AI data market
AI data TAM execution constraints: why a bigger market can add nothing
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
A large AI data TAM can coexist with a small company because output is the smallest of three limits: demand the company can win after substitutes and locked contracts, supplier datasets that pass rights and quality review, and review hours. In this article's illustrative model, doubling the market raises delivered datasets 20 percent, and quadrupling it adds nothing more once review capacity binds.
Value drivers
AI data TAM execution
Key takeaways
- Delivered datasets equal the smallest of three limits: winnable demand after substitutes, supplier datasets that pass review, and review hours; the TAM enters only the first.
- In the illustrative central case, doubling addressable needs lifts output from 30 to 36 datasets a year and quadrupling adds nothing more, so the company's share falls from 3.0 to 0.9 percent.
- Cutting review hours per dataset alone adds nothing at today's market; combined with a doubled market it doubles output, so market size and capacity must be valued together.
- Published examples show substitutes at work: over 98 percent synthetic alignment data at NVIDIA, opt-in user data at Anthropic and third-party data products listed on AWS Data Exchange.
- Ask any AI data company which limit capped output in each recent quarter, with requests, accepted datasets and review hours as denominators.
Why can a large AI data market still produce a small company?
Because a company's output is the smallest of three limits, and the market size enters only one of them. The other two are the supplier datasets that pass rights, privacy and quality review, and the review hours available to process them. When either supply limit binds, a larger total addressable market changes nothing about what the company can deliver this year.
Big adoption numbers make the gap easy to miss. The U.S. Census Bureau reported that 19.8 percent of businesses said they were using AI in any business function as of May 3, 2026, after revising its survey question in November 2025 from AI use in producing goods or services. That figure measures firms using AI. It does not measure firms buying data, the budget for data, or the share of any need that licensed business records could meet, and it says nothing about how many datasets one supplier can deliver.
The thesis an investor can test is narrow: for an AI data company at the stage when most market slides are shown, the binding limit may well be supply rather than the market. The test is not a bigger or smaller TAM estimate. It is a question about the company: which of the three limits capped last quarter's output, and how fast is it moving?
The three-limit bridge from addressable needs to delivered datasets
Delivered datasets per year equal the smaller of a demand line and a supply line. The demand line starts from addressable needs, removes the needs met by substitutes, keeps the share open to purchase this year, and applies the company's win share. The supply line takes supplier datasets offered for review, caps them at what review hours can process, and keeps the share that passes rights, privacy and quality review.
With the illustrative central inputs below, the demand line is 1,000 x (1 - 0.50) x 0.60 x 0.10 = 30 datasets a year. The supply line is 0.60 x the smaller of 120 offered and 6,000 hours / 100 hours each = 60 reviewed, which gives 36 accepted. Output is the smaller of 30 and 36, so 30 datasets, or 3.0 percent of the 1,000 addressable needs, and demand is the binding limit. Across every combination of the low, central and high values in the table, 6,561 in all, output ranges from 3 to 120 datasets a year; the extremes multiply several tail values at once.
Every value is an editorial assumption for a reader's own run, not an estimate of any market or of SourceX, and the model counts datasets rather than dollars. The last column names the evidence that would replace each assumption in a real diligence file:
| Input (Illustrative) | Low | Central | High | Evidence that would replace it |
|---|---|---|---|---|
| Addressable data needs per year in the segment | 500 | 1,000 | 2,000 | Dated buyer requests by task type, not a market report |
| Share of needs met by substitutes | 30% | 50% | 70% | The alternative each buyer tried before asking |
| Share open to purchase this year | 40% | 60% | 80% | Renewal and expiry dates of buyers' existing data contracts |
| Company win share of open purchases | 5% | 10% | 20% | Requests won divided by requests answered |
| Supplier datasets offered for review per year | 40 | 120 | 240 | Dated supplier submissions entering review |
| Share passing rights, privacy and quality review | 40% | 60% | 80% | Accepted divided by reviewed, by stage of rejection |
| Review hours available per year | 4,500 | 6,000 | 9,000 | Staff and partner hours assigned to review |
| Review hours per dataset reviewed | 60 | 100 | 150 | Time logs by review stage over several quarters |
What substitutes and open channels take out of the demand line
Substitutes remove needs before any supplier competes for them; in the model, moving the substitute share from 30 to 70 percent halves output from 36 to 18 datasets. Published examples show three kinds of alternative at work.
