AI data market
The next AI data bottleneck forecast: why verification may bind next
By SourceX Editorial · Updated
9 sections
Short answer
The next AI data bottleneck is moving down the supply chain. Public human text may be fully used between 2026 and 2032 (Epoch AI), yet synthetic data and better curation now stretch it, so access binds mainly for private records. For business tasks, the tighter limit is verification: knowing whether an output is right. GDPval's primary metric is still human expert comparison. Buyer integration comes after.
Record timeline
Next AI data bottleneck
Key takeaways
- Treat AI data supply as four stages, access, preparation, verification and buyer integration, and ask which one limits a specific capability today.
- Access is easing for general training as synthetic and curated data stretch public text, but it still binds for private business records that need permission.
- Model developers are industrializing preparation themselves: DataComp-LM found the choice of filtering model alone moved 7B-model MMLU accuracy between 35 and 44 percent.
- Verification is the likeliest next constraint for business tasks: leading benchmarks rely on human-confirmed answers, and developers cap verification cost during training.
- Buyer integration lags capability: a 2026 Census working paper found 18 percent of firms used AI in a business function in late 2025 and early 2026.
Where is the AI data bottleneck now, and where could it move next?
For general model training, the bottleneck has moved off raw access to public text; for business-task capabilities, it is moving toward verification, the ability to tell whether an output is right, with buyer integration the stage after that. This is a forecast, not a measurement, and the rest of the page sets out the evidence for it, the alternatives and the signposts that would change it.
The useful unit is not data in general but a specific capability: a claims-handling agent, a procurement assistant, a code-review model. For each, data passes through four stages before it is worth money, and the stage that fails first sets the investment case. Capital that funds yesterday's bottleneck earns yesterday's returns, because value accrues to whoever relieves the stage that binds next.
The page gives a four-stage evidence scorecard, four scenarios for 2028, six calls with the tests that would prove them wrong, and the case that the real limit is not data at all.
Four stages a dataset passes before anyone pays for it
Access, preparation, verification and buyer integration: a dataset creates value only after it clears all four, and each fails for different reasons. Naming them separately stops one stage's scarcity from being credited to another.
A stage binds when relieving it would change the outcome and relieving the others would not. That definition matters to investors, because a company can sit in a growing market and still be relieving a stage that no longer binds.
- Access: the records exist and someone may use them. It binds when the data is used up, restricted, or locked behind permissions nobody can grant.
- Preparation: the records are made usable through extraction, deduplication, de-identification, filtering and linking context. It binds when usable records cost more to produce than a buyer will pay.
- Verification: someone can tell whether a model's output on the task is right. It binds when correct answers need expert judgment that is scarce, slow or inconsistent.
- Buyer integration: a buyer can absorb the data, and the capability it produces, into a training pipeline or a deployed workflow. It binds when the capability exists but nobody can put it to work.
The stage-by-stage evidence, read on 26 October 2026
As of 26 October 2026, public evidence shows access easing for general training, preparation industrializing, verification tightening and integration still early. The scorecard below the text gives the strongest dated evidence for each stage and the reading it supports.
Access first. Epoch AI's study of human-generated data put the effective stock of quality-adjusted public text at about 300 trillion tokens and projected, with 80 percent confidence, that it would be fully used between 2026 and 2032. Restrictions rose at the same time: the Consent in Crisis audit of 14,000 web domains found that within a single year, 2023 to 2024, more than 5 percent of tokens in the C4 corpus and more than 28 percent of its most actively maintained critical sources became fully restricted. Developers report filling the gap anyway. NVIDIA's Nemotron training FAQ says synthetic data enters both pre-training and post-training, and that quality-filtered and synthetic web crawl made up about 49 percent of Nemotron 3 Ultra's Phase 1 tokens.
Preparation is now a lever that labs pull themselves. The DataComp-LM benchmark reported that a 7B model trained on its filtered DCLM-Baseline set reached 64 percent five-shot accuracy on MMLU with 2.6 trillion training tokens, and that the choice of filtering model alone moved 7B-scale accuracy between 35 and 44 percent. Nemotron's FAQ ranks data and training infrastructure first among what matters and describes an ensemble of three classifiers that sorts web documents into five quality tiers. Preparation of public data is being automated; preparation of private records, such as de-identifying them and linking business context, largely is not.
