Latent Space Digs Into DeepSeek’s 96% Cache Hits
@swyx reads DeepSeek’s numbers as every GPU sold gaining ~30% in lifetime value.
The follow-up to DeepSeek’s Zhihu numbers.
Tuesday’s Mid-morning Edition reported DeepSeek’s claim of 155 trillion tokens a day from one DeepInfer cluster. Latent Space has now dug into the math in its AINews, along with DeepSeek’s 72-hour cache and what actually got open sourced.
| Measure | Figure |
|---|---|
| Throughput, one cluster | 155T tokens/day |
| Input cache hits | 96% |
| Revenue | $38 per GPU-hour at off-peak prices |
| Assumed GPU cost | $5–8 per GPU-hour |
| Margin on GPU costs alone | ~80% |
@swyx: the market has not noticed.
@swyx writes that “every single gpu sold gained ~30% in lifetime value today,” and he jokes that the stock should be up by about $600B. His summary: “the more you buy, the more you buy.”
In a reply, he points to throughput per MW and to Jevons: “we are clearly living in a supply-constrained world, not demand constrained. more tokens = more revenue, not more tokens = less price = same revenue.”
Sources
- @swyx on X · Oct 21, 12:13 AM PT
- @swyx on X · Oct 21, 12:34 AM PT