The AI race is a data race

For half a decade, the AI race has been narrated as an architecture race: a better transformer, a cleverer attention variant, a flashier mixture of experts. Dwarkesh Patel's writeup of a controlled decomposition says that story is mostly wrong. From 2019 to 2025, roughly three times as much pretraining efficiency came from data work as from model work, and the two contributions were almost independent. Most of the variance was additive, not interactive, meaning a model improvement and a data improvement mostly just stack.

If data and models are separable levers, whoever owns the cleaner, more diverse, better-curated corpus collects a real industrial rent, while algorithmic novelty travels faster than a Hugging Face release. It also reorders the open-source debate. Open weights close the architecture gap visibly, but if the binding constraint is data, the gap that matters is the data gap, and closed labs with privileged access to high-quality text, code, and reasoning traces hold the durable moat.

The mechanism repeats: identify the data edge, hoard or source it, then let the next architecture wave ride on top. Capital and policy should read this as a data-policy story first and a model-policy story second. The caveat is honest: the result is pretraining-only and tops out at 1e19 FLOPs of training compute, and at frontier scale the ratio could compress or invert. Until it does, data is the load-bearing wall, and the model is the paint.