DeepSeek-V3/R1 inference system overview ('one more thing')
DeepSeek discloses serving stats of 608B input and 168B output tokens a day on about 227 H800 nodes, and a theoretical 545% margin at R1 prices.
Describes cross-node expert parallelism (EP32 prefill, EP144 decode), dual-batch overlap and load balancing. Daily cost at $2 per H800-hour was $87,072 against $562,027 of theoretical revenue if every token were billed at R1 list price; real revenue is far lower.
- Date
- Saturday, 1 March 2025
- Lab
- DeepSeek
- Kind
- infra
- Access
- research preview
Figures
| Measure | Value | Measured by |
|---|---|---|
| Theoretical cost profit margin | 545% If all tokens billed at R1 prices; actual revenue lower due to free tiers, V3 pricing and night discounts | company |
| Tokens per day / cache hit rate | 608B in (56.3% hit), 168B out 24 hours ending 2025-02-28 12:00 UTC+8; 226.75 average nodes of 8 H800 | company |
Commit 'day6: one more thing' is dated 2025-03-01 04:03 UTC. The 545% figure is a theoretical, company-calculated margin on one measured day that omits R&D, training and non-peak revenue, so it is not a reported profit.
Sources
This record was partly confirmed: some claims could not be checked on 6 October 2026. How we check