Frontier AI labs have spent the last two years racing to build gigawatt-scale compute clusters. Will Depue, the former OpenAI researcher known for shipping early agentic products, argues in a post on X published July 6 that the next scarce input isn’t chips at all. It’s data, and the industry has not organized itself to collect it at anywhere near the scale it needs.
Depue’s core claim is structural. Model performance has scaled with two inputs in tandem: model size and training data volume. Compute capacity has kept compounding, with trillions of dollars flowing into cluster buildout and no clear ceiling in sight. Public data has not kept pace. Depue puts the usable stock of public human text at roughly 300 trillion tokens, and says the open web no longer generates enough fresh high-quality material to match what scaling demands.
The dollar figures Depue cites are worth separating from his forecast. He states that vendor billing for data, excluding what labs spend on internal collection, already runs roughly $7 billion a year. His projection of more than $100 billion a year by 2030, over 10x growth, is his own extrapolation from that baseline, not a disclosed roadmap from OpenAI, Anthropic, or any other lab. Readers should treat it as an informed bet from someone close to the labs, not a confirmed budget line.
Depue points to Anthropic’s book-scanning project as an example of labs already building proprietary data pipelines rather than relying on scraped text. He also credits divergent, manually collected training corpora, not architecture, with explaining why OpenAI and Anthropic have pulled ahead in different specialties: math for one, cybersecurity for the other, in his framing. If that holds, data sourcing strategy becomes as consequential to a lab’s roadmap as its chip supply contracts.
The market angle Depue only gestures at is where this gets interesting for operators. He names Mercor, a data-labeling company founded three years ago, citing rumors that peg its revenue at roughly $2 billion with a labeler workforce numbering in the millions, as an early winner of this shift. Scale AI has run a similar playbook for years, supplying labeled and reinforcement-learning data to multiple labs simultaneously. If Depue is right that demand for niche, expert-generated data keeps rising even as demand for generic labeling flattens, the value transfers toward vendors who can recruit narrow domain expertise (radiologists, tax attorneys, senior engineers) rather than generalist labeling workforces.
The comparison to the compute buildout has a limit Depue does not fully address. Gigawatts and GPUs are fungible: a cluster announcement can be verified against chip shipments and utility interconnection filings. Data cannot be audited the same way. Much of what Depue describes, private licensing deals, “confidential training” arrangements, companies persuaded to turn off deletion policies, happens in bilateral contracts with no public disclosure standard. A “Stargate for data” would be a negotiation with hospitals, courts, and corporations over access rights, not a capital commitment to a fab.
Depue’s own piece contains the strongest source of skepticism about his thesis. He acknowledges that data quality control “can vary massively” and that the field has not ruled out synthetic data or more data-efficient architectures closing the gap without brute-force collection spend. If either materializes at scale, the 10x spending curve he projects overstates what labs will actually need to pay for.
For operators, the practical takeaway is to stop evaluating AI data vendors purely on labeler headcount or revenue growth. The vendors likely to matter in 2030 are the ones building defensible access to expert, hard-to-replicate data domains now, before that access becomes the thing labs are willing to overpay for.
Based on a post by Will Depue (on X), published July 6, 2026.
