Unverified50% confidenceFactExact time
Inference at large model companies continuously consumes GPU compute, and token generation has a real marginal cost, which differs from the situation at traditional SaaS companies where gross margins often reach 70%-80% or more.
1
Sources
50%
Confidence
Long-term
Relevance
10/20/2026
First Seen
Sources
Related Entities
Related Claims
Unverified单条推理通常只能占用GPU SM利用率的5%～15%，而合理的批次大小可将利用率提升至60%～90%72% similarVerifiedDuring large model inference, actual GPU utilization is often less than 10%.71% similarUnverified许多推理服务的GPU利用率不足30%，通过VGPU技术可将利用率提升至70%以上70% similarUnverifiedSome enterprises purchase GPUs ahead of schedule to seize the AI initiative, but actual compute utilization may remain below 50% for a long time.70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1078491API
curl https://kongchang.com/api/v1/knowledge/claims/1078491MCP
get_claim(id=1078491)