Why U.S. frontier AI keeps losing the open-source race

创业邦 · wechat · 2026-07-29

Why U.S. frontier AI keeps drifting away from open source

Jason Calacanis’s question — why the U.S. hasn’t produced a leading open model — is used here to argue that the problem is structural, not accidental.

Frontier training costs run into hundreds of millions of dollars, while model “shelf life” is only 12–24 months before a new generation devalues the previous one.
That makes closed APIs the only business model that can plausibly recover costs: companies must charge per token and prevent users from bypassing them.
The article argues that open-source AI is not like Linux: models are expensive both to train and to run, so every query burns real compute and electricity.
Meta is presented as the one major U.S. open-source champion, but its open strategy is framed as a competitive weapon rather than a public-goods project.
After Llama 4 reportedly underperformed, Meta is said to be shifting toward a closed model project codenamed “Avocado”, which the piece treats as proof that closed frontier AI is the only commercially viable path.
It also highlights a practical dead end for U.S. open-weight startups: they are constrained by model terms that forbid distilling from competitors, while well-funded labs and top researchers remain concentrated at OpenAI, Anthropic, and Google.

The conclusion is blunt: in AI, the economics of training and inference collide with the logic of open source, and someone has to keep paying the bill.

Original post →

Why U.S. frontier AI keeps losing the open-source race

More from AGI Musings