Companies
DeepSeek releases its AI engine, and Latent Space estimates an 80% margin
DeepSeek has made public the system it uses to answer its users and has shared usage figures. A newsletter uses them to calculate that it makes a lot of money, but no one has audited them.

The essentials
- DeepSeek has released DeepInfer as open source. It is the system that runs its V4.1 Flash model, though the company is holding back some important parts (TLDR AI, Latent Space)
- Claimed by one partyAccording to DeepSeek, a single one of its server clusters processes 155 trillion tokens (fragments of text) a day and brings in $38 per chip per hour (TLDR AI, Latent Space)
- Claimed by one partyThe Latent Space newsletter uses those figures to calculate a margin of more than 80%. This is its own estimate, not a figure from DeepSeek (Latent Space)
Why it matters
Until now, the best tools for serving AI at large scale were usually closed. If Latent Space's numbers are close to reality, answering users could be very profitable. But that calculation depends heavily on which costs are included (Latent Space).
The details
The team behind vLLM, the open program DeepInfer is based on, has published its own tests. It says that on the same machines, its system generates text at a speed only 6% below the one DeepSeek reports (Latent Space). DeepSeek has also not yet released two parts: the one that stores text it has already read and the one that splits the work across machines (Latent Space).
What we don't know
- Whether anyone independent has checked the usage and revenue figures DeepSeek has given
- How the margin changes if all costs are counted, and with uses that don't repeat as much text already read
- When DeepSeek will release the parts it has not yet made public
Sources
- DeepSeek DeepInfer 🐋, PyTorch on TPUs ⚡, Ramp’s agent framework 👨💻
- [AINews] DeepSeek DeepInfer: Open Source SOTA Inference running 155T toks/day/cluster making $22m/day at ~80% margin for a 81 day payback
Was this story useful? Yes Not really