Wayve CEO Says Tesla Is Following His AI-First Approach to Self-Driving

Alex Kendall says Elon Musk once told him end-to-end neural nets "would never work" for cars. That conversation happened years ago. Now Tesla's FSD v12 ships with exactly that architecture. Kendall couldn't resist the jab: "I told you so," he told Bloomberg this week.

The irony runs deeper. Wayve, a London startup backed by SoftBank, Microsoft and Nvidia, started with the pure AI approach Tesla just adopted. While Waymo maps every curb in Phoenix and Cruise burns cash on human safety drivers, Wayve trains one model on London, Boston and Pittsburgh data, then drops it into new cities cold. Same neural net, no hand-coded rules.

Kendall frames this as autonomous driving 2.0. Version 1.0 was robotics: lidar, HD maps, explicit programming. Version 2.0 is pure machine learning. The car watches YouTube-scale driving footage, learns to predict what humans do next, then acts. No ontology of stop signs. Just pattern matching at superhuman resolution.

Most self-driving stacks break the world into pieces: perception finds objects, prediction guesses their paths, planning picks a route, control steers. Each module speaks a different language. Wayve bolts these into one transformer that eats pixels and spits steering angles.

Training happens on 1,000+ cars already on UK roads. They drive normally, collecting edge cases. When the AI hesitates at a tricky roundabout, human intervention teaches it the right move. The model updates overnight. Next morning, every Wayve car handles that roundabout better.

This approach scales differently. Traditional stacks need fresh maps for each city. Wayve's neural net generalizes. Kendall claims they can deploy in a new city with 10x less data than Waymo needs for a single neighborhood. The key is diversity: train on chaotic London traffic, and Boston feels easy.

While Waymo builds a taxi empire and Tesla sells premium cars, Wayve wants to be the Android of autonomy. Kendall sees licensing as the fastest path to ubiquity. Car makers get Wayve's AI as software. No custom sensors, no mapping crews, just download the model.

This flips the economics. Robotaxis need 100% reliability and regulatory approval in every city. Licensed AI only needs to beat human drivers. That's a lower bar, and regulators move faster when existing automakers ask permission instead of new mobility companies.

SoftBank's $1.05B Series C, announced this week, funds this licensing push. The money scales cloud training and builds sales teams for Detroit, Seoul and Munich. Kendall won't name deals yet, but says conversations with major OEMs are "very advanced."

Kendall positions Wayve between two extremes. Tesla's approach is pure but hardware-limited. Their cameras see fine in California sun, struggle in London fog. Waymo's approach is safe but brittle. Hand-coded rules break when construction crews repaint lanes.

Wayve splits the difference. They use cameras plus cheap radar, no lidar. The AI learns weather patterns instead of programming around them. In snow, the model learned that tire tracks show where pavement actually sits. No engineer wrote that rule. The data did.

The real bet is computational. Wayve trains on thousands of GPUs, then compresses the model to run on a single car's computer. It's the same trick that made ChatGPT possible: massive training, efficient inference. Tesla now chases this balance. Waymo doesn't.

UK regulators wrote the first autonomous vehicle law with Wayve's help. The Automated Vehicles Act, passed last year, creates a path for AI systems that learn continuously. US rules lag, but Europe follows the UK's lead. That's why Wayve stayed in London instead of moving to Silicon Valley.

The licensing model sidesteps another problem. When an OEM's Wayve-powered car crashes, liability flows to the automaker, not the AI startup. Traditional robotaxi companies own every accident. Kendall calls this "the Intel model" - they make the chip, Dell owns the support calls.

China's the real prize. Local regulations favor domestic players, but Wayve's software-only approach could slip through partnerships. No factory footprint, no data center, just code. Kendall hints they're exploring joint ventures with Chinese automakers already shipping to Europe.

Wayve plans 10,000 cars on UK roads by 2027. That's tiny compared to Tesla's fleet, but each car trains the global model. Kendall's target: make Wayve AI standard on mid-tier cars by 2030. Think Honda Accord, not Mercedes S-Class.

The technical roadmap points to full autonomy, but staged carefully. First: highway driving with human oversight. Then: complex city centers. Finally: door-to-door trips with no safety driver. Each stage unlocks new licensing revenue from different car segments.

The wild card is compute cost. Training these models burns millions in GPU time. Wayve's betting that falling hardware prices and better algorithms keep them ahead of traditional stacks. If they're right, every car gets safer as the cloud model improves. If they're wrong, Waymo's mapping trucks start looking cheap.

Wayve's story mirrors what happened in language models. Rule-based translation died when neural nets learned to predict next words. Computer vision followed. Now it's driving's turn. The pattern is consistent: when AI eats a domain, it starts with weird edge cases, then becomes obviously better.

Kendall's insight isn't technical. It's strategic. Don't fight the old rules. Change the game so old rules don't apply. That's why Musk had to pivot. That's why Waymo might need to license competitors' AI. The companies still writing if-then statements for stop signs are already obsolete. They just don't know it yet.