Verification and integration have thinner public evidence, so their readings rest on fewer sources, as the table shows.
| Stage | Strongest dated evidence | Reading today | Signpost that would change it |
|---|---|---|---|
| Access | Epoch AI: public text fully used 2026 to 2032 on trend; Consent in Crisis: over 28 percent of critical C4 sources restricted in 2023-24 | Easing for general training; binding for private records | A developer reports a training run limited by a lack of usable data |
| Preparation | DataComp-LM: filter choice moved 7B MMLU between 35 and 44 percent; Nemotron: classifier ensemble sorts web text into five tiers | Industrializing for public data; manual for private records | Automated tools de-identify and link business records at scale |
| Verification | SWE-bench Verified: 500 engineer-confirmed problems beside a 2,294-instance full set; GDPval: human expert comparison is the primary metric | Tightening for business and agent tasks | Automated graders match expert agreement on occupational tasks |
| Buyer integration | Census working paper: 18 percent of firms used AI in a business function, November 2025 to January 2026 | Early; caps how much capability turns into spending | Census readings show most firms using AI in business functions |
Why verification looks like the next constraint for business tasks
Because business tasks rarely come with an answer key. A model can be trained on more records, but it improves on a task only when someone can say which outputs are right, and for multi-step business work that judgment is expensive, slow and often disputed.
Benchmarks show the cost. SWE-bench, a software engineering benchmark, lists 2,294 task instances in its full set; SWE-bench Verified, announced on August 13, 2024 as part of a collaboration with OpenAI's Preparedness team, is a 500-problem subset that software engineers confirmed are solvable. GDPval, OpenAI's evaluation of 1,320 tasks across 44 occupations, makes head-to-head comparison by human experts its primary metric because the tasks are hard to grade automatically; its experimental automated grader comes within about 5 percentage points of the agreement human experts reach with one another.
Training shows it too. NVIDIA's Nemotron FAQ says the Nano model capped competitive-coding verification at 50 unit tests to cut verification time, leaving 22,000 tasks, and that prompts the model already solved every time were dropped. Anthropic's engineering guidance on agent evaluations describes combining code-based, model-based and human graders, and running multiple trials because outputs vary from run to run. Each of those choices spends scarce verification on the cases that teach the most.
The implication for investors is that records carrying their own outcomes, such as a resolved ticket with its final disposition, an invoice with its approval trail or a claim with its settled amount, may become worth more than records that show only the input. They let a buyer check a model without paying an expert for every example. That is a hypothesis to test, not an observed price.
Four scenarios for where the bottleneck sits in 2028
The bottleneck in 2028 could sit at any of the four stages, and each scenario shifts value to a different kind of company. This article's base case is grader scarcity for business-task capabilities, with the absorption limit close behind.
The early signals are chosen so that a reader can see which scenario is arriving from public documents rather than from vendor claims.
| Scenario | What binds | What you would observe | Who captures value | Early signal to watch |
|---|---|---|---|---|
| Access wall | Permission and restrictions | Crawl restrictions keep rising and private records stay unlicensed | Holders of rights others cannot get | A repeat restriction audit shows another sharp rise |
| Curation race | Preparation | Gains come mainly from better filtering and synthetic rewriting | Labs and tool vendors that automate curation | Open curation results keep beating larger raw corpora |
| Grader scarcity (base case) | Verification | Human-confirmed subsets and expert graders remain the reported standard | Owners of outcome-labeled records and expert review capacity | Occupational benchmarks keep human comparison as the primary metric |
| Absorption limit | Buyer integration | Capable models outrun firms' ability to deploy them | Companies that connect models to workflows and their records | Census AI use stays a minority of firms while benchmark scores climb |
Six calls on the bottleneck, each with a test that would prove it wrong
These six calls run to the end of 2027 and resolve on public documents. They are this article's forecasts, written so they can fail, and each names the evidence that would prove it wrong.
Two calls test verification, two test access and two test integration and demand. If the two verification calls fail while the integration calls hold, the bottleneck skipped verification and moved straight to buyers.
| Call to the end of 2027 | Resolves on | Proved wrong if | Stage |
|---|---|---|---|
| Human expert comparison stays the primary metric of major occupational benchmarks | The next GDPval-style release from a major lab | Automated grading becomes the headline metric | Verification |
| Agent benchmarks keep reporting human-confirmed subsets | SWE-bench and similar leaderboards | Leaderboards return to unfiltered task sets | Verification |
| Public-text exhaustion does not stop a frontier release | Developer model cards and technical reports | A developer attributes a delayed or smaller run to a lack of usable data | Access |
| Web crawl restrictions keep rising | A repeat of the Consent in Crisis audit | Restricted shares fall below the 2023-24 readings | Access |
| Most firms still do not use AI in a business function | Census Business Trends and Outlook Survey releases | A majority of firms report AI use in a business function | Buyer integration |
| AI data licensing revenue stays concentrated among few buyers | Reddit annual reports and other licensor filings | Filings show contract value spread across many buyers | Buyer integration |
The case against: compute, cheap graders, synthetic data and few buyers
The binding constraint may not be data at all, and investors should weigh that before backing any data stage. Epoch AI's analysis of whether training can keep scaling through 2030 examined power, chip manufacturing, data scarcity and latency, and judged power and chip production more likely than data to bind first.
Improving models and synthetic data could also dissolve the verification thesis. If automated graders close the remaining gap with human experts on tasks like GDPval's, verification stops binding and the bottleneck jumps straight to integration. Nemotron's FAQ shows the synthetic path already carries large shares of training data, though it also says synthetic data is validated by training small models on it, which is itself a verification cost.
Buyer concentration is the third alternative. Reddit's FY2025 annual report attributes almost all contract value tied to its licensing revenue to two partners and calls AI data licensing a novel business model without an established track record. If a few buyers set demand, the bottleneck is their appetite, and no supply-side stage matters much. The risks below would each break the verification-next forecast:
- Compute and power bind first: Epoch AI judges them more likely than data to limit scaling through 2030.
- Automated graders reach expert parity: GDPval's experimental grader is already within about 5 percentage points of expert agreement.
- Synthetic data absorbs volume: about 49 percent of Nemotron 3 Ultra's Phase 1 tokens came from quality-filtered and synthetic crawl.
- Few buyers set demand: Reddit attributes almost all licensing contract value to two partners.
- Access tightens again through law or platform terms, pulling the bottleneck back to the first stage.
- Expert review becomes a cheap service, so verification capacity stops commanding a premium even while it binds.
Matching a data company to the stage that binds next
Ask which stage the company's revenue relieves, what evidence shows that stage binds for its buyers, and what happens to its margin when the bottleneck moves on. A company relieving access in a market where verification binds is selling into yesterday's constraint.
SourceX is a labeled application of the question, not an exception to it. Its public description has it working for companies that license operational records to AI developers, handling sourcing and qualification, rights review and contracting, then data preparation, delivery and payment, which places its work in the access and preparation stages. For buyers it holds no stock: datasets are sourced from U.S. businesses on request, with each buyer stating scope, volume, language, format, timing and licensing needs. The verification-next forecast implies an investor should ask how the records SourceX sources carry outcomes a buyer can verify against. SourceX shares company and commercial metrics with investors directly; it is raising from accredited investors under Rule 506(c), and materials are available on request through its investor page.
Questions that apply to any company in this market, SourceX included:
- Which of the four stages does each revenue line relieve, and how do you know that stage binds for the buyer?
- Do your records carry outcomes, approvals or final states that let a buyer verify a model without an expert?
- What share of deliveries passed the buyer's own acceptance test, and how is that share defined and counted?
- How many distinct buyers account for most of your contract value?
- If synthetic data or automated graders improved sharply, which of your services would buyers still pay for?
