Waymo Built Its Own Robotaxi Chip on TSMC's 5nm Process

Waymo has designed its own chip for its robotaxis, and the message to Nvidia is hard to miss: the largest AI customers keep concluding that at sufficient scale, the best supplier is yourself.

Bloomberg broke the story on August 20, with The Decoder and several industry outlets confirming the details. The custom ASIC delivers more than 1,000 TOPS (trillions of operations per second), is manufactured by TSMC on a 5-nanometer process, and handles both jobs that matter in an autonomous vehicle: crunching the sensor feed from lidar, radar, and cameras, and running the AI models that let the car understand its surroundings and react. Until now, that work ran on chips from Nvidia and AMD.

Already in cars, not on a roadmap

Two details separate this from the usual custom-silicon announcement. First, the chip is already installed in Waymo's latest generation of vehicles, the robotaxis being developed with Chinese manufacturer Zeekr. Second, Bloomberg reports performance comparable to Nvidia's current autonomous driving systems, at lower cost. Waymo isn't hedging with a pilot program; it has committed its newest fleet to in-house hardware.

The economics explain the confidence. A robotaxi fleet heading toward tens of thousands of vehicles turns per-car compute into a nine-figure line item. Paying a vendor's margin on every unit makes sense when volumes are small; once they aren't, vertical integration starts paying for itself. Google has run its own TPUs for a decade, Amazon has Trainium, OpenAI is co-designing chips with Broadcom, and Anthropic announced its own chip design team this month. Waymo extends the pattern from the data center to the edge.

What it means beyond Waymo

For Nvidia, the near-term revenue impact is minor; automotive silicon is a rounding error next to its data center business. The direction of travel is the harder problem. Customer by customer, the biggest buyers of AI compute are converting into competitors for their own workloads, and each defection normalizes the next one. Supply security is doing as much work in these decisions as cost: nobody who lived through the GPU shortages wants a single point of failure in their hardware chain.

There is a transferable lesson for any business building on AI, whatever its size. When an external dependency becomes your dominant cost driver, there is a volume threshold past which bringing it in-house beats renting it. For most companies that threshold arrives not with chips but with API bills, and the equivalent move is shifting high-volume, routine workloads onto open-weight models on infrastructure you control. Waymo's chip is the billion-dollar version of a calculation worth running at every scale: know where your threshold is before you cross it.

Sources: The Decoder, Bloomberg