SemiAnalysis (@SemiAnalysis_) on X

What it claims

SemiAnalysis argues that stock vLLM, with Wide-EP (wide expert parallelism) and EPLB (expert-parallel load balancing), reproduces DeepSeek's serving performance. vLLM matches the decode performance of DeepSeek's DeepInfer release within 6%. At DeepSeek's off-peak pricing, that equals $39 per GPU-hour in revenue, measured at 200 P90 tokens per second on the AgentX workload. The thread also gives a GB300 estimate: assuming 60% utilization, GB300 makes $50 per GPU-hour in profit at 125 TPS when priced at OpenRouter rates. SemiAnalysis sets this against DeepSeek's $39 per GPU-hour off-peak.

Why DeepInfer is not benchmarked

DeepInfer is not on SemiAnalysis's InferenceX benchmark because the release supports only offline, single-turn inference. A companion project, deepseek-recipe, adds API infrastructure. SemiAnalysis says it still lacks the "key load bearing components" for online serving. SemiAnalysis declined to build those components itself because doing so would add too many assumptions to the comparison. As a result, the 6% figure covers decode performance only. It is not a like-for-like online serving benchmark of DeepInfer.

Limits

The fetched text shows only parts 1, 2 and 5 of a five-part thread, and the main post is cut off at "Specifically:". It does not give:

• the vLLM version or flags;
• the GPU type for the 6% comparison;
• the model checkpoint;
• the AgentX workload definition;
• which serving components deepseek-recipe is missing.

The two dollar figures measure different things. The $39 figure is called revenue, and the $50 GB300 figure is called profit. They also use different throughputs (200 TPS against 125 TPS), and only the GB300 figure states a utilization assumption (60%). As stated, they are not directly comparable.