Srikanth Thirumalai's line to Axios, that physical AI has no click reboot and you deal with the consequences [7], is written for whoever decides what happens after a failure. In practice that is a city permit office and a fleet underwriter, not a rider choosing between apps. The technical content is real, and the post is also now a document Waymo has handed to anyone who later asks why it picked the sensors it picked [1][6].
The most useful number in it is not the mileage counter. Cumulative miles only rise, and 200 million of them [5] describe how much Waymo has driven rather than how much its service gets used. Divide 500,000 paid trips a week by roughly 4,000 vehicles [8] and you get about 125 paid trips per car per week, close to 18 a day [1]. That is a utilization figure, and it is the one that would embarrass Waymo if it sagged while the fleet and the city count grew.
Tesla's headline number sits on a different axis. A recent filing points to more than 125,000 Cybercabs a year [11], roughly 31 times Waymo's entire current fleet in a single production year [2]. The vehicle has no steering wheel and no pedals [11], so there is no partial-credit version to sell while the software matures. Musk's promise of a million robotaxis on the road in 2020 [12] is the useful calibration for which of these two figures is a target and which is a receipt.
The counter-argument deserves to be taken at face value. New Street Research's Pierre Ferragu wrote that Waymo built a driving gas plant that AI at scale makes irrelevant, and called the blog post incumbent rhetoric [9]. Waymo's evidence has a real limit that matches this: 200 million miles of a fused-sensor stack shows that a fused-sensor stack works at this volume [5][8], and it cannot show that a camera-only stack fails. What Waymo can do is make the alternative expensive to defend in front of a regulator, and announcing three more markets the same morning [3] does more of that work than the prose does.
So here is a test that travels to your own vendor evaluations. Ask which published number would have to quietly disappear for the vendor's architecture claim to be abandoned. For Waymo, it is weekly paid trips and the city count; if those keep appearing alongside the sensor argument, the claim is load bearing rather than positioning. For Tesla, the equivalent tell is the safety monitor. Pulling monitors from a majority of trial cars in Texas and Florida [13] is the first move that lets an outsider check the camera-only claim against something, and it happened before the reveal rather than at it.
