On Wednesday morning Palantir chief executive Alex Karp sat down with CNBC and delivered what one viewer called a televised nervous breakdown. He accused OpenAI and Anthropic of running an "effing insane" business model, of overselling the risks of their own technology while quietly handing the same powerful models to any company or government that would pay, and of charging enterprises what amounts to a wealth tax that transfers their competitive edge to a handful of labs. "This is the voice of American business that is being channeled through me," he told an incredulous Becky Quick. It was good theatre. It was also, underneath the heat, a fairly precise description of a shift already under way.
The specific grievance is the token. Frontier labs charge by the volume of text their models read and write, a meter that spins faster the more useful the tool becomes. Palantir's accompanying "AI sovereignty" manifesto gave the complaint a name, "tokenmaxxing," and argued that metered pricing rewards disposable scripts over durable software and, more damagingly, siphons an institution's hard-won data and model weights out to third parties. "Controlling your weights is controlling your fate," the company wrote. Strip away the swagger and the claim is straightforward: if the model that encodes your business knowledge lives on someone else's servers, so does your advantage.
What makes this more than a rival vendor talking his own book is that enterprises are already voting with their budgets. Coinbase says it cut internal AI spending by nearly half by routing engineers to Chinese open-weight models, Zhipu's GLM and Moonshot's Kimi, through an internal gateway. Microsoft is reportedly weighing a fine-tuned version of China's DeepSeek to power parts of its Copilot assistant. The fast-growing coding startup Cursor built its latest model on top of Kimi. Data from OpenRouter shows Chinese open models at times accounting for more than 60 percent of the tokens consumed by leading models. The appeal is not only price: a model you can download and run on your own infrastructure is a model no government can switch off.
That last point lands harder after the past month. When Washington briefly forced Anthropic to cut off its most capable models, enterprises got a live demonstration of what dependence on a metered, licensable American model actually means. Fortune's correspondent noted that even inside the United States, "a lot more enterprises are now talking about open source." The sovereignty argument stopped being abstract and became a procurement checklist.
There is a catch the manifesto glides past. The most capable open-weight models in the world are Chinese, which trades one dependency for another and carries its own reputational and regulatory risk, including the real possibility that Washington moves to restrict their use too. Owning your weights sounds like freedom until you notice whose weights are on offer. Karp's outburst will be remembered for the profanity, but the question he was really asking is the one every large buyer of AI is now forced to answer: is intelligence something you rent, or something you had better learn to own?