DeepSeek says its AI cost $294K to train. Seriously.

Chinese AI firm DeepSeek now claims it trained its reasoning-focused R1 model for just $294,000, disclosing the figure for the first time in a peer-reviewed paper in Nature. The company says training ran for 80 hours on a cluster of 512 Nvidia H800 chips designed for China.

The cost undercuts the nine-figure sums often floated in the U.S. AI race. OpenAI CEO Sam Altman has previously said “foundational model training” cost “much more” than $100 million, without giving detailed breakdowns.

DeepSeek also acknowledged in supplementary materials that it owns Nvidia A100 GPUs and used them in a preparatory phase before the R1 training run on H800s. That transparency comes after months of scrutiny over whether the startup had access to restricted H100 chips despite U.S. export controls. DeepSeek denies using H100s; U.S. officials have alleged the firm found ways to access them via shell companies in Southeast Asia.

The disclosure arrives as DeepSeek basks in rare visibility: R1’s work made the cover of Nature, and the paper outlines a reinforcement-learning-first approach to boost “reasoning” without massive pretraining budgets. That technical angle—plus the headline $294,000 price tag—will rekindle debates over whether smarter training methods can beat sheer compute scale.

Skeptics will zero in on what the figure includes—and what it doesn’t. “Training cost” typically refers to the main compute run, not months of upstream data curation, engineering salaries, or earlier experiments. The 80-hour H800 sprint may be accurate for the final push while still masking a larger all-in R&D bill.

Even so, the signal is clear: if a competitive reasoning model can be brought up for low six figures of direct training spend, incumbents betting only on bigger clusters could face margin pressure—and investors will keep asking whether the biggest budgets are still buying the biggest advantages. That conversation already rattled markets earlier this year when DeepSeek’s low-cost pitch first broke through.