The $8.6B Self-Driving AI Backed by Nvidia and Uber | Alex Kendall, Wayve
Summary
Wayve’s decade-long contrarian bet is that autonomous driving scales through one general-purpose AI driver, not fleets dependent on HD maps, retrofit sensors, and infrastructure. Kendall began with $1.5 million while the industry pursued an “AV 1.0” paradigm he says required $100 billion-plus; Wayve’s resulting model is now “capable of driving any vehicle anywhere,” though it has yet to launch a driverless service.
Consumer vehicles, not robotaxis alone, are the scale opportunity Wayve is building toward. Roughly 100 million cars are produced annually versus “less than 10,000 robotaxis in the world,” and Kendall argues that native integration into $30,000-$50,000 mass-market vehicles could create tens of millions of vehicle opportunities before every vehicle eventually becomes capable of driverless operation.
The bottleneck has shifted from proving the model can drive to validating, integrating, and commercially deploying it. Kendall says “performance is no longer the critical path,” while carefully adding that it still has “orders of magnitude” of headroom; near-term constraints include safety evaluation, regulatory proof, multi-year automotive sales cycles, vehicle integration, and deployment operations.
Wayve’s strongest technical evidence is zero-shot operation across more than 500 cities and over 10 different cars spanning electric vehicles, vans, and SUVs, with no prior training data from more than half the cities. Tests ranged from 22-hour darkness north of the Arctic Circle to a Tokyo typhoon that shut trains and buses; during a day of journalist drives, Kendall reports “no disengagements, perfect autonomy the whole day,” but resists calling Wayve superhuman without scaled safety evidence.
Language became a performance input, not merely a conversational interface. A 2021-22 project trained one vision-language-action model to “see the world, drive a car, and understand language”; language pre-training helped it act on a flashing oncoming car yielding, while also enabling explanations and selectable driving styles for consumers and automotive brands.
Kendall says the AV industry, including Wayve, should reject the trolley problem’s forced binary and instead preserve “always a third good option.” That minimal-risk maneuver reduces kinetic energy, stays predictable, and stops or pulls over; Kendall says Wayve has had no incidents since beginning operations in 2018, while emphasizing that difficult failures usually combine darkness, speed, adversarial driving, bad weather, and perhaps sensor loss.
Kendall’s sharpest retrospective lesson is that deep-learning execution depends more on systems than novelty: “It’s 1% algorithms, 99% infra.” Clean data, reliable learning loops, measurement, evaluation, simulation, and robust vehicle platforms mattered more than “shiny sexy algorithmic innovation”; Wayve spent millions and months on approaches it later deleted after finding something better.
Deep dive
Not yet available upstream; scheduled sync will retry.