Fastest - Higher is better
Median tokens per second while the model is generating. Higher is better.
309.1
247.5
172.8
83.5
75.6
55.4
Lambda 2xB200CrusoeNebiusDeepInfraDeepSeekDigitalOcean
STAC benchmark
DeepSeek V4 Pro - providers
Recorded workload
The recorded workload* is a long-context, long-trajectory agentic system in which a DeepSeek LLM iteratively develops, debugs, and evaluates trading models.
- Median input tokens
- 172K
- Peak input tokens
- 334K
- Turns
- 300
Lowest latency - Lower is better
Median time until the first token. Lower is better.
799
1.10
1.76
1.98
2.05
2.25
CrusoeLambda 2xB200DigitalOceanDeepSeekDeepInfraNebius
Highest throughput - Higher is better
Aggregate output tokens per second for the whole run. Higher is better.
231.5
148.4
132.2
63.7
61.6
40.3
Lambda 2xB200CrusoeNebiusDeepSeekDeepInfraDigitalOcean
Request latency - Lower is better
Median time to finish a request. Lower is better.
1.57
2.09
3.63
5.44
5.89
7.63
CrusoeLambda 2xB200DeepInfraDigitalOceanNebiusDeepSeek
Prefill speed - Higher is better
Median prompt processing speed. Higher is better.
84.5k
3.1k
1.1k
1.0k
901.6
403.3
Lambda 2xB200NebiusCrusoeDigitalOceanDeepSeekDeepInfra
Inter-token latency - Lower is better
Median gap between output tokens. Lower is better.
0.06
18.3
20.7
26.0
28.6
37.0
DeepSeekCrusoeNebiusDigitalOceanLambda 2xB200DeepInfra
Comparison
Click a column title to sort. Click a row to expand p90/p99 spreads. Missing values stay at the bottom.
| STAC - Lambda 2xB200 | 309.1 | 1.10 | 2.09 | 231.5 | n/a |
|---|---|---|---|---|---|
| Crusoe | 247.5 | 0.799 | 1.57 | 148.4 | 98.5 |
| Nebius | 172.8 | 2.25 | 5.89 | 132.2 | 92.3 |
| DeepInfra | 83.5 | 2.05 | 3.63 | 61.6 | 96.8 |
| DeepSeek | 75.6 | 1.98 | 7.63 | 63.7 | 97.1 |
| DigitalOcean | 55.4 | 1.76 | 5.44 | 40.3 | 96.2 |
Compare two providers
Click two names. Click again to drop one. A third click replaces the older pick.
Select a second provider to see the spread.
*The recording is contributed by a large US market maker. An agentic coding system (Codex with DeepSeek V4 Pro) iteratively developed, debugged, and evaluated trading models for Bitcoin/USDT market data, producing 300 sequential requests. Input token count is not strictly monotonic across the conversation: it comprises three contiguous segments separated by context-checkpoint compactions, after which accumulated input resets.