DeepSeek, the Hangzhou-based lab whose V3 release shipped at a fraction of frontier training costs, says a single cluster serving its V4.1 Flash model now handles 155 trillion tokens per day. The company puts overall revenue at $38 for each card, each hour, calculated on what it calls “valley” pricing, a term the post does not define. Both figures come from DeepSeek’s own write-up on Zhihu, published 20 October 2026, and no outside party has checked them.
The revenue number is a claim, and an unusually specific one. US labs rarely publish what a GPU-hour earns them, so there is nothing public to set it against. What the post does give is the engineering that DeepSeek says produces it, and that is where the useful detail sits.
Start with the hardware. DeepSeek splits each request into stages that run on different machines. Prefill, which reads the prompt, runs on ordinary 8-card servers. Decode, which writes the answer, runs on 64 to 128 cards grouped as one unit. A separate vision stage handles images and, the post suggests, can run on a single card, ideally the cheapest one available. The framework targets chips from both NVIDIA and Huawei (the Ascend line), so the economics are not tied to one supplier.
Caching carries much of the economic weight. DeepSeek keeps the stored attention data from earlier turns so that a returning conversation skips most of the reading work. It holds that data on a separate storage cluster of roughly 100 PiB of SSD space, with a retention time of three days for agent traffic and seven days for chatbot traffic. Agent sessions, which resend long histories, are the workload this is built for. The post does not publish an overall cache hit rate for the main store, so the savings are described but not measured in public.
One design choice stands out as a reversal. The earlier V4 setup saved a short-lived “sliding window” cache to disk as well, and it ate close to half of the persistent capacity while delivering little benefit, because that data is only useful for minutes. V4.1 drops it. When that window is missing, the system recomputes only the last 128 tokens, an approximation that DeepSeek says cuts total prefill computation nearly in half in the decoder layers. Where it does keep the window, in server memory, the post reports a lifetime of about 180 seconds and a hit rate above 98 percent.
Guessing ahead is the third lever. DeepSeek’s DSpark method has a small draft model propose five tokens at a time, then spends verification effort only on the guesses most likely to be accepted. Against checking all five every time, the post reports a 6 percent throughput gain in live traffic. That is modest, but it stacks on top of everything else.
The post also covers load balancing across cards, with measured imbalance between the slowest and average expert groups of 1.09 in decode. Automatic scaling shifts machines between prefill and decode as traffic moves, and it refuses to shrink the fleet when quota use passes 0.8. Faults are handled by keeping spare capacity and rerouting requests, not by restarting faster, since restarts take minutes.
A separate AI Insiders article covers the open-source DeepInfer repository. DeepSeek notes that its cache management and load-balancing code is tightly bound to internal infrastructure and has not been released, so outsiders get the engine but not the parts that, by DeepSeek’s own account, drive the cost savings.
For anyone pricing inference against DeepSeek’s API, the practical comparison is cache retention: a three-day window on agent traffic is the number to ask other providers about, because it decides how much of a long agent session you pay to reread.
Reported by DeepSeek, in its own technical post on Zhihu, on 20 October 2026.
