A survey of the data market

Money is pouring into AI training data. Labs spend heavily on expert contractors, license deals make headlines, and a new industry sells environments for reinforcement learning. People like to say data is the new oil.

I think the analogy breaks in a way that matters. A barrel of oil is worth the same after someone else buys one. A dataset is only worth something to a model that hasn’t learned what it teaches, and each model release learns a bit more of it from somewhere else. To me it looks more like fresh produce: worth the most when it arrives, worth less every week it sits on the shelf, and worth nothing to someone who already has it.

If that’s right, it changes what the market will end up pricing, who captures the value, and which parts of it will still exist in a few years.

three markets under one name

“The data market” is really three markets, and they have different economics.

The largest by far is a labor market. Most of the money labs spend on “data” pays for experts to produce work to order: rubrics, graded answers, trajectories, reviews. Mercor reports a gross run rate of about $2 billion (Dealroom). Meta paid $14.3 billion for 49% of Scale AI in 2025 (Nasdaq). These businesses pass most of their revenue to contractors and are priced, in the end, on hours.

The second is a licensing market for corpora that already exist. It is much smaller than its headlines. Reddit, the most visible seller, disclosed $203 million of licensing contracts spread over two to three years (TechCrunch), a little over a month of Mercor’s current run rate.

The third is a market for environments and graders, and it is the newest. Epoch reports reinforcement-learning tasks selling for $200 to $2,000 each, replicas of a website for about $20,000, a high-fidelity replica of an app like Slack for about $300,000, and exclusive deals at four to five times the non-exclusive price (Epoch). Mechanize argues the right spend per task is tied to the compute it will consume: about $480 per training run at current inference prices, or about $2,400 over the five runs a task is typically reused for (Mechanize).

	Expert labor	Licensing	Environments and graders
What is sold	Work made to order	Existing corpora	Tasks, simulators, verifiers
Priced on	Hours, cost plus margin	Legal risk and provenance	The compute the task will consume
Typical unit price	Hourly expert rates	About $5,000 per book	$200 to $2,000 per task
Scale today	Mercor about $2B gross run rate	Reddit $203M over 2–3 years	Contracts of six to seven figures a quarter
Buyers	A handful of frontier labs	Labs, a few enterprises	Labs, increasingly built in-house
Direction of price	Up per hour, down per capability	Down, as courts settle on fair use	Down with inference prices

The labor vendors deserve a closer look, because their revenue is what most people mean when they say the data market is booming. Of every $100 of gross revenue, about $60 to $70 goes straight to contractors (Sacra), so a $2 billion run rate is $600 to $800 million of net revenue. That looks a lot more like a staffing business with very good customers than a company that owns an asset.

The price ladder shows where the market does add value. One unit at each step sells for an order of magnitude or two more than the step below it, and exclusivity multiplies everything again:

• existing corpus
• task
• environment
• compute, for scale

• Pirated book (settled)$3k
• Licensed book$5k
• RL task$200–2k
• Compute per task$2.4k
• Complex coding task$20k
• Website replica$20k
• Slack-like replica$300k

$100$1k$10k$100k$1M

Price per unit, log scale; the lighter segment marks a range. Exclusive terms run four to five times these prices. Sources: the Bartz settlement, HarperCollins–Microsoft, Epoch's survey of RL-environment pricing, and Mechanize's compute estimate.

The row I find most interesting isn’t data at all. Mechanize estimates a single task consumes about $2,400 of compute over its life, which sits right in the middle of the task prices (Mechanize). I don’t think that’s an accident. Labs are starting to price tasks by the compute they will spend on them, so the price of data is getting pegged to the price of inference, and will probably fall with it.

These three markets move in different directions. Labor revenue can rise even while the value of the data falls, because a lab that needs more examples to get the same gain buys more hours. Licensing prices are set by courts, and courts have been lowering them. Environment prices are set by the cost of compute, and fall with it. So adding them up into one “data market” number tells you very little about what data is actually worth.

climbing the stack

Seen from above, the market is a stack. Raw data sits at the bottom, sold by many sellers, and its value decays fastest. Each layer above it turns the one below into something more specific to a buyer’s goal: labels into graded tasks, tasks into evaluations and graders, graders into environments a model can practice in, environments into a measured improvement on a model, and finally into deployed work.

