Data Is Worth Half of AI. It’s Paid Two Percent.
How Big Is Data’s Share of the AI Economy? Three Independent Roads to an Answer
Returns to Data — Article II of a Vector1 Research Program
By Brian Curry · Vector1 Research
In Article I we argued that data is the fourth factor of production and the only one paid almost nothing. That raises the obvious question — how much should it be paid? This paper estimates the size of data’s share, γ, three independent ways: from the physics of how models learn, from the market that is forming to buy data, and from the courtroom that just priced it. The three roads do not fully agree — and the distance between them is the most important number in this series.
1. The Number the AI Field Already Published
Here is the problem with measuring data’s share of the AI economy: you cannot look it up. Capital’s share and labor’s share appear in the national accounts because capital and labor are paid, and payments leave records. Data is not paid, so it leaves no trace in the accounts. Its share, γ, is a missing number.
But it is not unknowable. It is latent in evidence we already have — we just have to invert the right relationship to recover it. And the most powerful piece of evidence was published not by economists but by machine-learning researchers, who had no idea they were measuring a factor share.
When you train a large model, its performance follows a startlingly regular law: error falls as a power of how much compute you spend on parameters and how much data you feed it. Fit that law and you get two exponents — one on parameters (call it α, the compute term) and one on data (call it β). These exponents are the empirical heart of modern AI. And when the canonical fit is done carefully, they come out almost equal.
Figure 1. The two exponents of the Chinchilla scaling law, from the Besiroglu et al. (2024) reanalysis: the compute/parameter exponent α ≈ 0.35 and the data exponent β ≈ 0.37. In the regime AI is actually trained in, data and compute contribute to model capability at nearly the same rate.
Read that as an economist and it is astonishing. The machine-learning field’s own best-fit law says that data and compute are roughly co-equal engines of model capability. Compute is paid for — lavishly, in hundred-billion-dollar data-center commitments. Data is paid for at roughly the rate of a rounding error. Two co-equal inputs; one of them is free. That asymmetry is the subject of the rest of this paper.
2. Road One — The Scaling Law, Read as a Production Function
The relationship in Figure 1 is the Chinchilla scaling law (Hoffmann et al., 2022). For a model with N parameters trained on D tokens of data, expected loss is:
E is the irreducible floor — the entropy you can never train away. The two terms after it are the reducible loss: the part of capability that parameters buy and the part that data buys. This is, structurally, a production function. Output (capability, measured as loss removed) is produced by two inputs (compute and data) with diminishing returns to each. An economist would recognize it instantly; it simply uses loss instead of dollars.
So let us treat it like one. The marginal product of data — the capability bought by one more token — is the derivative of loss with respect to D:
It is strictly positive and diminishing, exactly as a factor’s marginal product should be. We can go further and ask the question a cost-minimizing lab actually faces: how much compute is one unit of data worth? That is the marginal rate of technical substitution between the two inputs — the data-compute exchange rate:
This matters because it means data and compute are priced in the same units — capability. A lab choosing between buying more GPUs and buying more data is, whether it says so or not, comparing their marginal products. Compute’s side of that comparison is settled in cash. Data’s side is settled by scraping.
Now the step that gives us γ. In a competitive market, each input earns a share of output equal to its output elasticity. The scaling law’s exponents are those elasticities — that is what an exponent in a power law means. Normalizing data’s exponent by the total gives the scaling-law-implied data share:
With the reanalyzed Chinchilla exponents, this lands near one-half; with the original 2022 fit it lands near 0.45. Either way the conclusion is the same and it is not subtle: by the AI field’s own measurements, data is responsible for something close to half of what makes a model good. Call this estimate γ̂ ≈ 0.40–0.51, and treat the top of that band with caution — capability is not identical to revenue, an objection we take seriously in §5. But the floor of the band is already an order of magnitude above what data is paid.
3. Road Two and Three — What the Market and the Courtroom Actually Pay
The scaling law tells us what data is worth inside the model. The second and third roads tell us what it is paid outside the model — and they use real transactions, not theory.
Start with the market. A licensing economy for training data has formed in the last two years. The reference points are now public: Amazon pays the New York Times roughly $20–25 million a year for training access; News Corp’s deal with OpenAI runs on the order of $50 million a year; Reddit’s data-access arrangements are in the same neighborhood; and the upper end of bespoke frontier deals reaches a few hundred million.
Figure 2. Representative annual values of known AI training-data deals. These are real prices — and they are tiny. The largest are low hundreds of millions per year, set against an AI sector earning tens of billions per year.
Sum every known licensing arrangement and amortize the one-time settlements across the years they cover, and the total flowing to data owners is, generously, low single-digit billions per year. Divide by AI-sector revenue and you get data’s paid share:
About two percent. The third road — the courtroom — gives the same order of magnitude from a completely different direction: the largest data settlement in history, Anthropic’s $1.5 billion, is a one-time payment against an enterprise value measured in the hundreds of billions; annualized against the value it helped create, it too rounds to a couple of percent. Two independent measurements of the paid share, one from contracts and one from litigation, agree: data is currently paid on the order of 2% of the value it creates.
4. The Wedge
We now have both numbers the series was built to compare. Data’s value share, from the scaling laws, is γ̂ ≈ 0.45. Data’s paid share, from the market and the courts, is about 0.02. The Data Capture Ratio from Article I is simply their quotient:
Φ ≈ 0.04. Data’s owners receive about four cents of every dollar their data earns. Drawn on a single axis, the gap between what data is worth and what data is paid is not a discrepancy at the margin — it is most of the distance.
