Waymo Details Custom 5-Nanometer Chip and Compute Stack Inside Its Robotaxis
Waymo has disclosed the hardware architecture powering its driverless fleet, including a custom 5-nanometer chip that processes sensor data at one quadrillion operations per second.
Autonomous vehicle company Waymo has released technical details about the in-vehicle computing systems powering its driverless fleet, offering a rare look at its custom silicon, hardware architecture, and key component suppliers. As reported by The Verge, the system housed in each vehicle trunk can execute up to one quadrillion operations per second.
The disclosure centers on a custom 5-nanometer Application-Specific Integrated Circuit (ASIC) built specifically to handle the immediate torrent of incoming sensor data from 13 high-resolution cameras, radar, and lidar units. Delivering 1,000 trillion operations per second (TOPS) of front-end processing power, the chip filters noise, fuses sensor feeds, and performs initial machine learning tasks before passing structured data to the primary computing system.
Waymo executives Satish Jeyachandran and Daniel Rosenband outlined three core engineering principles behind the hardware design: low latency, ruggedized durability against road vibration and temperature extremes, and dual-system redundancy to maintain operation during hardware failures. The company said its onboard compute capacity has grown twentyfold over the past eight years to support real-time sensor fusion and neural network inference.
Waymo does not build all of its silicon in-house. Its heterogeneous compute architecture draws on a multi-vendor supply chain that includes AMD, Micron, Nvidia, Samsung, SanDisk, Socionext, and TSMC. Waymo's proprietary silicon handles front-end data ingestion and sensor synchronization; commercial accelerators, GPUs, and processors take on broader machine learning workloads and vehicle orchestration.
The company currently operates approximately 4,000 autonomous vehicles across more than 10 cities and delivers around 500,000 paid rides per week. Industry estimates put the hardware cost of its sixth-generation compute and sensor suite at between $20,000 and $25,000 per vehicle, down sharply from the estimated $100,000 to $125,000 cost of its fifth-generation setup. Hardware deployment expenses remain a major factor in overall fleet economics.
The Cognarah Angle
Waymo's technical disclosure lands in the middle of an unresolved philosophical divide in autonomous driving. Waymo bets on multi-modal sensor fusion: cameras, radar, and lidar working together, supported by heavy trunk compute and custom front-end silicon. Its main competitors argue the opposite, that vision-only, end-to-end neural networks eliminate unnecessary hardware complexity and sensor contention. By publishing the specifics of its ASIC, Waymo is making a public case that sensor fusion is not just the safer choice but also the computationally defensible one.
The architecture backs that argument up. Processing and aligning raw multi-sensor feeds at the silicon edge reduces latency while preserving the fail-safe redundancy that commercial transit authorities require. The design works in the field. But building it requires an engineering team large enough to design custom silicon, a supply chain spanning seven major chip and memory vendors, and a balance sheet capable of absorbing the full cost at scale. That is not a playbook most entrants can replicate.
Which raises the more uncomfortable question: getting hardware costs to $25,000 per vehicle is a genuine engineering milestone, but it is not a solved problem. Scaling to hundreds of thousands of vehicles means tying up billions of dollars in computing assets that depreciate quickly, face constant thermal stress, and will be outpaced by the next generation of silicon within a few years. High-density, multi-vendor compute stacks are not designed for cheap, fast refresh cycles.
Waymo has demonstrated that a computer can drive a car safely. What it has not demonstrated is that a company burdened with $25,000 in onboard hardware per unit can sustain operating margins against human-driven fleets running on a smartphone mount and a data plan.
Reporting sourced from The Verge. Analysis and Cognarah Angle are Cognarah's own.
Written by
StacyAI-assisted news curation. Every story is reviewed by our editors before publication.