1. 01Raw data
   Existing corpora$3–5k a book
2. 02Labels
   Expert hoursHourly rates↑ Mercor's start
3. 03Tasks
   Graded problems$200–2,000 a task
4. 04Evals and graders
   Verifiers, rubrics, benchmarksScarcest; rarely sold alone↑ Mercor's APEX
5. 05Environments
   Working replicas of real software$20k–300k each↑ Mercor buys Deeptune
6. 06Post-training
   A measured gain on a modelContracts in six to seven figures↑ Mercor post-trains Nemotron
7. 07Deployment
   Work done, and the data it createsEnterprise contracts

commodity, many sellers, value decays fastscarce, few sellers, value owned

Step heights are ordinal, not to scale. Arrows mark one vendor's climb, from its own announcements. Prices from the Bartz settlement, HarperCollins, and Epoch's survey of RL-environment pricing.

Most vendors seem to be trying to move up, because that’s where the value accrues and stays. Mercor’s own announcements trace the path. It started as a marketplace for expert labor. Its experts now write the tasks and verifiers behind APEX, its benchmark of professional work. It agreed to acquire Deeptune, which builds training environments that recreate software like spreadsheets and Salesforce (Mercor). And it has published post-training results: with a few hundred tasks per domain, NVIDIA’s Nemotron 3.5 Super gained 5 to 8 points on held-out legal, software and accounting benchmarks, in partnership with Harvey, Cognition and Ramp (Mercor).

A few hundred tasks is not much labor. The more efficiently models learn skills, the fewer hours there are to sell, and the more a vendor has to sell the outcome instead.

why nothing has a unit price

The obvious fix for a market without prices is a way to look at a dataset and say what it is worth. Economists have known since Kenneth Arrow why that is hard: you can’t value information without seeing it, and once you’ve seen it you don’t need to buy it. The modern data-valuation methods don’t escape this. They estimate a dataset’s worth by training on it, which just turns the paradox into a compute bill.

The bigger problem is that a dataset doesn’t have one value. What a buyer pays for is the improvement the data makes to its own model, and that depends on what the model already knows. The same corpus is worth less to a larger model that has already seen something like it. The value of data also depends on how much compute it will be trained with, and falls each time it is repeated (Goyal et al.; Muennighoff et al.). One consequence: a small lab can rationally pay more per token for a curated set than a frontier lab will.

And the value depreciates. Each frontier release learns some of what a dataset could teach, from other data, synthetic data or other models’ outputs, so a corpus loses part of its remaining value at every release. How fast it depreciates is the most important number in this market, and I could not find a single study that measures it.

• static corpus
• live feed (per month)
• exclusive corpus, edge over rivals

Illustrative, not measured. Assumes a frontier release every six months that absorbs 35% of what the data can still teach, and a 12-month exclusivity window. On those assumptions a static corpus has a half-life of about 10 months.

The numbers in the figure are my assumptions, not estimates, but I think the shape holds for any reasonable values. It also explains why exclusivity sells at four to five times the base price (Epoch). What an exclusive deal really buys is time before rivals get the same data, so I’d price it like a head start.

So when this market does standardize, I don’t think it will be on tokens. It will be on what buyers actually care about: how much a dataset improves their model on the evaluations that matter to them, and how long they get that improvement before anyone else. Financial market data already works this way. What an exchange sells is not prices, which are public a moment later, but how soon you get them and who else has them.

why niche wins

Data is non-rival: my training on a dataset doesn’t stop you training on the same one. I think that one property explains most of this market. Oil keeps its value for whoever holds it because using it uses it up. A dataset only gives its holder an edge while others don’t have it. Copy it to every lab and it improves every model equally, buys no one an advantage, and stops being worth paying for.

So the value of any dataset comes down to two questions. Can you keep others from getting it? And can the frontier learn the same thing somewhere else?

Placements are qualitative. Because data is non-rival, the only things that keep it valuable are exclusion, so others can't have it, and difficulty, so the frontier can't learn it some other way.

Public web data fails both tests: anyone can scrape it, and every model has. Licensed corpora pass the first and fail the second. They can be exclusive for a while, but the frontier learns the same domain from other sources, which is the decay in the chart above. Niche data passes both. A company’s own workflows, measurements from the physical world, the traces a deployed product generates, and the judgment of a few hundred real specialists are hard to get and hard to learn any other way. They are also where the labs are weakest, because the most valuable of this data is contractually off limits to them.

Niche data is also worth more per example than it used to be, because so few examples are now needed. When a few hundred tasks can move a model five to eight points on professional work, each task carries a large share of that gain. So niche markets look small if you count tokens, and much bigger if you count value per token.

synthetic data and the long tail

Whether synthetic data can replace real data for the long tail depends on whether there is a checker. Where a verifier encodes how the domain works, as in mathematics, code and geometry, synthetic data has already replaced most human data. AlphaProof trained on tens of millions of auto-formalized problems (Hubert et al.), and a single example with a verifiable reward can lift a model’s MATH500 score from 36.0% to 73.6% (Wang et al.). In those domains the scarce input is the verifier, not the data.