Synthetic data is the first. NVIDIA's research page for Nemotron-4 340B, published in June 2024, says over 98 percent of the data used in its model alignment process was synthetically generated. Microsoft's Phi-4 technical report, from December 2024, describes 50 broad types of synthetic datasets totaling about 400 billion unweighted tokens. Neither figure measures any buyer's spending, and both describe one developer's models, but they show that a need for training or alignment data does not have to become a purchase.
First-party data is the second. Anthropic's 2025 update to its consumer terms lets users of its Free, Pro and Max plans choose whether their chats and coding sessions may be used to train models, with a five-year retention period for those who allow it. Data a developer collects from its own users competes with data it would otherwise license.
Open channels are the third. AWS Data Exchange publishes third-party data products to the AWS Marketplace catalog, where subscribers can find products and review public or custom offers. Where a buyer can subscribe to a listed product, a direct supplier competes on fit and rights rather than on access, which is what the win-share input stands for.
Why review capacity does not grow with the market
Review work is done per dataset and often per recipient, so it scales with deliveries, not with the market. A market four times larger brings four times the requests to the same review desk.
U.S. health rules give a concrete case. Under HIPAA's expert determination method, 45 CFR 164.514(b)(1), a qualified expert must determine that the risk of identifying anyone is very small for an anticipated recipient, alone or with other reasonably available information, and must document the methods and results. That is a judgment about one dataset and one recipient, which cannot be done once for a whole market. Rights review works the same way: each supplier's contracts, notices and software terms are read for that supplier.
Capacity can grow, through hires, partners and tooling, but it grows in steps and with a lag, while a TAM estimate can double on a slide overnight. That asymmetry is why the review-hours inputs sit beside the market inputs in the bridge rather than in a footnote. The question of which hire or partner to fund belongs to a separate decision; this model only shows when capacity, rather than the market, sets the ceiling.
AI data TAM execution constraints in one chart: output as the market grows
Output rises with the market only until a supply limit binds, then goes flat. In the central case, demand binds at 1.0 times the market and review hours bind above 1.2 times; doubling addressable needs raises delivered datasets from 30 to 36, a 20 percent gain, and quadrupling them adds nothing more. The company's share of addressable needs falls from 3.0 percent to 1.8 percent at twice the market and 0.9 percent at four times.
The two levers multiply each other's value. Cutting review hours per dataset from 100 to 60 at today's market leaves output at 30, because demand still binds; doubling the market alone gives 36; doing both gives 60, double the starting output, with demand and supply meeting exactly. A team that funds capacity without demand evidence, or cites market growth without capacity evidence, is pulling one lever of two.
The chart draws the same line for three values of review hours per dataset: 60, 100 and 150. The table below varies each input on its own between its low and high value, with every other input held at the central case, and names the limit that binds:
| Input varied alone | At the low value | At the high value |
|---|---|---|
| Addressable needs (500 to 2,000) | 15 datasets, demand binds | 36 datasets, review hours bind |
| Substitute share (30% to 70%) | 36 datasets, review hours bind | 18 datasets, demand binds |
| Share open this year (40% to 80%) | 20 datasets, demand binds | 36 datasets, review hours bind |
| Win share (5% to 20%) | 15 datasets, demand binds | 36 datasets, review hours bind |
| Supplier datasets offered (40 to 240) | 24 datasets, supplier pipeline binds | 30 datasets, demand binds |
| Review pass share (40% to 80%) | 24 datasets, review hours bind | 30 datasets, demand binds |
| Review hours per year (4,500 to 9,000) | 27 datasets, review hours bind | 30 datasets, demand binds |
| Hours per dataset (60 to 150) | 30 datasets, demand binds | 24 datasets, review hours bind |
Is a bigger TAM still worth paying for, if supply binds today?
