What Is Your Data Worth to AI? Now We Can Actually Compute It
The technology to meter one person’s contribution to a trillion-dollar model arrived in the last 18 months — and it changes who can get paid
Returns to Data — Article III of a Vector1 Research Program
By Brian Curry · Vector1 Research
Article I argued data is the fourth factor of production. Article II measured its share: roughly 45% of AI value created, roughly 2% paid — a ~$34B annual gap. This paper answers the question those numbers force: whose money is it? We show the obstacle that made individual payment impossible — attribution — has quietly been solved by the machine-learning community itself, derive the payment formula it enables, and confront the two facts every honest data-dividend proposal must hold at once: your individual share is small today, and the claim it represents is the fastest-growing asset in the world.
1. The Objection That Just Died
Every conversation about paying people for their data eventually hits the same wall, and it sounds unanswerable: you can’t trace whose data did what. A trained model is a smoothie — billions of contributions blended beyond recovery. No meter, no market. For years this was not just rhetoric; it was true. The principled method for valuing a single training point required retraining the model on astronomically many subsets of the data. For a frontier model, the universe does not contain enough compute.
Then, over about eighteen months, the wall came down — and the people who demolished it were not economists or activists but machine-learning researchers solving their own engineering problems. They wanted to find mislabeled data, prune bad examples, and debug training runs. To do it they built, step by step, exactly the meter the economics was missing: approximation methods cut the cost by orders of magnitude; by 2025, “In-Run Data Shapley” computed every training point’s contribution during a single ordinary training run, at negligible overhead; by 2026, federated variants were marketed for real-time data pricing.
Figure 1. The compute cost of attributing a model’s performance to individual training contributions, 2019–2026. It has fallen by roughly six orders of magnitude — from “physically impossible” to “a rounding error on the training bill.” The standard objection to paying data’s owners was an engineering constraint, and it no longer holds.
Understand what this means for the series so far. Articles I and II established that data earns roughly 45% of AI value and is paid roughly 2%, leaving a gap on the order of $34 billion a year. The last defensible reason that gap couldn’t be distributed to individuals — no meter — has expired. What remains is arithmetic, and arithmetic is what this paper does.
2. The Fair-Division Formula
Suppose we accept the gap is owed. How should it be split among the people whose data created it? Equal shares feel democratic but are obviously wrong: a radiologist’s annotated scans did not contribute the same as a reposted meme. We need a division rule that rewards actual contribution — and game theory settled this question in 1953.
The Shapley value asks, for each contributor: across every possible order in which the dataset could have been assembled, how much did your contribution improve the model when it arrived? Average that over all orderings, and you get a single number — your marginal product, fairly computed:
The Shapley value is not one option among many; it is the unique division rule satisfying four axioms anyone would demand of a fair split — symmetry (identical contributions earn identically), null player (useless data earns nothing), additivity, and, crucially, efficiency:
Efficiency says the individual shares sum exactly to the whole — no value created by data is left unowned, and none is invented. Combine this division rule with Article II’s measurement of data’s total slice, and a person’s annual payment falls out as a two-term product: the size of data’s slice of AI revenue, times your fraction of the attributed total:
This is, to our knowledge, the first payment formula for individual data contributors in which both terms are now empirically computable: γ̂ from scaling laws (Article II), and φᵢ from in-run attribution (Figure 1). The economics and the engineering have finally met.
3. Your Data Is Not Average
The formula’s most important property is that it does not pay everyone the same — and the distribution of what it pays is the most counterintuitive result in this paper.
Attribution studies keep finding the same shape: a model’s capability owes a disproportionate debt to rare data. Feldman and Zhang’s influence work showed that neural networks depend heavily on memorizing the long tail — the unusual examples, the expert annotations, the underrepresented languages, the niche technical documentation. Common knowledge is learned from any of a million redundant sources, so no single source earns much credit for it. Rare knowledge has one teacher, and the Shapley value pays that teacher in full.
Figure 2. Attributed value per contributor, ranked. The distribution is a power law: a small head of rare, expert, or unique contributions earns large per-person value, while the broad base of common data earns little individually — but is enormous in aggregate.
Decompose any individual’s contribution and the same split appears inside it:
The economic implications cut in an unexpected direction. The people with the strongest individual claims are not influencers with large followings but the specialists: the rare-disease patient whose case notes taught the model something no textbook contains, the speaker of a low-resource language, the engineer whose niche documentation is the model’s only source on a topic. A data market priced by attribution is, structurally, a market that pays expertise and rarity — the opposite of the attention economy, which pays volume and virality. That alone makes it worth building.
4. The Multiplier Most Analyses Miss
Here is where individual data valuation breaks from every physical-resource intuition — and where most existing analyses, which stop at one model, undercount.
Your data is nonrival (Jones & Tonetti, 2020): using it to train one model does not use it up for the next. The same contribution can sit in the training corpus of every frontier model, every fine-tune, every distilled variant, every agent built on top — simultaneously. A rival input’s value gets divided across uses. A nonrival input’s value gets summed:
Figure 3. The same data contribution, valued across the number of models trained on it. Oil burns once; data teaches forever. As the model population grows — and in an agent economy it grows fast — the value of a single contribution multiplies rather than dilutes.
This is the deep reason data’s owners are underpaid even by the standards of Article II. Our $34B gap was computed against today’s model population. Every new model trained on the same corpora silently re-borrows every contribution in them, at a marginal price of zero. The liability compounds with each training run — which is also why settling it early is in the industry’s own interest, a point Article IV will make precise.
