DeepSeek Processes 155 trillion tokens a Day and Earns $38 per GPU Hour
Until now, outsiders could mostly only guess whether AI services make money. DeepSeek has published its own revenue figures, so anyone can run the numbers with real data.
01What happened
DeepSeek open-sourced its inference engine. It also released several sets of revenue figures at the same time. Put simply, an inference engine is the software that makes a model actually run and answer your questions.
02Key numbers
A single cluster currently processes 155 trillion tokens per day. Revenue is $38 per GPU per hour, calculated at off-peak pricing. DeepSeek itself published these figures.
03In short
A token is a small piece of text that a model reads and writes. A cluster handles this many pieces every day, and each GPU brings in $38 per hour. That is the math behind the business. The $38 figure uses off-peak prices, so it should not be applied directly to every time period.
04Why it matters
Many AI companies talk about how strong their models are. Few open their books. DeepSeek released its revenue figures along with its code. Outside observers now have a real benchmark to compare against.
Liked it? Get the next one by email
Free. Every new post lands in your inbox and the Substack app (posts are written in Chinese first). Unsubscribe in one click. Or subscribe on Substack.