DeepSeek 开源推理引擎 DeepInfer 并披露成本数据

Aakash Gupta @aakashgupta原推发布于 2026.10.21

在 X 上打开原推

For one dollar, DeepSeek's AI will write you more words than all seven Harry Potter books put together. And DeepSeek just revealed it still pulls in serious money at that price. Each chip in one of its clusters brings in about $38 an hour, and renting that kind of chip runs about $8. Both of those numbers use DeepSeek's off-peak prices, the half-price hours. They come from a document DeepSeek posted today alongside DeepInfer, the software it uses to run its own AI, which it just made free for anyone to download. That same cluster handles 155 trillion tokens a day. Back in May, Google said every product it runs combined was doing about 107 trillion a day. Google's number has surely grown since then, but we're comparing one DeepSeek cluster to all of Google. So how does a company charging pennies pull in almost 5x what the chip costs to rent? It remembers. Every time you send an AI a message, the model re-reads the whole conversation from the top. A coding agent might re-read an entire codebase on every single step, and that re-reading is the expensive part. DeepSeek saves that work to disk and holds onto it for 3 days on agent traffic and 7 days on chatbot traffic, on about 100 petabytes of SSDs that do nothing else. Most AI companies keep that stuff around for minutes or hours. When DeepSeek already has your context saved, reusing it costs $0.003 per million tokens versus $0.15 to process it fresh, and the GPU gets to spend its time writing new answers, which is the thing people are paying for. Then read the fine print on what they released. The engine and the kernels are in there. But DeepSeek's own README lists three things you need for a full service that aren't included, and two of them are the memory system and the load balancer that keep those multi-day memories working. DeepSeek says that code is too wired into its internal systems to release, so it wrote up an explanation of how it works instead. You also can't run this in your garage. DeepSeek says the smallest setup that makes sense is 128 GPUs, 64 to read prompts and 64 to write answers. At rental prices that's about $1,000 an hour before a single customer shows up. So the real audience here is cloud companies. DeepSeek shipped a second tool, deepseek-recipe, that makes anyone else's setup respond exactly like DeepSeek's official API. Tianyi Cui from DeepSeek wrote that the hope is for companies all over the world to deploy DeepSeek models and make money doing it. That $38 figure is a sales pitch to every cloud sitting on a rack of Blackwells. If enough of them bite, DeepSeek's model ends up running in data centers around the world that DeepSeek never had to build, and the code behind the $38 is still sitting in Hangzhou.

这条是引用着另一条写的：「DeepSeek 开源推理框架 DeepInfer，支撑 V4.1 Flash 生产服务」