Waymo has designed its own AI chip for its robotaxi fleet, pulling a piece of its onboard compute stack off Nvidia hardware. For a company that runs paid, driverless rides in multiple U.S. cities, that's not a side project — it's a change to the silicon sitting between a sensor detecting a pedestrian and the car deciding what to do next.
According to The Decoder, the custom chip is built specifically for Waymo's robotaxi operations and reduces the company's reliance on Nvidia in what the outlet describes as critical infrastructure. Waymo hasn't detailed the chip's architecture or performance numbers publicly, and the value of the story isn't in a spec sheet — it's in the decision itself.
That decision puts Waymo in company that used to be reserved for the largest hyperscalers and chipmakers: Google already builds TPUs, Amazon has Trainium and Inferentia, Tesla has Dojo, Microsoft has Maia, Meta has MTIA. Waymo, an Alphabet company, now designs its own silicon too — a step up from simply buying whatever Nvidia, AMD, or Qualcomm sell.
Why companies this size stop buying and start building
Designing a chip is expensive and slow — tape-out cycles run years, not quarters, and the upfront cost only pays off if you're deploying at a volume that amortizes it. That's why custom silicon has historically been a hyperscaler-only move. Three pressures tend to push a company over that line:
- Cost at scale. Buying merchant GPUs at fleet volume means paying Nvidia's margin on every unit, indefinitely. A custom chip converts that recurring cost into a one-time design investment.
- Latency and fit. General-purpose GPUs carry capability a specific workload doesn't need. A chip built for one job can strip that out and shave the milliseconds that matter when the job is real-time driving decisions.
- Supply control. Nvidia's allocation goes first to whoever buys the most, and AI labs training frontier models have been outbidding almost everyone else for GPU capacity. Owning your own chip means your roadmap no longer depends on someone else's allocation queue.
Waymo checks the first and third boxes clearly: it operates a large, homogeneous fleet — the same compute problem repeated across thousands of vehicles — and it has been racing to expand into new cities, which makes hardware supply a real operational risk, not an abstract one.
What's different about "critical infrastructure" here
Most companies swapping out Nvidia are optimizing a data center bill. Waymo's chip runs, at least in part, on the vehicle — powering perception and decision-making systems that have to work every time, with no fallback to a bigger data center if latency spikes. That's a different risk profile than a chatbot backend timing out for a second. In our estimation, this is likely why Waymo chose to build rather than simply negotiate better GPU pricing: a supply disruption for a robotaxi fleet isn't a cost problem, it's a fleet-grounding problem.
What this means if you're not Waymo's size
Almost nobody building AI products today operates at a scale where custom silicon is a first move — it took Waymo, backed by one of the largest cloud businesses in the world, over a decade of robotaxi operations to get there. For most teams, the practical takeaways sit one level down from "build your own chip":
- Don't assume GPU pricing and availability are fixed variables — the biggest AI infrastructure buyers are actively working to reduce their exposure to any single vendor, and that reshapes the market everyone else buys into.
- If a product depends on predictable, low-latency inference — robotics, autonomous systems, real-time agents — model that dependency on general-purpose compute as a risk, not a given, even if you'll never design a chip yourself.
- Watch which workloads specifically get pulled off Nvidia hardware by companies like Waymo, Tesla, and the hyperscalers. That's a signal for where merchant silicon is genuinely a bottleneck, versus where it's just convenient.
AiiN's takeaway
This isn't a story about Nvidia losing a customer — Waymo almost certainly still buys Nvidia hardware for training and other workloads. It's a story about what happens once a company's AI deployment is large and repetitive enough that owning the silicon becomes cheaper than renting it. Waymo joins a short but growing list of companies that have crossed that line, and it's a reminder that vertical integration in AI hardware isn't just a hyperscaler flex — it follows directly from cost and latency math that any company can eventually run, once it has enough scale to make the numbers work.