Waymo Just Revealed Its Own AI Chip. Here's Why That Matters More Than the Robotaxi Itself
Waymo Just Revealed Its Own AI Chip. Here's Why That Matters More Than the Robotaxi Itself
Updated September 26, 2026. Every figure below comes from Waymo's own published pages, the California DMV, the California Public Utilities Commission or NHTSA, all linked at the end and all read that day. Rates per million miles and per-camera figures are arithmetic done here from those published numbers. Where Waymo has announced something without publishing a number, this article says which number is missing rather than estimating it.
One blog post, one headline number, and a conspicuous set of blanks
On August 20, 2026 Waymo published a post describing the compute stack inside its vehicles. The disclosed facts are these: a purpose-built 5nm ASIC, unnamed and unbranded; "over 1,000 TOPS of ML performance dedicated to front-end processing" of raw lidar, radar and camera data, including temporal denoising for low light; specialised accelerators feeding a purpose-built inference engine running sensor fusion; a heterogeneous design with two independent engines for redundancy; integration into the vehicle's liquid cooling loop; and simultaneous processing of thirteen high-resolution cameras.
Waymo does not claim it has replaced anyone. The post names AMD, Micron, Nvidia, Samsung, Sandisk, Socionext and TSMC as partners and positions the ASIC as one component of a mixed stack.
Now the blanks, which are the more interesting half. Not published: the chip's name, its power draw in watts, its thermal design point, its cost, any latency figure despite "ultra-low latency" being the stated design goal, die specifications, a deployment date, or which generation of the Waymo Driver carries it. As of this writing that single post is the entire public technical record. A reader who wants to know whether this silicon is in the car that picks them up cannot find out.
What 1,000 TOPS is being spent on, converted into something countable
The one usable arithmetic in the announcement is the pairing of the compute figure with the camera count. Thirteen high-resolution cameras against more than 1,000 TOPS of dedicated front-end processing works out to roughly 77 TOPS per camera stream, before any of the planning, prediction or control work that happens downstream of sensor fusion. That is the part of the system Waymo chose to build silicon for: not the driving decisions, but turning photons and returns into a usable picture fast enough that the decision layer is looking at the present rather than the recent past.
Which is why the missing latency number matters more than the missing wattage. "Pixels to actuation" is the metric the design is optimising, Waymo says so explicitly, and it is the one metric not given a value.
The only outcome the public can actually check
Compute claims are unfalsifiable from outside. Crash rates are not, and Waymo publishes them against a human benchmark for the same cities and road types. As of data current through June 2026, Waymo reports 271.3 million rider-only miles — Phoenix 92.121 million, San Francisco Bay Area 82.421 million, Los Angeles 67.098 million, Austin 21.064 million, Atlanta 8.624 million.
| Crash category | Reduction vs human benchmark | Fewer events over 271.3M miles | Per million miles (calculated here) |
|---|---|---|---|
| Serious injury or worse | 95% | 55 | 0.01 vs 0.21 — a gap of 0.20 |
| Any injury-causing crash | 82% | 841 | 0.67 vs 3.77 — a gap of 3.10 |
| Airbag-deployment crash | 82% | 358 | — |
| Pedestrian injury | 93% | 86 | — |
| Cyclist injury | 86% | 56 | — |
| Motorcyclist injury | 82% | 36 | — |
Waymo also reports that 41% of its collisions involved a delta-V below 1 mph — contact events at parking-lot speeds, which is worth knowing before reading any raw collision count.
The conversion that makes the compute argument concrete: at a gap of 3.10 injury crashes per million miles, every additional million rider-only miles corresponds to roughly three injury crashes that the benchmark would have produced and this system did not. A fleet driving a million miles a week is therefore operating at about 160 avoided injury crashes a year against the benchmark — arithmetic from Waymo's published rates, not a Waymo claim.
The percentages moved as the miles piled up, which is the real trend line
In September 2024, at roughly 22 million rider-only miles, Waymo published 84% fewer airbag-deployment crashes, 73% fewer injury-causing crashes and 48% fewer police-reported crashes. The current page, at 271.3 million miles, reports 82% fewer airbag crashes and 82% fewer injury crashes.
| Measure | Sept 2024 (~22M miles) | Through June 2026 (271.3M miles) |
|---|---|---|
| Airbag-deployment crashes | 84% fewer | 82% fewer |
| Injury-causing crashes | 73% fewer | 82% fewer |
| Police-reported crashes | 48% fewer | Not stated in the same form |
Injury reduction improved by nine percentage points while mileage grew more than twelvefold; airbag reduction slipped two. Neither movement is attributable to any particular hardware change, because Waymo does not publish which fleet software or silicon generation covers which mileage. That is precisely the gap a dated chip disclosure could have closed and did not.
Two of Waymo's own pages do not agree on the mileage
The safety impact page reports 271.3 million rider-only miles through June 2026. Waymo's FAQ, undated, says "more than 200 million miles of fully autonomous experience." Both are Waymo's. They are not reconcilable into one figure and should not be averaged; the safety page carries a date and the FAQ does not, so the dated one is the one to use, and the discrepancy is a reminder that even first-party numbers need a timestamp attached before they are quoted.
