The Information reports DeepSeek generated $70.7M in 7 months while its API gross margin reached 82.9%.

that revenue was roughly 10 times DeepSeek’s entire 2025 total.

Company wide gross margin was 44.6%. Profitability stayed high because AI infrastructure optimization reduced the chips needed to run models while keeping inference costs low.

That infrastructure spending was roughly 23 times revenue, although it includes rented servers plus purchased chips and equipment rather than only operating costs.

---
IMO, the $70.7M 7-month total understates how fast DeepSeek was reportedly growing by midyear. A $400M-$500M annualized run rate implies roughly $33M-$42M of monthly revenue, vs. about $10M averaged from January through July.

Overall, as compared to OpenAI/Anthropic, DeepSeek appears able to serve API traffic cheaply enough to preserve substantial gross profit at prices far below many premium models.

The V4-Pro increase from $0.87 to $3.96 per 1M output tokens at peak hours is a 4.6x price increase, yet the cited price remains far below Kimi K3 at $15 and Claude Opus 5 at $25. DeepSeek is testing how much of its cost advantage it can keep instead of passing it through to developers.

DeepSeek can run high-gross-margin API serving and still require heavy external capital. AI infrastructure spending through July was about 23x revenue and more than 9x its 2025 infrastructure spend.
The API-margin figure does not capture those fixed and expansion costs. Utilization and pricing now have to rise fast enough to absorb the compute buildout before the next model cycle requires another large spending increase.