Methods note: how the stage scorecard and calls were built
The four-stage model, the scenarios and the calls are editorial frameworks. Evidence comes from primary sources checked on 26 October 2026: two Epoch AI publications, arXiv papers for DataComp-LM, Consent in Crisis and GDPval, NVIDIA's Nemotron training FAQ and the SWE-bench repository on GitHub, Anthropic's engineering post on agent evaluations, a 2026 Census Bureau working paper and Reddit's FY2025 annual report.
A stage's reading reflects the strongest dated public evidence, not a weighted score, and verification and integration rest on fewer sources than access and preparation. Calls were chosen because each resolves on a public document by the end of 2027. The Census figure comes from a working paper whose AI-use measures changed between releases, so it should not be compared directly with earlier Census AI-use rates.
Not measured: market sizes, prices, the cost of verification per task, any company's revenue, or SourceX metrics, none of which are implied. Pages verified through search excerpts should be opened by a reviewer before publication.
Frequently asked questions
What is the next bottleneck for AI training data?
For general model training, raw access to public text is easing because synthetic and curated data stretch it. For business-task capabilities, the likeliest next constraint is verification, knowing whether an output is right, followed by buyers' ability to put new capabilities to work. That is a forecast with dated signposts, not a measurement.
Will AI run out of training data?
Epoch AI projected that the stock of public human text, about 300 trillion tokens, would be fully used between 2026 and 2032 on trend. Developers now use synthetic and filtered data heavily, so exhausting public text does not stop training by itself. Private business records are a separate supply that depends on permission.
Why is verification a data problem rather than a compute problem?
Verification needs correct answers, and for business tasks those come from experts or from records that carry their own outcomes. Compute can run more graders, but it cannot supply the judgment that defines what correct means, which is why GDPval still uses human expert comparison as its primary metric.
What does buyer integration mean for AI data?
It is a buyer's ability to absorb a dataset into training or evaluation and then put the resulting capability to work. It binds when models can do a task but firms cannot deploy them; a 2026 Census working paper found 18 percent of firms used AI in a business function between November 2025 and January 2026.
Does synthetic data end the data bottleneck?
It eases access for general training but moves cost into verification, because synthetic data has to be checked before use. NVIDIA's Nemotron FAQ says its team validates synthetic data by training small models on it, and checking synthetic outputs still depends on real tasks with known answers.
Sources
- Epoch AI estimated the effective stock of quality-adjusted public human text at about 300 trillion tokens and projected, with 80 percent confidence, full use between 2026 and 2032. Source
- Epoch AI's analysis of scaling through 2030 examined power, chip manufacturing, data scarcity and latency, and judged power and chip production more likely than data to bind. Source
- The Consent in Crisis audit of 14,000 web domains found that in 2023-24 restrictions left more than 5 percent of C4 tokens and more than 28 percent of its critical sources fully restricted. Source
- DataComp-LM reported a 7B model reaching 64 percent five-shot MMLU with 2.6 trillion tokens on DCLM-Baseline, and filtering-model choice moving 7B MMLU between 35 and 44 percent. Source
- NVIDIA's Nemotron training FAQ describes synthetic data in pre- and post-training, a five-tier classifier pipeline, a 50-unit-test verification cap for Nano, and data first among what matters. Source
- The SWE-bench repository lists 2,294 instances in the full benchmark and describes SWE-bench Verified, announced August 13, 2024, as 500 problems engineers confirmed are solvable. Source
- The GDPval paper covers 1,320 tasks across 44 occupations, uses human expert comparison as its primary metric, and reports an automated grader within 5 points of human inter-rater agreement. Source
- Anthropic's engineering post on agent evaluations describes combining code-based, model-based and human graders and running multiple trials because outputs vary between runs. Source
- A 2026 Census Bureau working paper found 18 percent of firms used AI in a business function from November 2025 to January 2026, with measures revised from earlier releases. Source
- Reddit's FY2025 Form 10-K says almost all contract value tied to its licensing revenue comes from two partners and calls AI data licensing a novel business model. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.