Running DeepSeek R1 on H100
Quick answer
DeepSeek R1 sustains 74.2 tokens/s per GPU on H100 at 50 tokens/s per user, which works out to $4.38 per million tokens at hyperscaler pricing, served by Dynamo SGLang. Fastest measured TTFT: 0.2 ms; fastest TPOT: 0.0 ms (each the best across all configs, not one run).
Benchmarked configs
72
Serving engines
Dynamo SGLang, Dynamo TRTLLM
Precisions
fp8
Run dates
2026-02-08 → 2026-02-13
Throughput at every interactivity target
Serving is a trade-off: push more concurrent users through a GPU and each user's tokens arrive slower. The ladder below reads the measured frontier at each per-user speed target on a single-turn chat workload (8k input / 1k output), using the best engine and precision at that point.
| Per-user target | Tokens/s per GPU | $ / 1M tokens | Engine | Precision |
|---|---|---|---|---|
| 30 tok/s | 116 | $2.81 | Dynamo SGLang | fp8 |
| 50 tok/s | 74.2 | $4.38 | Dynamo SGLang | fp8 |
| 75 tok/s | 49.5 | $6.57 | Dynamo SGLang | fp8 |
| 100 tok/s | 25.7 | $12.64 | Dynamo SGLang | fp8 |
| 150 tok/s | - | - | Dynamo SGLang | fp8 |
| 200 tok/s | - | - | Dynamo SGLang | fp8 |
What serving actually costs
Converting the 50 tokens/s per user operating point to $ per million total tokens across rental pricing tiers from the SemiAnalysis AI Cloud TCO model.
| Pricing tier | $ / GPU / hr | $ / 1M tokens |
|---|---|---|
| Hyperscaler | $1.17 | $4.38 |
| Neocloud | $1.55 | $5.81 |
| Retail | $1.78 | $6.67 |
Frequently asked questions
- How fast is DeepSeek R1 on H100?
- At an interactivity target of 50 tokens/s per user on a single-turn chat workload (8k input / 1k output), H100 sustains 74.2 tokens/s per GPU serving DeepSeek R1 with Dynamo SGLang in FP8. Peak measured throughput across all configs is 1,376 tokens/s per GPU.
- How much does it cost to serve DeepSeek R1 on H100?
- $4.38 per million total tokens at hyperscaler $/GPU/hr pricing, at 50 tokens/s per user. Neocloud and retail rental tiers are tabulated above; slower interactivity targets lower the cost further.
- Which serving engines run DeepSeek R1 on H100?
- The runs behind this page used Dynamo SGLang, Dynamo TRTLLM in FP8, including disaggregated prefill and multi-node serving. Engines are rebuilt and re-benchmarked continuously, so the best config can change between visits.
- How are these DeepSeek R1 numbers measured?
- Every number is measured on real H100 hardware by the InferenceX fleet, sweeping concurrency on a single-turn chat workload (8k input / 1k output) to trace the throughput-versus-interactivity frontier; the newest run landed on 2026-02-13. The same derivation powers the InferenceX overview leaderboard.