Robotaxis are settling into partnerships, not winner-takes-all stacks

Two robotaxi launches happened within a day of each other, and between them they set out the only two industry structures currently on offer. Tesla opened Cybercab rides to the public in Austin on 4 September. The vehicle is purpose-built with no steering wheel and no pedals, runs Tesla's own software on Tesla's own hardware with camera-only sensing, and is dispatched through Tesla's own network. At least 51 have been registered in Texas since production began in early 2026. Tesla owns every layer. Uber and Wayve began carrying passengers in London on 3 September with 15 vehicles. The cars are Ford Mustang Mach-E electric vehicles, the driving system is Wayve's, the demand comes through the ordinary Uber app where riders requesting UberX, Uber Electric or Uber Comfort may be matched to one at no extra cost, and a human safety supervisor remains behind the wheel because the UK pilot framework requires it. Three companies, three layers. The vertically integrated model has real advantages and they are not trivial. Controlling the vehicle lets you remove the controls and design for cost, which is the whole economic argument for a purpose-built two-seater. Controlling the software and the hardware together removes integration overhead. Controlling the network means no commission to anyone. What it costs is optionality. Every market Tesla wants to enter, it enters alone: its own regulatory approvals, its own fleet, its own local operations, its own demand generation from zero. In ride-hailing, demand generation from zero is expensive and slow, which is why almost nobody has successfully entered a city against an incumbent network. The partnership model inverts that. Wayve supplies driving capability and has said it plans to expand to Tokyo, where it is working with Uber and Nissan, and then to a further ten cities with Uber. Uber supplies demand, in cities where it already has riders, drivers, regulatory relationships and a brand. The vehicle manufacturer supplies vehicles. Each participant does the part it is best at and none has to build the others. For Uber specifically this is the strategy it has been assembling for years across multiple autonomy partners: remain the demand layer, let others own the driving stack, and add autonomous capacity to the network without owning it. The company's exposure is then not to which autonomy company wins, but to whether autonomy arrives at all. The reason the partnership structure is likely to dominate is that the layers have genuinely different economics. Building a driving system is a research problem with enormous fixed costs and near-zero marginal cost per city once it generalises. Manufacturing vehicles is a capital-intensive industrial business with thin margins. Operating a demand network is a marketplace business with strong local network effects and weak global ones. Those three are rarely good at each other, and a company trying to do all three is subsidising two of them. Wayve's technical position sharpens the point. Its approach is built around learning to drive in a way intended to generalise to roads the system has not seen, rather than relying on high-definition prior maps of every street. If that holds, the cost of entering a new city falls sharply, because the expensive validation and mapping work does not have to be redone everywhere. A driving system that generalises is a supplier to many networks. A driving system that requires per-city mapping is closer to a local operator. The counterargument for integration is coordination. When the vehicle, the software and the network are one company, the vehicle can be designed around the service rather than adapted to it, and there is no negotiation over data, liability or economics. Tesla's ability to build a car without controls is a direct consequence of owning the stack, and no partnership would have produced it. Neither structure has yet demonstrated the thing that matters. London still has a safety supervisor in every vehicle, because the UK framework requires approval to operate without one. Austin has dozens of vehicles in one permissive jurisdiction. Both are pilots. For the Gulf the structural question is the practical one. Saudi Arabia and the UAE have both been building toward autonomous transport, and the choice a city makes is between hosting an integrated operator, which brings a complete service and takes the whole relationship, and assembling a partnership, which requires more coordination and leaves more of the value and the operational knowledge locally. HUMAIN and Applied Intuition announced work on a national autonomous trucking network, which is a partnership structure applied to freight rather than passengers. The evidence that would resolve which model works is the same for both: a vehicle carrying a passenger with no supervisor and no controls, in a second city, with published intervention data. Tesla has the vehicle and one city. Uber and Wayve have the cities and the supervisor. Whoever