Wayve Is Selling the Driver, Not the Fleet

Wayve launched its first public robotaxi service with Uber in London on September 3 — a small fleet, with a trained, TfL-licensed private hire driver in every vehicle to supervise and intervene. Tokyo is next.

Compared against Waymo, which as of September 2026 operates in 14 US metros with roughly 4,000 vehicles serving about 500,000 paid rides per week, that is a rounding error.

Compared against what Wayve is actually selling, it is a reference deployment.

The model is licensing

Wayve's strategy, laid out again this week, is to license its autonomous-driving AI to other companies rather than operate its own fleet. The product is the AI Driver; the customer is whoever owns the cars.

This is a genuinely different bet from everyone else in the category, and the distinction is not cosmetic.

Waymo is a vertically integrated operator: it builds the stack, buys and retrofits the vehicles, runs the depots, manages the fleet, and takes the ride revenue. Tesla is vertically integrated in a different direction — it builds the cars and the stack and intends to run the network. Both models require enormous capital per city and produce a business whose unit economics are fleet economics: vehicle depreciation, depot real estate, cleaning, remote assistance staffing, insurance.

Licensing inverts that. If the AI Driver is a software product sold to automakers and fleet operators, the marginal cost of the next city is close to zero and the capital burden sits on whoever owns the metal. The company scales like software instead of like a taxi company.

The catch is equally structural: you do not control the hardware, the sensor suite, the maintenance quality, or the operational discipline of the entity deploying you — and you carry reputational and probably liability exposure for all of it.

Why the technical approach makes licensing possible

Wayve's system, AV2.0, is an end-to-end deep-learning driver that learns from experience rather than executing against pre-built high-definition maps. Wayve's claim is that this lets it adapt to new roads, vehicles, weather and cities efficiently.

Mapless is the load-bearing word.

The map-first approach that defined the first decade of this industry produces excellent behavior inside the mapped region and requires a survey, a validation pass and ongoing maintenance for every new area. That is why Waymo's expansion is measured in metros and why each one takes quarters. The map is an asset and a tax.

An end-to-end learned driver has the opposite profile: weaker guarantees in any specific location, but a generalization story that makes "works in a city you have not surveyed" a plausible claim rather than a category error. If it holds, it is the only approach that could support licensing at all — because you cannot license a driver that requires you to map your customer's territory first.

London is a reasonable place to test it. Dense, chaotic, asymmetric intersections, aggressive cyclists, bus lanes, and roundabouts. A learned policy that handles central London is a stronger generalization claim than one trained on Phoenix.

The safety driver is the honest part

Every vehicle has a TfL-licensed driver who can take over. Riders request through UberX, Uber Comfort or Uber Electric at standard pricing and may be matched with a Wayve vehicle — no separate app, no premium, no novelty framing.

Transport for London granted the Private Hire Vehicle licenses by August, which is the regulatory mechanism that makes this work: it is legally a private hire service with an unusually capable driver assistance system, not a driverless deployment.

This will be read as Wayve being behind. It is more accurately Wayve being early in a jurisdiction that has not created a driverless category yet, and choosing to accumulate real-world miles under supervision rather than wait.

The relevant comparison is what supervision is costing everyone. Tesla's robotaxi fleet crossed one million miles of unsupervised operation, announced at the Cybercab event in Austin on September 3 — and NHTSA has an open look at the Cybercab, which ships without a steering wheel, mirrors, or brake pedals. The agency has separately directed AV developers to submit remediation plans for robotaxis interfering with active emergency scenes, characterizing it as a functional insufficiency rather than an edge case.

Nobody in this industry is operating without a supervision story. The variation is whether the supervisor sits in the car, in a remote operations center, or in a regulatory filing.

The scale everyone is building toward

Nevada's Transportation Authority unanimously approved permits for Tesla, Uber and Waymo to run commercial robotaxis in Clark County, collectively enabling up to 7,000 vehicles over twelve months.

That is the shape of the next year: multiple operators, one metro, permits issued in parallel. Which means the differentiator stops being "who has permission" and becomes "who can fill the permit."

Filling 7,000 slots requires capital for vehicles or a partner who already has them. That is precisely the gap a licensing model is designed to sit in — and precisely why Wayve's London fleet being tiny is not the metric to judge it by.

What to watch

The first licensee who is not Uber. An automaker or a large fleet operator adopting the AI Driver is the proof that this is a licensing business. Without one, it is an operator with unusual marketing.

Tokyo. A second city on a different continent, different road rules and different driving culture is the real test of the mapless generalization claim.

Whether the safety driver comes out, and when. TfL creating a driverless authorization would reset the comparison entirely.

Waymo's response. The incumbent has never licensed its stack. If it starts, the vertical-integration thesis has a crack in it.