H200 loses money at full utilization while B300 breaks even at 35%: GPU hosting economics for open-source models
AccBalanced · x · 2026-11-03
SemiAnalysis highlighted Yingjun Wu's analysis on whether hosting open-source inference engines is profitable. Running the numbers on DeepSeek V4.1 Flash: the H200 can't cover its rent even at 100% utilization, while the B300 breaks even at roughly 35%.
Key takeaway: hosting open-source models can be profitable, but it isn't simple — which GPU card you rent matters more than which inference engine you run, countering the popular 'hosting open models prints money' narrative on X and LinkedIn.
More from Venture
- Apple says memory price hikes are 'unprecedented,' purchase obligations jump $33B in one quarter — firstadopter · 2026-11-03
- A one-night AI-built project nets 333 points and ~22k visitors — Aizkmusic · 2026-11-03
- Palantir CTO cites Anthropic's S-1: AIG's AI underwriting runs entirely on AIP — eliano · 2026-11-03
- Replit CEO Amjad Masad: next cycle's Anthropic won't look like Anthropic, VCs should back full-stack paradigm over nth LLM lab — beffjezos · 2026-11-03
- The Circuit: Anthropic's S-1, Intel's questions, Marvell and custom silicon bets — BenBajarin · 2026-11-03
- SEO veteran Lily Ray mocks influx of newbies selling 'AI search as a service' webinars — lilyraynyc · 2026-11-03