A $38,000 van lands in an American port and, after a 127.5% tariff, costs Waymo roughly $86,500 before a single sensor is bolted on. Add the Gen 6 hardware-and-software stack at about $25,000, and each of these Chinese-built robotaxis crosses $100,000 on its way to the curb.
This is the vehicle the market leader chose. Not despite the tariff—because even at $100,000 it undercuts the Jaguar I-Pace, which Waymo reportedly pays upwards of $200,000 to put on the road.
Sit with that. The most successful robotaxi operator in America is financing its fleet on Chinese hardware it pays a nine-figure import penalty to bring in, and that's still the cheap option.
The most successful robotaxi operator in America is financing its fleet on Chinese hardware it pays a nine-figure import penalty to bring in—and that's still the cheap option.
Last week the open question was which regulator would first permit an empty driver's seat for paying passengers. This week the picture sharpened into something more useful: the field has stopped being one race and split into three tiers, sorted not by vehicle or vision but by what each operator will put in public.
Start with the winner. Waymo just received CPUC approval to expand across the SF Bay Area and LA and to bring service to Sacramento and San Diego. It runs 200,000 rides a week, reportedly up from 10,000 a year ago and 100,000 by August 2024, per reporting on Alphabet's updates.
And it shows its work. Over 25.3 million miles, Waymo recorded eleven accidents against the 104 expected of human drivers over the same distance—88% fewer property-damage incidents, 92% fewer injuries, per Waymo's internal analysis.
Waymo's expansion is gradual, "guided by our safety framework."
That is tier one: scale earned on evidence already on the table.
But the tariff line is the tell. It reveals that even the evidence-leader has no clean domestic path to scale. More than 3,200 Zeekr vans have reportedly been imported since 2024, more than 2,600 of them this year alone—and Waymo says "more than 100" are actually in service.
More than 100. Out of 3,200-plus. The rest are testing, sitting in lots, or waiting on a deployment date nobody will name.
So the leader's real constraint isn't the software. It's a supply chain that runs through Chinese docks and a tariff regime designed to keep Chinese cars out—a contradiction Waymo is eating by the boatload because domestic hardware costs twice as much.
Now the middle tier. This week Zoox reportedly published its Safety Case Framework publicly for the first time, "outlining how we define, evaluate, and validate safety" across its purpose-built platform, software, and operations.
Publishing a framework is real. It's more than a rendering and more than a promise. One sharp reply nailed the operative detail: the whole thing hinges on the ODD boundary—the operational design domain—and "scale depends on where that boundary becomes approved service, charging and support."
The problem is what Zoox pairs it with. In the same news cycle, the company promoted CTO Jesse Levinson discussing "thousands of Zoox robotaxis, in multiple cities, commercially."
Thousands. Multiple cities. Commercially. And underneath that post sits the question the company has not answered: "What is the total vehicle count today?"
That's the tension in one screen. A framework describing how safety will be validated, a promise of thousands of vehicles, and no published fleet count, no fare, no cost-per-mile for paid service.
The framework is a down payment on a promise Zoox hasn't yet priced. Meanwhile it's staging an installation at SFMOMA for one night so people can "step inside our robotaxi." The experience is real; the economics are still unpublished.
Which advances the standing open thread on this beat: whether Zoox will ever attach an actual fare or cost-per-mile figure to paid service, or keep letting the framework and the museum stand in for one.
That's tier two: a published methodology, an unpublished business.
Then there's Tesla, and here the inversion completes itself.
Tesla reportedly has roughly 250 Cybercabs spotted in factory lots as of late July, built at Gigafactory Texas since February. The two-seat vehicle has no steering wheel and no pedals, which means that outside its small geofenced validation zones, it "cannot be driven at all."
Read that again. Hundreds of finished cars that are, by design, undrivable everywhere except the narrow patches where the software has been signed off.
Musk himself framed the caution on the Q2 call: the company needs to build up Cybercab-specific driving data on a new platform before putting many on the road. That's a validation argument—and it's an argument for why you don't mass-produce ahead of validation, made by the company mass-producing ahead of validation.
The public rollout is being sold through a sweepstakes reportedly running through August 23, offering riders a chance to attend the launch event. An end-of-August target hovers over a timeline that, in Tesla's own history, has a habit of slipping.
And then the Starlink dish. The latest units carry an integrated satellite terminal, reportedly ~375 Mbps per Motor1's reporting, first shown August 10—positioned as a second communications path for navigation, customer service, and fleet management, explicitly not the autonomy stack, which runs on the AI4 computer offline.
Here's the quiet part, stated plainly in the reporting: in Tesla's launch cities, where cellular is already dense, "the practical benefit is limited." The satellite payoff, per the reporting, more likely arrives in remote, low-coverage territory Tesla isn't operating in yet.
So the dish is hardware for a future that isn't here, bolted onto cars that can't leave the geofence, promoted via a contest, ahead of a date that may move. What's missing from all of it is a public safety case—the audited miles-and-incidents evidence that got Waymo four California metros.
Tesla is substituting geofences and Starlink and a sweepstakes for the thing regulators actually reward. Testing in Austin, Chicago, and San Francisco still runs with a remote-monitoring layer where a human can take over—the safety net Tesla hasn't yet published its way past.
That's tier three: hundreds of vehicles, a launch event, and no public evidence standard behind them.
Now stack the three tiers and the pattern is impossible to miss.
Waymo promises the least in language and shows the most in data: 200,000 weekly rides, 25.3 million audited miles, four approved metros. Zoox promises thousands and shows a framework. Tesla promises multiple cities by end of August and shows a parking lot and a raffle.
The more a company promises, the less it has published. That's the inversion, and it holds cleanly across all three.
Which brings the whole sector back to that $86,500 van. The tariff line isn't a quirk of Waymo's procurement—it's the structural fact underneath the entire US robotaxi buildout.
Because the one operator with the evidence to scale has no cheap domestic way to build the fleet, and the operators with cheaper stories haven't shown the evidence. The buildout is being financed on either Chinese hardware at a 127.5% penalty or American promises at zero published proof.
Pick your poison. There is currently no third door.