Waymo builds custom silicon for autonomous vehicle inference

Waymo hasn't disclosed whether this chip targets inference-only workloads or includes training capability, nor has it committed to a production timeline or stated whether the silicon will remain proprietary or eventually license to other AV makers. The announcement is light on the actual performance gains relative to current Nvidia or Qualcomm solutions.

This is largely disconnected from recent activity in the autonomous vehicle space covered by Modelwire, since we have no prior related coverage to anchor against. However, it belongs to a broader pattern of AI-heavy companies (Tesla, Apple, Meta) building custom silicon to escape vendor lock-in and reduce per-unit costs at scale. The move signals that Waymo believes it has reached sufficient technical maturity and deployment volume to justify the R&D and manufacturing complexity of bespoke hardware. It also suggests the company is preparing for a competitive environment where hardware efficiency becomes a margin lever, not just a nice-to-have.

If Waymo deploys this chip in production vehicles within 18 months and publicly reports latency or power consumption improvements of 20% or more over its current stack, that confirms the bet was sound. If the timeline slips beyond 24 months or if the company remains silent on performance metrics after launch, the project may be more about optionality than necessity.