"It should make you angry": DeepSeek's docs show $38 of revenue per GPU-hour, and X ran the margin math at 82%

"DeepSeek just published the price of a token, and it should make you angry." A post by Alexey Fateev opening with that line passed 4,100 likes and 190,000 views on X. It picked apart the numbers in a document DeepSeek published alongside its DeepInfer inference engine.

Per DeepSeek's deployment Q&A on GitHub, a single cluster serving DeepSeek V4.1 Flash handles 155 trillion tokens a day, "with overall revenue of US$38 per GPU-hour (at valley pricing)." The document says the system keeps KV cache for more than 72 hours to avoid redoing prefill on cache hits.

Citing math by Yuchen Jin, Fateev argued that even renting Nvidia B300s at $7 an hour works out to roughly an 82% gross margin. "Now remember every time a US lab told you that intelligence is expensive," he wrote, calling part of $200-a-month pricing "a business model built on you not knowing what a token actually costs."

The vLLM project said its open-source engine can match DeepInfer's decode performance within 6% and generate $39 per GPU-hour at DeepSeek's off-peak pricing. DeepSeek has not disclosed its costs, and the 82% figure is an outside estimate.