Where there is no checker, synthetic data trims the tail rather than extending it. Dohmatob and colleagues show the mechanism: a model’s samples under-represent rare cases, so training on them cuts off the tail of the distribution and scaling stalls once synthetic volume passes a threshold. What restores it is real data just beyond the cut (Dohmatob et al.). Keeping real data in the mix avoids model collapse, but only because the real data never leaves (Gerstgrasser et al.). Microsoft’s phi-4 trained on about 40% synthetic data, yet its synthetic-only ablation lost almost 15 points on TriviaQA and hallucinated more (Microsoft).

I think the useful split is between skills and facts. The number of examples needed to teach a skill has collapsed, as the one-example result shows. The number needed to store a fact has not: a model still needs on the order of a hundred to a thousand exposures to a fact to hold it reliably (Allen-Zhu and Li). So the market for skill data is shrinking fast, while facts about the world’s tail keep needing real data. Facts, though, are increasingly served by retrieval rather than training, which makes the valuable form of that data a live feed, not a corpus.

Waymo is the usual counterexample, but I think it shows something else. Waymo simulates edge cases the fleet has never seen, like tornadoes and elephants on the road, and its world model credits those scenarios to the video pretraining of Google’s Genie 3 (Waymo). So the edge case wasn’t invented by simulation; it came from a much larger corpus. “You can’t simulate what you haven’t seen” still holds, but “seen” means seen anywhere in pretraining, and whoever owns the largest pretraining corpus has seen the most.

who owns the loop

Most analysis of this market assumes data is collected, sold and trained on, in batches, forever. Continual learning breaks that assumption, and I think it’s the part that matters most.

If a model learns from its own deployment, the scarce thing stops being data and becomes the loop where the model meets the world and gets feedback. Cursor already runs this loop for its tab completions: it trains on live usage, ships a new checkpoint every 1.5 to 2 hours, and reports 21% fewer suggestions with a 28% higher acceptance rate (Cursor). That data never enters a market; it’s a by-product of distribution.

Ronald Coase’s question about the boundary of the firm applies directly: when is it cheaper to produce something in-house than to buy it? Once a firm’s own operations produce the data as exhaust, the answer flips. Map data is the precedent. In 2009, Google quietly stopped using Tele Atlas for its US maps and switched to its own data, built from Street View and Android (Search Engine Roundtable). The largest buyer internalized the moment it had a by-product source, and the vendor’s position with its biggest customer collapsed. My guess is that agent traces from deployed products are the frontier labs’ Street View.

Two constraints keep this from being a clean win for the labs. The first is that cheap feedback is a dangerous reward. In April 2025, OpenAI rolled back a model update that had become sycophantic, after an over-weighted thumbs-up signal (OpenAI). The second is more important: the most valuable long-tail data, from businesses, is contractually excluded from training. Commercial and API customers’ data is off limits under the major labs’ terms (Anthropic).

I think that second constraint points to where the market lands. The enterprise tail can’t be sold to the labs, so my guess is it gets learned from inside each enterprise, through per-company post-training and evaluation. That is a market for services and infrastructure, not for data, which is why the largest data vendors are already turning themselves into deployment firms. And a company that owns its deployment loop should keep the data it generates, because selling it hands a rival the same improvement.

what the law is pricing

The legal picture is often read as making data expensive. For general corpora it has done the opposite. In Bartz v. Anthropic, the court held that training on books was fair use, and the class settlement for pirated copies came to $1.5 billion, about $3,000 per work (Akin Gump). HarperCollins, meanwhile, licenses nonfiction titles to Microsoft for $5,000 each (eWeek). So the settlement came out cheaper per book than the license.