On scale, Waymo's 2025 year-in-review reports reaching one million fully autonomous rides per month by spring 2025, more than 14 million trips during 2025, and over 20 million lifetime rides by the end of that year, with a stated goal of one million rides per week by the end of 2026 and expansion to more than 20 additional cities. The FAQ lists thirteen directly served cities plus Austin and Atlanta through the Uber partnership. Fleet size is not published anywhere on Waymo's site.
Tip: Waymo publishes its own crash data; for the car you drive, nobody does. A dash cam is the closest a private driver gets to the same record, and it costs less than one insurance deductible. (These are Amazon Associate links — we may earn a small commission on qualifying purchases.)
What regulators compel, and what they leave out
Three regimes touch this and none of them requires a chip disclosure.
The California DMV collects annual disengagement and mileage reports from every autonomous-vehicle permit holder; its most recent release, dated February 20, 2026, reports that permit holders together logged more than 9 million test miles between December 1, 2024 and November 30, 2025 — an aggregate across companies, not a Waymo figure.
The California Public Utilities Commission granted Waymo a driverless deployment permit in November 2022 and fared passenger service authorisation for San Francisco in August 2023. Deployment carriers file quarterly data: vehicle miles travelled during passenger service including the electric-mile subset, miles from start location to pickup, daily average and monthly total idle time, passenger occupancy per trip, wheelchair-accessible-vehicle request fulfilment, and — once a carrier exceeds 300 trips in a quarter — incident-level and fleet-level stoppage events.
NHTSA's Standing General Order requires reporting of crashes where an automated driving system was engaged within 30 seconds, covering property damage, fatality, vulnerable-road-user strikes, airbag deployment or hospital transport, and publishes the raw incident data as downloadable CSV files, current through August 17, 2026 at the time of writing. There is no published NHTSA summary statistic for Waymo specifically; that figure exists only inside the dataset.
So the regulatory record covers miles, disengagements, trips, occupancy and crashes. It does not cover compute, power, or which hardware revision was in the car. A company can therefore ship new silicon into a passenger-carrying fleet without any filing that names it.
The energy question nobody answers
No published figure for the onboard compute power draw of an autonomous vehicle could be found — not from Waymo, not from the Department of Transportation, not from the national laboratories. Waymo's own compute post gives TOPS and mentions liquid cooling without stating a single wattage. A chip that is liquid-cooled and delivers over 1,000 TOPS is not a trivial load on a battery-electric vehicle's range, and the absence of any number, first-party or federal, is a real hole in the public record rather than an oversight by this article.
Why the chip is the more interesting story, if it is one
The robotaxi is a product; a lot of companies have one. Custom silicon at the sensor front end is a statement about where the remaining engineering problem lives — not in deciding what to do, but in seeing quickly and cheaply enough that the deciding has time to work. Building an ASIC for that layer only makes sense if perception latency and unit cost are the binding constraints on scaling to the twenty-plus additional cities Waymo says it wants.
The honest counterweight: on the published record this is one blog post with one number. The safety figures, which are extensive and dated, cover a period during which the chip's deployment status is unknown, so nothing in the 271.3-million-mile record can be attributed to it. Anyone treating the announcement as an explanation for the safety numbers is connecting two things Waymo itself has not connected. What would change that assessment is small and specific: a deployment date, a latency figure, and a power number. Until those exist, the chip is a signal about strategy, not a measured improvement.
What to check before believing the next claim about this
- Ask whether a mileage figure is dated. Waymo's own site currently carries two that differ by about 70 million miles.
- Read reductions as rates, not headlines: 0.67 versus 3.77 injury crashes per million miles is the same fact as "82% fewer" and is easier to compare against anything else.
- Note that 41% of reported collisions were under 1 mph delta-V before comparing raw collision counts with any other source.
- Treat the CPUC quarterly filings and the NHTSA CSV as the primary record; press releases are downstream of them.
- For the chip specifically, watch for three numbers — watts, latency, deployment date. None exists yet.
Related reading: what Nvidia's quarter says about who is buying AI silicon, the music-industry lawsuit against Anthropic, and what the federal projections say about entry-level tech jobs.
Sources for every figure above
- Waymo, the compute stack announcement — August 20, 2026. 5nm ASIC, 1,000+ TOPS, thirteen cameras, two engines, liquid cooling, partner list.
- Waymo safety impact — 271.3 million rider-only miles through June 2026, per-category reductions, IPMM rates, the 41% sub-1-mph figure.
- Waymo safety data hub, September 2024 — the ~22-million-mile comparison used for the trend table.
- Waymo 2025 year in review — rides per month, 2025 trip total, lifetime rides, 2026 targets.
- Waymo FAQ — city list and the undated "more than 200 million miles" statement.
- California DMV autonomous vehicle disengagement reports — the aggregate 9-million-mile figure released February 20, 2026.
- California Public Utilities Commission autonomous vehicle programs — permits and quarterly reporting requirements.
- NHTSA Standing General Order crash reporting — reporting triggers and the public CSV dataset.
Safety comparisons above are Waymo's own published analysis against its chosen human benchmark for the same service areas and road types; they have not been independently replicated here. This is general information, not advice about any vehicle purchase or transport decision.
Comments
Post a Comment