A Stargate for Data

There's a new bottleneck in AI, and it isn't GPUs. It's data. For years the industry lived in a compute-limited world: infinite internet text, never enough chips to chew through it. That era is ending. Public human text tops out around 300 trillion tokens, and even the reinforcement-learning trick of using gradable math and coding tasks is starting to run dry. Compute keeps scaling smoothly into the trillions of dollars because everyone can just buy more chips. Data can't scale that way, because most of what's useful was never digitized in the first place — it's locked in private archives, undocumented workflows, or simply in people's heads.

The internet, as one essay from a data-industry insider puts it, was a one-time civilizational subsidy: decades of free, unintentionally curated human output that happened to be perfect training material. That subsidy is spent. So labs are pivoting to buying it. Total third-party data spend already runs around $7 billion a year, and the argument is that figure could top $100 billion annually by 2030 as labs pour compute-scale budgets into human-generated data, licensing deals, and projects like Anthropic's book-scanning effort.

This shift changes who wins. Compute is a commodity — everyone buys the same Nvidia chips and builds similar clusters. Data isn't. Frontier models have looked similar because they all trained on roughly the same internet; as labs build proprietary, hand-collected datasets, that convergence breaks down. OpenAI's edge in math and Anthropic's in cybersecurity, the argument goes, trace back to which lab bothered to collect the right specialized data first, not some secret architectural sauce.

That's also why data-labeling companies look like the last ones standing in an AGI race. Mercor, just three years old, is reportedly doing something like $2 billion in revenue running armies of expert contractors. The logic: by the time frontier labs stop needing human-generated data, AGI has basically already arrived, and everything else in the economy has been upended anyway. Marginal datasets don't get cheaper as models improve — they get more valuable, because closing that last 1% gap between a model that mostly does a job and one that fully does it is worth a fortune.

The bigger claim here is geopolitical as much as technical. If economic and scientific progress is now rate-limited by data coverage rather than raw compute, then data collection deserves the kind of national mobilization the U.S. gave to compute with the Stargate cluster buildout. And a state with more centralized authority to marshal data — China gets named explicitly — could theoretically out-collect more market-driven rivals, compounding an advantage that has nothing to do with who has more GPUs.