Gavin Baker: post-trained open models plus router deliver frontier performance at 40% cost
GavinSBaker · x · 2026-10-14
Investor Gavin Baker lays out a hybrid multi-model thesis:
- Open-weight models are increasingly post-trained on enterprise private data by companies like Fireworks AI, producing highly optimized specialized models
- Combined with a frontier model behind a router, they deliver frontier-level performance at 40% of the cost with the same underlying compute
- Cursor's Composer, Cognition's SWE-2, and Harvey's Tenet are early examples; every large enterprise will eventually post-train its own models for data sovereignty
- He warns an OpenAI/Anthropic model-layer oligopoly would be negative for nearly every company — a monopsony over hyperscalers and the compute supply chain, with eventual vertical integration
- "Own your weights, control your future" — a view echoed by Satya Nadella and Alex Karp
More from Companies & People
- MIT: Universities Must Rethink Credit and Infrastructure for the AI Era — nordicinst · 2026-10-14
- OpenAI says hidden recording device found in office, employee arrested by FBI — OpenAINewsroom · 2026-10-14
- OpenAI and Anthropic tell UK lawmakers to regulate them; DeepMind and Meta decline — nordicinst · 2026-10-14
- Stanford HAI names 15 new Data Science Scholars across nine disciplines — StanfordHAI · 2026-10-14
- China's GigaAI Opens First Provincial-Level General World Model Lab in Beijing — pstAsiatech · 2026-10-14
- After reading ~150 investor updates, YC alum shares what the best startups do post-batch — ycombinator · 2026-10-14