–based AI and Machine Learning researcher, data scientist, and the founder of Vector1 Research, where he works on the intersection of cognitive AI architecture, causal inference, and economic systems. His research introduced Memory-Node Encapsulation (MNE), an original data structure for artificial episodic memory, and the NeoCortex-M architecture for memory-driven AI. He is the creator of Papilon, an open-source framework for complex system optimization with agentic orchestration, PyCausalSim, a Python framework for causal discovery through simulation, and Daedalus, a platform for causal and economic analysis.

His practitioner work focuses on building production intelligence systems — causal marketing analytics, agentic systems, and the serving infrastructure that makes them deployable inside enterprise environments. He writes about that work at vector1.ai and on Medium.