In October 2022, Sam Altman told OpenAI's board he wanted to ship an open language model "soon, before Stability or someone else does." The pitch was not generosity.
In general, we think this helps discourage others from releasing similarly-powerful models, and makes it harder for new efforts to get funded.Simon Willison
The email surfaced in Musk v. Altman discovery and hit Simon Willison's link blog Monday morning, landing in the middle of the argument it settles. Monday's top essay on Hacker News — 900 points and climbing — argues that America's locked-down AI is losing to China's open-weights strategy. Ben Thompson proposed that Washington declare training-data collection fair use and bar the contract terms that forbid distillation. Alibaba shipped Qwen 3.8 Max as open weights — a reversal from May, when it kept Qwen 3.7 Max closed — two days after Xi Jinping urged China to "seize this rare, historic opportunity to encourage open source." And Moonshot's Kimi K3 stopped taking new subscribers because demand outran capacity.
The debate keeps getting framed as values — safety against freedom, control against openness. That is the wrong ledger. Openness in AI is a pricing decision: you open the layer where your competitor makes money and you don't. Altman wrote it down four years ago — an open release as a way to deny oxygen to whoever comes next. Xi's speech is the same move from the other side of the board.
Openness in AI is a pricing decision: you open the layer where your competitor makes money and you don't.
Tyler Cowen's read is the compact version: commoditize your complements. Xi ties open source explicitly to AI "moving from the digital world into the physical world" — factories, robots, devices — where China's advantages compound and the model is an input, not the product. The werd.io essay supplies the mechanism: the model layer has almost no moat beyond brand and switching costs, and US export controls already prevent Chinese firms from selling global-scale centralized AI services. Giving the weights away converts a compute disadvantage into a distribution advantage — by Martin Casado's estimate, any given startup is 80 percent likely to be running Chinese models already. For OpenAI and Anthropic the same math runs in reverse. The model is the product; the valuations are the capitalized value of keeping it closed. They cannot commoditize their own revenue line, so they defend it — with terms of service.
The standing answer is the frontier premium: closed American models stay far enough ahead that position doesn't matter, and capability is the moat. It has held so far, and it is the bet every closed-lab dollar is priced on. But Monday's tape cuts against it. Qwen 3.8 Max is a 2.4-trillion-parameter open release from a company that made the opposite call two months ago. Kimi K3 spent the day turning customers away. And Thompson's proposal exists because the labs' anti-distillation clauses are unenforceable — stopping distillation, he notes, is "nearly impossible" when it is "literally just querying the API." When the defense of your product layer is a contract term you cannot enforce, that is not a moat. It is a request.
So the plumb-line question for every openness announcement, from either country: what is the model to the company releasing it? If it is the product, "open" is a concession. If it is a complement, it is a weapon. Altman answered the question in one email, four years before discovery made it public. The fight was never open versus closed. It is who can afford to give the model away.