While the big models keep swallowing ever more data centers, Google went the other way at I/O: a tiny board that runs AI entirely without the cloud. The new Coral Board is a compact single-board computer for on-device AI.
What’s inside
At its core sits the Coral NPU — an open-source machine learning unit built on RISC-V and developed by Google Research. Alongside it is a Synaptics Astra SL2619 chip with a 2 GHz dual-core processor, 2 GB of RAM, and 1 TOPS of compute.
That’s enough to run Google’s compact open-source language model Gemma 3 270M entirely on the board — no cloud access required. And that’s the whole point: the data never leaves the device.
What it’s for
The board targets small devices: headphones, AR glasses, smartwatches. The Coral NPU also aims to tackle an old problem — the fragmentation among AI accelerators, where every vendor does its own thing. With an open, unified architecture, developers could build once and run on more hardware.
At I/O, Google showed off fitting demos: real-time translation, voice-controlled hardware, and a generative music performance where a vision model turned jellyfish movements into sound. All the demos are open source on GitHub. The board is expected to ship this summer, though Google hasn’t announced a price yet.
My take
Edge AI is the unflashy counterpart to the frontier-model arms race — and that’s exactly what makes it interesting. Not every task needs a $900-billion data center. Translation in your ear, voice control on your glasses, small classifiers on your watch: all of that runs better locally, faster, and without uploading data anywhere. The genuinely exciting part is that Google built the NPU as open source and on RISC-V — because that could finally clean up the patchwork of edge accelerators. For Claude fans this isn’t a direct topic, but it shows where the other half of the AI world is heading: not up into the cloud, but down into the device.
Sources: The Decoder · Google Developers · Synaptics