Yes, but for the ceiling after capacity grows, not for this year's output. The objection says a large market is what makes capacity worth building, and that is right: in the model, the value of cutting review hours is zero at today's market and large at twice the market. Market size and capacity are complements, so neither should be priced alone.
The weaker version of the objection treats broad pools as demand. AI adoption rates, AI spending on compute, global data volume and venture funding are all real, but each counts something other than purchases of licensed records, and the Census adoption figure above is a good example of a large number that measures firms, not budgets. The live methodology page on data licensing market reports sets out how a market figure should define its population and sources; this page starts where that one stops, at the company.
The right use of a large TAM in diligence is as an upper bound on the demand line, paired with evidence for the other inputs. A TAM slide with no win share, no substitute share and no capacity numbers tells an investor where the company would like to be, not where its output can go.
Comparing two AI data companies on execution limits, not market size
Compare companies on which limit binds and how fast it moves, using denominators that a market estimate cannot supply. Two companies citing the same TAM can sit in opposite regimes: one capped by demand it cannot yet win, the other capped by review hours with requests waiting.
The criteria below turn the bridge into a diligence comparison. Ask both companies for the same evidence, over the same quarters, and score each row on whether it was provided with the stated denominator:
| Criterion | Evidence to request | Denominator | Warning sign |
|---|---|---|---|
| Obtainable demand | Dated buyer requests with scope, by month | Requests received, not market size | Pipeline described only as a share of TAM |
| Substitution exposure | For each request type, the alternative the buyer tried | Requests where an alternative was tried | No answer to what the buyer uses instead |
| Open share | Renewal and expiry dates of buyers' data contracts, where known | Purchases open in the year | Assumes all spending is open every year |
| Accepted supply | Supplier datasets offered and accepted after review | Datasets entering review | Counts signed suppliers instead of accepted datasets |
| Review capacity | Review hours per dataset by stage, over time | Datasets reviewed | Hours per dataset unknown or not falling |
| Binding limit | Which limit capped output in each recent quarter | Quarters observed | Growth claims that never name a binding limit |
Risks that could make the three-limit model wrong
Several developments could make the bridge mislead, and each changes a specific input, so a reader can watch for it rather than distrust the whole model.
The model is static: it shows one year and ignores how fast capacity ramps or how requests age while waiting. It also treats inputs as independent, although a buyer with large needs may impose stricter acceptance tests, which would lower the review pass share exactly when demand rises.
- Automation cuts review hours per dataset sharply. Then supply limits loosen and the market becomes the binding limit sooner, which favors companies with demand evidence.
- Substitutes spread further. NVIDIA's report that over 98 percent of its alignment data was synthetic shows how far one developer has gone; if that pattern extends to more tasks, the demand line shrinks for every supplier at once.
- Buyer concentration. When a few developers hold most needs, one buyer's change of plan moves addressable needs and win share together, and the range in the input table understates the swing.
- First-party collection grows. Programs that let users opt in to training use, such as Anthropic's consumer setting, can absorb needs that licensed data would otherwise meet.
- Per-dataset review requirements rise. More documentation per dataset raises hours per dataset and moves the kink in the chart to a smaller market multiple.
Testing the bridge on SourceX: questions for the investor overview
SourceX is a direct application of the bridge, because its own pages describe output as project throughput rather than inventory. Datasets for AI buyers are sourced on request from U.S. businesses, not held in stock, and buyers look for years of consistent history and tens of thousands of records or more in one system, assessed deal by deal. Its how-it-works page adds that positive interest is not yet a purchase, and that qualifying does not by itself mean an offer will follow.
SourceX publishes no demand, pipeline, supplier or buyer figures, and none are implied here; its investor page says, "We share company and commercial metrics with investors directly." The company's current raise, under Rule 506(c), is open only to accredited investors, and materials are requested through that page. The questions below map each limit of the bridge to evidence an investor can ask for:
- Which of the three limits, demand, supplier pipeline or review capacity, capped delivered datasets in each of the last four quarters?
