H200 loses money even fully busy: GPU choice beats inference engine for hosting economics
teortaxesTex · x · 2026-11-03
Yingjun Wu's analysis of hosting DeepSeek V4.1 Flash shows GPU choice matters far more than the inference engine: H200 (8 GPUs) earns only $1.6 per GPU-hour at full load and needs 244% utilization to break even — unprofitable at any utilization. B300 (8 GPUs) earns $21.4 and breaks even at 35% utilization (31%/51% gross margin at 50%/70% busy). GB300 NVL72 ranges from $18.1 full-load revenue (84% break-even) standard, to $32.0 with NVIDIA Dynamo (48% break-even), to $64.7 with rack-tuned vLLM (24% break-even, 66% margin at 70% busy). teortaxesTex adds: to profit on Hopper you must own compute — rental margins price out resellers.
More from Infra
- ByteDance crawler spotted scraping websites at massive scale — gefei55 · 2026-11-03
- 1e29 FLOPS would take 200k GB200s 3.2 years and 6.2 TWh, dev notes — mike64_t · 2026-11-03
- Fine-tuned open-source models beat frontier models, cut costs 95% — pdamodaran · 2026-11-03
- Open-source trainer beats ms-swift by 58%, hits 12,523 tokens/s on one GPU — incorporo · 2026-11-03
- OpenAI's Grid Protectors program lets utilities cut up to 1,000MW from its Georgia data center — rohanpaul_ai · 2026-11-03
- Nearly $1 trillion erased from US stock market as tech selloff deepens — IanArawjo · 2026-11-03