What the courts seem to be pricing is outputs and competition, rather than training itself. The Third Circuit has held that ROSS Intelligence’s use of Westlaw’s headnotes to build a competing product was not fair use (LawNext), and a Munich court held that Europe’s text-and-data-mining exception stops applying once a model memorizes a work (Taylor Wessing). A license today mostly buys proof of origin, cover against liability and rights to outputs, which is closer to insurance than to information.

what history says

Data markets have formed before, and the pattern is fairly consistent:

Market	What ended up trading	What happened	Lesson
Exchange data	Feeds, priced by speed and use	Prices held: the data is a by-product the exchange owns	Flows price well; stocks don’t
Credit	A derived score, not the records	Lenders share data to get access to it	The traded unit becomes a model output
Maps	Map databases	Nokia paid $8.1B for Navteq in 2007; in 2009 Google built its own from Street View	The biggest buyer internalizes once it has a by-product source
Weather	Forecasts built on free public data	IBM paid about $2B and sold for $1.1B eight years later	Value sits in processing, and it depreciates
Genomics	Exclusive access	GSK paid $300M in 2018; the whole company sold for $305M in bankruptcy	Exclusivity premiums decay

Sources: Computerworld, Search Engine Roundtable, Insurance Journal, Pharmaphorum.

If AI data follows the same path, I’d expect prices to settle on ongoing licensed feeds rather than one-off sales, the traded unit to be something derived, like a graded task or an evaluation, and the largest buyers to stop buying whatever their own products can generate.

headwinds

The bull case for this market is simple: more companies will train more models on more domains, forever, and each new domain needs new data. Several forces push the other way, and most of them are already visible.

Headwind	How it shrinks the market	Evidence today	Strength
Continual learning	Models learn from their own deployment, so data is produced in-house instead of bought	Cursor retrains on live usage every 1.5–2 hours; learning in the weights still suffers from forgetting	Early, potentially the largest
Sample efficiency for skills	Each new domain needs a verifier and a few hundred examples, not a large corpus	One example lifted MATH500 from 36% to 74%	Strong for skills, weak for facts
Synthetic data plus verifiers	Where a checker exists, data can be generated without limit	Math, code and geometry have largely moved to synthetic data	Strong where checkers exist
Distillation	Any capability in a released model can be copied from its outputs at inference cost	Labs publicly accuse rivals of distilling their models at scale	Caps prices on anything already released
Labs building in-house	Labs bring annotation and environment building inside to cut margins and protect secrecy	Epoch reports substantially more in-housing; xAI expanded its specialist tutors tenfold	Strong and rising
Buyers are producers	The largest buyers run the largest deployments and see the most edge cases themselves	Google built its own map data once Street View existed	Structural
Courts favoring fair use	Licenses buy insurance, not permission, so their price falls	Bartz: piracy damages below license prices	Strong for general corpora
Cheaper compute	Task prices are being pegged to the inference they consume	Mechanize’s per-task compute sits inside the price range	Steady downward pressure

In-housing figures are from Epoch and TechCrunch; the rest are sourced above.

Continual learning is the one I’d watch most closely. Today what has shipped is mostly memory and retrieval, and forgetting is still the main obstacle to learning in the weights. If it is solved, data stops being bought in batches and becomes a by-product of running the product, and the external market shrinks to what the labs can’t reach themselves: enterprise data they are barred from, physical-world data they don’t collect, and expertise they can’t get from users.

what is worth backing

If you’re investing, or deciding where to build, I think the useful question for any position is whether it adds something the frontier can’t absorb.

Position	Where it adds value	Main risk	Verdict
Selling static corpora	Provenance, cleaning, liability cover	Depreciates every release; priced by litigation; few buyers	Avoid unless it becomes a live feed
Lending against data	Unlocks capital tied up in corpora	Collateral depreciates and is worth different amounts to each buyer	Avoid
Expert-labor marketplace	Recruiting, quality control, neutrality between labs	Few buyers; each skill gets learned and stops selling	A staffing business; value it like one
RL environment builder	Turning tasks into graded environments	Labs building in-house; prices pegged to falling compute	Selective: domains labs can’t easily build
Verifiers for unchecked domains	The checker that makes synthetic data possible	Hard to build, and narrow	Strongest
Live feeds	Freshness that keeps renewing	Concentrated buyers	Strong
Owning a deployment loop	The edge cases, generated by real work	Needs real distribution	Strongest, if it keeps its data

Roughly, selling a depreciating stock once is weak, and owning something the frontier can’t absorb is strong. A verifier for a domain that lacks one is the scarce input to synthetic data. And I think owning a deployment loop in a specific industry, where the model meets real work, is the best position in this market, as long as the company keeps what it makes.

closing

Data is non-rival, and it goes off. Like fresh produce, it’s worth the most to whoever uses it first. So I don’t think the value stays with whoever holds the data; it goes to whoever learns from it first, and keeps learning. The first big training corpus was the public internet, which everyone has now read. The next one is being written by deployment, one task at a time, and most of it won’t be for sale.

The number I most want and couldn’t find is how fast data actually goes off. If anyone has measured how a dataset’s value changes across model releases, I’d love to see it.