Image: Google

Google wants the AI chips inside watches, earbuds and small robots to follow one open design. Google Research has completed its Coral NPU blueprint, released a free compiler for it and, with chipmaker Synaptics, announced the first commercial chips built on the newest version. Google AI shared the news on X.

How we're bringing powerful AI to small devices without draining the battery.

— Google AI (@GoogleAI) October 22, 2026

An NPU is a small processor built just for AI math. Your phone already has one. The problem Google describes is power. Small devices waste much of their battery moving data back and forth between memory and the processor. Google says Coral NPU cuts that waste by doing more of the math close to where the data sits.

Image: Google

The new piece is a matrix unit, the part that handles the heavy grid math behind modern AI models. Google says it can be configured for up to 1 TOPS, a measure of how many trillion operations a chip runs each second, while targeting about 10 milliwatts on average. That is a tiny fraction of what a phone chip uses. Google says this lets small gadgets run compact models such as Gemma, live voice translation on wearables and visual perception on robots. Those are Google's targets, not measured results from shipping products.

Image: Google

The bigger idea is openness. Most small AI chips use their own closed designs and their own software tools. Google says developers pay a "fragmentation tax" by rewriting and retuning models for every chip. Coral NPU is built on RISC-V, an open chip standard that anyone can use without licence fees to a single owner. Google has put the full design on GitHub and released version 1.0 of its open compiler. The compiler takes models from JAX, PyTorch and TensorFlow and turns them into instructions for the chip.

If it works as planned, a developer could tune one model and run it on chips from different makers. That is roughly what Arm did for phone processors, but with an open blueprint instead of a licensed one.

Real hardware is starting to arrive. Synaptics launched its Astra SR200 series of microcontrollers, rated from 200 to 400 billion operations per second and aimed at wearables, cameras and sensors. It also made its Coralboard SL developer kit generally available through DigiKey and announced a new Coralboard SR kit. TechRepublic reported that the SR200 chips are still sampling, with no firm production date or pricing yet.

For regular buyers, nothing changes today. Coral NPU is a building block for companies that make gadgets. The payoff, if it comes, is smarter earbuds, glasses and home devices that work without the cloud, keep data on the device and last longer on a charge.

Several things are still unproven. Google's power and speed figures are design targets, and no independent tests have been published. Only one chip partner has announced products so far. An open standard only matters if many chipmakers adopt it, and Google has not named other device brands using Coral NPU. Google says a 64-bit version and extra security features are next on its roadmap.