5. The Honest Number — and Why It Grows
Now the part most advocates flinch from, which is exactly why we will not. Divide data’s current slice of AI revenue by a realistic contributor base and the average annual payment, today, is small:
About thirty-six dollars a year for the average contributor. Critics will call that a rebate, not a revolution, and on today’s numbers they are right. But three forces push the figure upward, and all three are structural rather than speculative. First, the base: AI-sector revenue is the fastest-growing line item in the world economy; the same arithmetic at a $1–3 trillion sector yields hundreds to thousands of dollars per contributor per year. Second, the multiplier: §4’s nonrivalry effect compounds the claim with every model trained. Third, the distribution: §3’s power law means the average is the wrong statistic — for contributors of rare or expert data, attribution-based payments are large today.
Figure 4. The average annual data dividend per contributor as AI-sector revenue scales from $80B toward $3T, holding the Article II data share constant. The honest message: small today, structural tomorrow. The claim grows with the entire AI economy — which is rather the point.
The right way to think about a data dividend, then, is not as a check but as a birthright share in the fastest-compounding asset class in history. Nobody dismissed land reform because the first year’s harvest was modest.
6. Objections, Taken Seriously
“Attribution at frontier scale is still approximate.” True — in-run methods compute contributions for a specific training run, and estimates carry error bars. But factor payments have never required perfection: labor markets pay salaries on noisy performance signals; spectrum auctions price an invisible commons. The bar is defensible, not exact, and a meter with error bars beats no meter at all. The error bars, moreover, shrink every conference cycle.
“People will game the meter — flooding it with junk designed to attribute well.” Some will try, and the Shapley value’s null-player axiom is the first line of defense: data that doesn’t improve the model earns nothing, and redundant data earns next to nothing, automatically. Adversarial attribution-farming is a real research problem — but it is the same problem as data poisoning, which labs already invest heavily against. The meter and the immune system are the same machinery.
“Doesn’t this just incentivize people to generate data for money?” Yes — and that is a feature. A market that pays for rare, useful, verified contributions is an incentive to produce exactly the data the data wall (Article I) says is running out. Paying the fourth factor is not charity; it is how you grow its supply.
“Why should the firms agree to any of this?” Because the alternative is the courtroom, and the courtroom has already priced one corpus at $1.5 billion — for past use alone, from one company. Article IV takes up the mechanism-design question in full: which institutions move Φ from 0.04 toward 1, at what cost, and who moves first.
7. What This Sets Up
We now have the three load-bearing results of the program: data is a factor of production (I); its share is large and its payment is negligible (II); and the technology and the formula to pay its individual owners now exist (III). What’s missing is the institution — the mechanism that actually routes the money.
Article IV — Closing the Gap: Five Mechanisms, Scored compares licensing collectives, statutory royalties, sovereign data trusts, equity transfer, and attribution-metered micropayments on a single criterion: how far does each move the Data Capture Ratio, and at what administrative, legal, and incentive cost? The answer, we will argue, is not one mechanism but a stack — and the bottom layer is the meter this paper described.
The meter exists. The formula exists. The money exists. What remains is plumbing.
References and Sources
- Shapley, L. S. (1953). “A Value for n-Person Games.” Contributions to the Theory of Games II. The original fair-division theorem and axioms.
- Ghorbani, A., & Zou, J. (2019). “Data Shapley: Equitable Valuation of Data for Machine Learning.” ICML. First application of Shapley values to training-data valuation; retraining-based, exponential cost.
- Wang, J. T., Mittal, P., Song, D., & Jia, R. (2025). “Data Shapley in One Training Run.” ICLR 2025. Per-point Shapley contributions computed within a single training run — the metering breakthrough.
- Zeng, D., et al. (2026). “Ripple Shapley: Data Influence Attribution in One Federated Training Run.” AAAI 2026. Real-time federated attribution; explicitly framed as enabling real-time data pricing; ~62× speedup.
- Ilyas, A., Park, S. M., et al. (2022). “Datamodels: Predicting Predictions from Training Data.” ICML. Scalable attribution via surrogate modeling.
- Park, S. M., et al. (2023). “TRAK: Attributing Model Behavior at Scale.” Efficient attribution estimators for large models.
- Feldman, V., & Zhang, C. (2020). “What Neural Networks Memorize and Why: Discovering the Long Tail via Influence Estimation.” NeurIPS. The long-tail dependence underpinning §3.
- Koh, P. W., & Liang, P. (2017). “Understanding Black-Box Predictions via Influence Functions.” ICML. Foundational influence-function attribution.
- Jones, C. I., & Tonetti, C. (2020). “Nonrivalry and the Economics of Data.” American Economic Review, 110(9). The nonrivalry framework behind §4.
- Vector1 Research (2026). Returns to Data, Articles I–II: the fourth-factor production function; γ̂ ≈ 0.45 from scaling laws; paid share ≈ 0.02; Φ ≈ 0.04; gap ≈ $34B/yr.
Methodological note: Figure 1’s cost curve is a stylized synthesis of the attribution literature’s reported complexity improvements, on a relative log scale — direction and order of magnitude are literature-grounded; exact points are illustrative. Figure 2 is a stylized power-law consistent with long-tail influence findings, not a measured frontier-model distribution. The per-capita arithmetic uses Article II’s γ̂ and run-rate revenue, with a 1-billion-contributor base chosen deliberately as the conservative (payment-minimizing) assumption; the revenue path in Figure 4 is a scenario, not a forecast. Code and derivations: 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 #DataShapley #DataValuation #AIEconomics #DataOwnership #DataDividend #Nonrivalry #AIandLabor #MachineLearning #Vector1Research