- How many buyer requests arrived in each quarter, and what alternative had each buyer tried first?
- Of supplier datasets entering review, what share passed rights, privacy and quality review, and at which stage did the rest stop?
- What are review hours per dataset by stage, and how have they moved over the last year?
- If addressable needs doubled next year, which limit would bind first, and what evidence supports that answer?
Methods note: how the bridge was computed and what it leaves out
The bridge, inputs, sensitivity table and chart were built for this article on 2026-10-28 with a short Python script using exact fractions; every output in the text is computed by that script, and the arithmetic for the central case is shown in full above. Output = min(D x (1 - s) x c x w, p x min(q, H / h)), where D is addressable needs, s the substitute share, c the share open this year, w the win share, q supplier datasets offered, p the review pass share, H review hours available and h review hours per dataset. All eight inputs are illustrative editorial assumptions in dataset counts and percentages; none is an estimate of a market or of any company, and the model contains no dollar values.
The chart holds every input at its central value except addressable needs, which run from zero to four times the central 1,000, and review hours per dataset, shown at 60, 100 and 150. The full-grid range of 3 to 120 datasets covers all 6,561 combinations of each input's three listed values. Public figures are dated in the text: the Census Bureau's May 2026 story on business AI use, NVIDIA's June 2024 Nemotron-4 340B page, Microsoft's December 2024 Phi-4 report, Anthropic's 2025 consumer terms update, the AWS Data Exchange user guide and the current eCFR text, all checked on 2026-10-28.
Not measured: any company's actual requests, acceptance rates or review hours; correlations between inputs; how capacity ramps over time; spending in dollars; and the share of any buyer's needs that business records, rather than publisher, platform or synthetic data, could meet.
Frequently asked questions
What is an execution constraint in AI data market sizing?
It is a limit inside the company that caps output regardless of market size, such as the supplier datasets that pass rights, privacy and quality review or the review hours available. When one of these binds, a larger TAM does not raise this year's delivered datasets.
How should an investor use a TAM figure for an AI data company?
As an upper bound on the demand line, paired with evidence for substitute share, open share, win share and supply capacity. A TAM with none of those numbers says where a company would like to be, not what it can deliver.
Why does doubling the market not double output in the model?
Because in the central case review hours bind once addressable needs pass 1.2 times the starting level. Doubling the market lifts output from 30 to 36 datasets a year, and further growth adds nothing until review capacity grows.
Does synthetic data shrink the market for licensed business data?
It shrinks the demand line for some tasks: NVIDIA says over 98 percent of Nemotron-4 340B's alignment data was synthetic. It does not remove needs that depend on real operational history, which is why the substitute share is an input to measure rather than a constant.
Does SourceX publish its demand or pipeline numbers?
No. SourceX publishes no demand, pipeline, supplier or buyer figures. Its investor page says those company and commercial metrics go to investors directly, and materials are requested there.
Sources
- The U.S. Census Bureau reported a national rate of 19.8 percent of businesses using AI in any business function as of May 3, 2026, after revising the question in November 2025 from AI use in producing goods or services. Source
- NVIDIA's research page for Nemotron-4 340B (June 2024) says over 98 percent of the data used in its model alignment process is synthetically generated. Source
- Microsoft's Phi-4 technical report (December 2024) describes 50 broad types of synthetic datasets totaling about 400 billion unweighted tokens. Source
- Anthropic's 2025 consumer terms update lets Free, Pro and Max users choose whether chats and coding sessions are used to train models, with five-year retention for those who allow it. Source
- AWS Data Exchange documentation says data products are published to the AWS Marketplace catalog, where subscribers find products and review public or custom offers. Source
- 45 CFR 164.514(b)(1) requires a qualified expert to determine that identification risk is very small for an anticipated recipient and to document the methods and results. Source
- SourceX's how-it-works page says positive interest is not yet a purchase and that qualifying does not by itself lead to an offer. Source
- SourceX's investor page says it shares company and commercial metrics with investors directly. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.