AI Is Leaving the Single Data Center.

That Is the Anti-Fragile Move.

This week’s tech news is not really about a foldable phone or another model launch. It is about where intelligence will live.

OpenAI is deepening work with Samsung on its own chips. Apple is arguing the device—not the cloud—is the right place for everyday AI. Mistral signed a billion-dollar deal with Loft Orbital and Marlan Space to put compute on a constellation of fifty satellites. Hugging Face’s co-founder, writing after the OpenAI-related breach of the platform, said commercial defense tools failed and that open-weight models are part of the real fix.

Those are four versions of the same idea: stop putting the whole system in one building, on one vendor, in one region.

Fragile AI stacks look efficient. One frontier API. One chip supplier. One hyperscale campus. They work until they do not—outages, capacity caps, a security incident, a grid constraint, a delayed interconnection. Then knowledge work freezes.

Anti-fragile stacks keep optionality. On-device models for work that cannot wait. A second chip path so one foundry or one GPU line cannot halt the company. Open weights you can run locally when the API is dark. Compute that is modular enough to live in a phone, a closet, or orbit—not only in a 500-megawatt hall.

The companies that treat the cloud as a default, not a monopoly, will get stronger when the next shock hits. The ones that optimized for a single pipeline will discover they built a single point of failure with a chat window.

Distribute the intelligence. Keep the option to run without permission from one status page.