Cao vs Yangyi: private deployment of open-source models is a real business — data, not models, holds the value

yangyi · x · 2026-09-28

In a X thread, @caozlog argues private deployment of open-source models is becoming a hot trend: rising data-security/compliance needs, open models past the usability threshold (he tested Zhipu's coding model), and controllable enterprise token costs — with caveats on iteration costs, low-frequency usage economics, and deployment reliability. He also flags 'sovereign AI' as a third-party opportunity for non-US/China countries, where Chinese open models lead.

@yangyi pushes back: models will be localized, but open models themselves won't be valuable — the value lies in data-driven corrections that encode human-AI feedback, which commands a premium. Frontier models solving future-era problems won't get cheaper. He agrees open-source deployment is a viable business (power-to-token arbitrage, data resale, cheap access, cloud transformation) but requires resources.

Related event: Debate over private deployment of open-source models(2 posts)→

Original post →

More from Venture

Venture channel →