Can host, can't host performantly

Xiaoyin Qu, a former Meta product manager, posted a clean thesis that traveled: US and European companies will abandon OpenAI and Anthropic and run Chinese open-weight models on their own machines — and not only to save money. Her sharper charge is a trust failure. Anthropic recently disclosed it will hold customer prompts for 30 days even under contracts that promised zero retention, citing safety, while it pushes into healthcare and legal — the same verticals its enterprise customers work in. Hand a lab your data and then watch it build your product, the argument goes, and self-hosting an open model you control starts to look like the safer bet.

"For Lindy, switching off Claude was a matter of survival." — on the ~90% cost cut

The cost half is already happening. Lindy, an automation startup, moved all of its traffic off Claude to the Chinese model DeepSeek and cut the bill on those routes by roughly 90%, calling the switch a matter of survival; the bank UBS reckons about 60% of the companies it monitors are now curbing AI spend by routing to cheaper, often Chinese, models. The trust grievance is real and the price gap is real.

What no one has shown is the part the prediction leans on hardest. The most knowledgeable replies in Qu's own thread are practitioners who would like it to be true and still doubt it: downloading open weights is easy, but serving them fast and reliably to thousands of employees is a different, unsolved engineering problem. That gap — between owning a model and operating one at scale — is where the bold call meets the wall.