For three years the default answer to "where does the AI run?" was simple: in the cloud, behind an API, billed per token. That default is quietly breaking. A growing share of inference in 2026 happens on the device itself, a phone, a laptop, a camera, a shop-floor sensor, and never touches a data center at all. One widely cited figure puts as much as 80% of AI inference now running locally rather than in the cloud. Whether the exact number holds, the direction is not in doubt.
This matters for anyone building software, not just chip designers. The question is no longer only "which model," it is "where should this particular call run." Get that decision right and you cut latency, cut cost, and sidestep a whole category of privacy problems at once. Get it wrong and you pay a cloud round-trip for something that should have happened in 5 milliseconds on the user's own hardware.
What edge AI actually means
Edge AI is inference that runs at or near where the data is created, instead of shipping that data to a central server and waiting for an answer. The model lives on the phone, the browser, the gateway, the camera. Nothing about the request leaves the local network unless you decide it should.
