Open-source models account for just ~3.2% of global AI training spend, analysis finds
0xBekket · reddit · 2026-09-12
The author parsed Hugging Face releases (base models vs adapters/fine-tunes), estimated training FLOPs with the 2×3×N×D formula, and priced them at H100 rental rates: open-source models represent only 3.2% of global annual AI training spend ($250M each for LLMs and diffusion models, $56M for adapters). Training a 70B model from scratch costs $1.6-2.1M on cloud vs $30M for owned hardware — a 15x premium — implying breakeven in 21+ months at 100% utilization, longer than hardware stays relevant (compared to Bitcoin ASICs that paid back in 3 months). He asks whether open-model trainers rent or build, and shares he's building a P2P GPU marketplace with crowdfunding; a friend argues the $600M market is already monopolized by big labs building their own datacenters.
More from Venture
- One-person companies are about the org chart, not headcount: humans judge, agents execute — ciguleva · 2026-10-28
- SemiAnalysis: OpenAI reportedly negotiating ~1GW TPU commitment while seeking sites — dylan522p · 2026-10-28
- Big Short's Eisman: the whole market is now 'one trade' on AI — and it 'better succeed' — SumitGup · 2026-10-28
- Directory tracks 174 AI training-data vendors as Anthropic's data costs jump $251M — rbhar90 · 2026-10-28
- SK Hynix says memory demand will outpace supply through 2030 — Chemical_Hawk_6307 · 2026-10-28
- SK Hynix posts $62.8B quarterly operating profit at 80% margin as AI memory demand booms — mhdfaran · 2026-10-28