Waymo's 5nm ingest chip moves autonomy economics into the silicon design budget

Pre-processing is the part of the stack nobody demos. Thirteen high-fidelity cameras on the Ojai ^[5], plus whatever else is bolted to the roof, produce a raw feed that has to be conditioned and aligned before a planner can do anything with it, and Waymo says its 5 nm part sits precisely there, ahead of the core driving brain ^[4]. That placement is the choice worth reading. Ingest is fixed, well-specified and repetitive work, which is exactly the workload where fixed-function silicon beats general-purpose parts on watts and dollars per frame, and exactly the workload you can freeze in hardware without freezing the models that ride on top of it.

Then the number. More than 1,000 TOPS ^[6], which TechCrunch puts roughly in the range of Nvidia's DRIVE AGX Thor, a processor sold as the compute for automated driving ^[7]. Read those two together and Waymo is spending a Thor-class performance envelope before inference begins ^[1]. Either the sensor load is heavier than outsiders price it, or Waymo concluded the cheapest headroom in the vehicle is upstream of the expensive silicon rather than inside it.

The economics of that decision are not software economics. Custom silicon converts a recurring purchase from a merchant vendor into a one-time design and tapeout bill plus a lower per-unit charge, and it only clears if the units arrive. Waymo has repeatedly said the next-generation vehicle is cheaper to build, operate and maintain, and has framed that as a precondition for eventually making money ^[3]. It has now opened the Ojai to all riders in Los Angeles, Phoenix and San Francisco ^[2], which is the volume side of the same equation. What it has not published is a cost figure. The only quantity on the table is a TOPS rating, and TOPS is a throughput claim, not a bill of materials. Waymo's own summary of the work is that the system has "unmatched efficiency and performance" ^[8], which is a sentence, not a metric.

The partner list complicates the vertical integration story in a useful way. Waymo named AMD, Micron, Nvidia, Samsung, Sandisk, Socionext and TSMC, some for the first time ^[10], without saying who contributed what. A company that designs its own ingest ASIC and still lists Nvidia as a compute partner is not exiting the merchant market; it is picking the one block where its sensor suite is idiosyncratic enough to justify a mask set, and buying the rest.

The other half of the per-vehicle bill showed up in the same newsletter. According to TechCrunch's sources, Idaho National Laboratory is evaluating whether Chinese lidar sensors would pose a security risk if widely deployed on US vehicles, and the work is funded by a company or group of companies in the electric and autonomous vehicle industries ^[11]. TechCrunch says Rivian, General Motors, Ford, Kodiak, Lucid Motors, Nuro and Uber all reported no knowledge of the review, while Aurora, Nvidia and Zoox did not respond ^[12]. Compute is the half of the sensor stack an operator can design its way out of. The sensors themselves remain someone else's supply chain, and someone in this industry is paying a national lab to ask who that someone should be.