DeepSeek Is Free Now but What’s the Catch?

DeepSeek open-sourced DeepInfer and says one cluster serves 155 trillion tokens a day at $38 per GPU-hour. Almost none of those tokens are computed, and the part that skips computing them is the part DeepSeek kept. Every take this week ran the same math: $38 of revenue against $3 to $18 an hour to rent an Nvidia B300, so an 80% margin. The math holds, but it misses the story. DeepSeek's own benchmark says one GPU writes about 26,300 tokens a second, so 155 trillion a day would take roughly 68,000 GPUs and $22.7 billion a year. DeepSeek's revenue run rate is $1 billion. The gap is the KV cache: