Open-source models shift profit back to GPU and RAM rents, the thread argues
davidmanheim · x · 2026-07-27
The thread argues that hardware providers do not just profit from open-source models—they also benefit from faster model iteration. Frontier AI labs can pressure cloud vendors on price, switch providers, or even build their own chips, but open models and repeated leapfrogging push more value back into GPU and RAM rents.
It also says the DeepSeek panic faded for a similar reason: more efficient open models lower compute per task, which is bad for hardware in the short run, but the long-run bet is a Jevons-paradox effect where cheaper AI usage drives much larger total demand. If the lab also owns the hardware stack, as with Google/DeepMind, it may prefer proprietary models because it captures both the model margin and the infrastructure margin.
More from Infra
- LightRAG v1.5.3 adds a one-time Milvus migration and hardens production edge cases — JeremyCMorgan · 2026-07-28
- llama.cpp adds DSpark speculative decoding and asks for speed results — pmttyji · 2026-07-28
- Nvidia puts an open vision-language-action model on Hugging Face — theteknosaur · 2026-07-28
- “Model eats harness” is really about deployment-driven harness evolution — m_wulfmeier · 2026-07-28
- Agentic AI system reportedly cost $1.2M a month before being cut to $100K — emmanuelvivier · 2026-07-28
- Wistron’s Early Nvidia Bet Turned It Into One of AI’s Biggest Supply-Chain Winners — pstAsiatech · 2026-07-28