Waymo designs 5nm AI chip, mirroring Tesla's in-house silicon bet

Waymo disclosed Thursday that it has built a custom 5nm AI chip, a cutting-edge manufacturing process from TSMC, for the front-end sensor stack on its robotaxis. The in-house ASIC pairs with off-the-shelf CPUs, GPUs, and accelerators from AMD, Micron, NVIDIA, Samsung, Sandisk, and Socionext inside a liquid-cooled, dual-redundant compute module that processes 13 high-resolution cameras in real time (Waymo blog).

The chip, introduced by engineering leads Satish Jeyachandran and Daniel Rosenband, is rated at more than 1,000 TOPS, or trillions of operations per second, the standard throughput metric for AI accelerators. Waymo says its in-vehicle compute has scaled 20x in eight years, and the system runs two independent compute engines for redundancy.

The move mirrors Tesla's vertical-integration bet on in-house compute rather than renting merchant silicon from Nvidia and others. The Register framed the disclosure as a chip designed "to stay ahead of Tesla"; Waymo's own blog names no competitor.

Independent reporting by The Verge puts the operational stakes in scale: roughly 4,000 vehicles in 10-plus cities and about 500,000 paid trips per week, on sixth-generation hardware that a Forbes estimate, as cited by The Verge, values at $20,000 to $25,000 per vehicle, down from $100,000 to $125,000 for the prior generation.

Waymo says it will present more detail at Hot Chips 2026 at Stanford Aug. 23–25.