Figure 3. Data’s value share (teal band, from scaling laws) against its paid share (red line, from market and settlements). The distance between them is the Data Capture Ratio’s complement — the wedge — and it spans almost the entire axis.
Put the wedge in dollars. Multiply it by the revenue of the pure-play AI sector and you get the annual compensation gap — the return data earns but does not receive:
Figure 4. Owed versus paid, per year. On these figures the gap is on the order of $34 billion annually — and that is computed against today’s revenue, before the data wall of Article I tightens the market further.
Roughly thirty-four billion dollars a year, on conservative inputs, earned by data and kept by the firms that hold it. To our knowledge it is the largest persistent factor-pricing anomaly in any modern industry. It is not a market failure in the usual sense — markets fail quietly. This one publishes its own exponents.
5. Four Objections, Taken Seriously
A number this large invites pushback, and it should. Four objections are worth answering directly.
“Capability isn’t revenue.” Correct, and it is why we band γ̂ rather than assert a point. The scaling exponents measure data’s share of capability, and the mapping from capability to dollars is imperfect. But notice the size of the safety margin: you could halve our estimate — insist data is worth only a quarter of model value, not a half — and the paid share of 2% is still off by a factor of ten. The conclusion survives almost any reasonable haircut.
“Synthetic data makes real data substitutable, so its share is small.” The data-constrained scaling literature (Muennighoff et al., 2023) says the opposite: repeated and synthetic data hit sharply diminishing returns after a few epochs. Real, fresh, human data is not freely substitutable — the elasticity of substitution is low — which is precisely the condition under which a scarce factor’s share is high, not low. Substitutability is an argument for the data wall, not against the data share.
“The labs add the value — compute, algorithms, fine-tuning.” They add real value, and we never claimed otherwise. That value is the α term and the labor share; it is paid, correctly, in salaries and chips. Our claim is narrower and harder to dodge: of the value created, roughly half traces to data, and that half is the part paid nothing.
“The data was public and free for the taking.” Public is not the same as valueless — the licensing market and the settlement both put a positive price on public and published material. As Article I argued, data’s nonrivalry means a firm can use it without using it up; that makes free access a windfall for the holder, not evidence that the input is worthless. A thing can be freely takeable and still be owed for.
6. What This Sets Up
If data’s share is real and large — call it a third to a half of AI value, paid at four cents on the dollar, a gap of tens of billions a year — then the next question is unavoidable and it is not aggregate. It is personal. That $34 billion is owed to someone. To whom, and how much each?
Article III — The Marginal Product of a Person takes the aggregate γ measured here and pushes it down to the individual contributor, using data-attribution methods (Data Shapley and its successors) to ask what a single person’s data is worth at the margin — and how nonrivalry, which let one dataset train a thousand models, multiplies that answer rather than dividing it.
We measured the size of the fourth factor’s paycheck in this paper. Next we find out whose name should be on it.
References and Sources
- Hoffmann, J., et al. (2022). “Training Compute-Optimal Large Language Models” (the Chinchilla paper). NeurIPS. Scaling law L(N,D) = E + A/Nᵅ + B/Dᵝ; compute C ≈ 6ND; ~20 tokens/parameter compute-optimal.
- Besiroglu, T., et al. (2024). “Chinchilla Scaling: A Replication Attempt.” Reanalyzed exponents α ≈ 0.348 ± 0.039, β ≈ 0.366 ± 0.039.
- Kaplan, J., et al. (2020). “Scaling Laws for Neural Language Models.” Earlier exponent estimates and the original scaling-law framework.
- Muennighoff, N., et al. (2023). “Scaling Data-Constrained Language Models.” Diminishing returns to repeated/synthetic data beyond a few epochs.
- Jones, C. I., & Tonetti, C. (2020). “Nonrivalry and the Economics of Data.” American Economic Review, 110(9), 2819–2858.
- Goodreads / industry reporting (2025). “Amazon Pays NYT $20–25M/Year for AI Training Data.” Market benchmark for premium training-data licensing.
- Quartz (May 2026). “The Price of AI Training Data, From $5M to $250M.” Range of licensing-deal values; News Corp–OpenAI ~$250M / 5 years.
- MediaPost / The Information (May 2026); Yahoo Finance / 5WPR (May 2026). AI-sector revenue: ~$80B annualized across pure-play AI firms; Anthropic ~$30B; OpenAI ~$25B.
- Anthropic author class-action settlement (preliminary approval September 2025): $1.5B; ~$3,000 per work.
- Vector1 Research (2026). Returns to Data, Article I — The Fourth Factor. Definitions of the augmented production function, the data share γ, and the Data Capture Ratio Φ.
Methodological note: the scaling exponents are taken from the published literature; γ̂ = β/(α+β) is a modeling estimate that treats the scaling-law exponents as local output elasticities and assumes capability tracks economic output — an assumption we band and stress-test in §5 rather than assert. The paid share and dollar gap use run-rate revenue and publicly reported deal values, are nominal, and are intended as order-of-magnitude estimates, not audited figures. Full derivations and code: github.com/vector1-ai (forthcoming).
Brian Curry is a Kansas City–based data scientist, AI researcher, and founder of Vector1 Research. The Returns to Data program develops the economic theory of data ownership in the age of artificial intelligence.
Contact: brian at vector1.ai · License: CC BY 4.0
#ReturnsToData #DataEconomics #ScalingLaws #AIandLabor #FactorShares #DataOwnership #DataDividend #AIEconomics #Chinchilla #Vector1Research
