DeepSeek open-sources its serving engine and reveals $38 of revenue per GPU-hour, an anchor for cheap Chinese tokens
What happened? DeepSeek has open-sourced DeepInfer, the inference framework that runs its own production API, under an Apache 2.0 licence, along with deepseek-recipe, the token encoding layer behind its official API. The release on 20 October came three days after the Hangzhou lab published the weights of DeepSeek V4.1 Flash.
The numbers nobody else publishes: In its announcement, DeepSeek disclosed that a single DeepInfer cluster serves 155 trillion tokens a day and earns about $38 of revenue per GPU per hour, measured at its off-peak prices. It also reported a 96% input cache-hit rate, helped by keeping cached conversation history for three days for agent traffic and seven days for chatbots on a storage pool of roughly 100 PiB. These are company-reported figures for one cluster, not audited averages.
Why it matters: Western labs such as OpenAI and Anthropic do not disclose per-GPU serving economics. DeepSeek's figure gives the market a public price anchor. One outside researcher, cited by BigGo Finance, estimated hardware-level inference gross margins of roughly 84% to 96%, a calculation that leaves out training, R&D, bandwidth and staff costs.
Nvidia and Huawei, side by side: DeepInfer runs on both Nvidia's Blackwell GPUs and HuaweiDealroom has a profile for this one. Try Dealroom →'s Ascend 950 chips. That matters as Beijing pushes its labs towards domestic silicon and US agencies move to strip DeepSeek out of federal systems.
The bigger picture: Investor and author Nicolas Colin argued on X that China will flood the world with cheap tokens. If open weights, an open serving stack and a published cost base become the norm, they put pressure on the API pricing of Western model labs and on the neoclouds that are borrowing heavily to rent out GPUs.
What's missing? The open-source release only covers offline inference. The KV-cache system, prefill/decode disaggregation and load balancing that make the production numbers possible stay in-house.
Sources: 36Kr / Wallstreetcn, MarkTechPost, Latent Space, BigGo Finance
