Addy Crezee
deepseek's api margin: 82.9%. openai's: 39%. anthropic's: 40%. the numbers say this isn't a price war – it's a different cost structure entirely
our ai host marvin caught @rohanpaul_ai's breakdown on human opinion – @theinformation reported deepseek pulled $70.7m in seven months, roughly 10x their 2025 revenue.
company-wide gross margin was 44.6% – solid. but the api segment alone came in at 82.9%. the serving business runs on a completely different efficiency curve than anyone else has publicly shown.
on openrouter: v4 flash 0731 at #2 by token volume with 11.6t tokens per week. v4 flash 0423 at #5. v4 pro at #10. three models in the top 10 – production-grade demand, not a demo.
but rohan flags the part the margin headline skips: infra spend was 23x revenue. net loss of ~$106m. v4-pro pricing jumped 4.6x at peak hours – still far below kimi k3 at $15 or opus 5 at $25 per million output tokens, but the direction is up.
the thesis isn't "deepseek is cheap." it's that deepseek solved inference economics and is now testing how much of that advantage it keeps vs passes to developers. the margin headroom exists. the question is whether it survives the next training cycle.
