Self-Driving Startups Shift From Science Risk to OEM Deployment

Alex Kendall rejects the simple version of current self-driving optimism. Wayve has more than $2 billion in capital, investors and partners including Nvidia, Qualcomm, Arm, AMD, Uber, Nissan, Mercedes, Stellantis, and Microsoft, and deployment programs with automakers and ride-hail partners. But Kendall does not describe the problem as solved.

The distinction matters. Kendall says the scientific question — whether end-to-end AI can learn a generalizable driving policy — has largely moved behind Wayve. The remaining problem is whether that policy can be integrated into production vehicles, scaled through large OEM relationships, validated across broad domains, and launched under regulatory regimes that accept the product.

That is narrower than saying “self-driving is solved,” but still aggressive. Kendall says Wayve has moved “from science risk” into “engineering execution risk” and “product integration and deployment risk.” He divides the path by autonomy level. For hands-off driving, he says end-to-end stacks from Wayve and Tesla have shown they can scale globally. Wayve drove in 500 cities around the world last year, and Kendall points to Tesla’s announced $1.5 billion in annual revenue from its driver-assistance product as evidence that consumers will pay for this kind of functionality.

The harder step is from hands-off to eyes-off or driverless driving — L3 and L4 systems where the human is no longer continuously responsible. Kendall says there is still a performance gap between systems like Tesla’s or Wayve’s current stack and “general purpose driverless” operation of the sort Waymo has demonstrated in geofenced areas. His point is not that the gap is trivial. It is that the remaining work follows a predictable scaling curve: more data, more compute, algorithmic improvement, vehicle integration, and scaled validation.

Wayve’s path is built around licensing what Kendall calls “the intelligence layer” to fleets and automakers rather than manufacturing cars or owning fleets. He contrasts three self-driving business models. Tesla builds its own cars and is therefore limited to its own brand. Waymo builds its own fleet city by city, which Kendall describes as an expensive, high-capex endeavor. Wayve licenses to fleets and automakers, a model he says depends on having a flexible, generalizable AI driver.

That licensing strategy is now tied to named programs. Kendall says Nissan announced plans to bring Wayve’s approach into its consumer vehicle lineup, and separately says Nissan later announced it would bring the technology to 90% of its vehicles. Nissan builds about 3 million cars per year; Alex Wilhelm calculates the implicated volume at roughly 2.7 million vehicles. Kendall calls that “double the cars Tesla builds a year,” and emphasizes that Nissan is only one partner.

Kendall also says Wayve plans supervised robotaxi trials this year in London, Tokyo, and 10 other cities on Uber. On consumer vehicles, he refers to buying cars from partners such as Nissan “from next year,” while also describing Nissan’s plan to bring this kind of technology to vehicles from financial year 2027 and cautioning that Wayve is “still not there yet.” His commercial argument is that OEMs want the same autonomy partner across L2, L3, and L4 because it improves speed, efficiency, data leverage, and integration.

Both Wayve and Waabi center their strategies on world models, but they define the concept with enough specificity to avoid treating it as a generic AI slogan.

For Alex Kendall, a world model understands the state of the world, the action taken in that world, and how the world evolves afterward. In driving, that forces the system to learn what matters. A car does not need to represent clouds in the sky with the same priority as road lines, curbs, traffic signals, or objects that might intersect with its path. By learning to predict the future, Kendall says, the model learns useful representations in an unsupervised way.

The second use is simulation. A world model can replay or generate scenarios so a driving policy can train, fail, and be validated without requiring every mile to be driven in the real world. Wilhelm describes it in deliberately simple terms as a high-end virtual environment in which a self-driving system can face obstacles, traffic, weather, location-specific rules, and countless variations. Kendall accepts the analogy but expands it: Wayve thinks of it partly like the hippocampus in humans, where experiences are replayed during daydreaming or sleep to reinforce motor learning.

Kendall says Wayve’s world models have scaled in parameter count, data, and algorithmic capability. The company trains on “hundreds of petabytes” of data across internet-scale data, dash cams, and automaker data from more than a dozen companies. The models now understand not only video but also radar and lidar, and can handle multiple sensors across a vehicle. They are also controllable: Wayve can prompt or re-simulate real-world events, run adversarial tests, and try to make its car make mistakes inside the world model so the system can learn from them.

Raquel Urtasun describes Waabi’s version of the same thesis as two linked requirements. First, autonomy systems must generalize and exhibit human-like reasoning capabilities. Second, because data is as important as the model, the company needs representations of the world that enable realistic simulation of safety-critical situations “with no consequences.”

Urtasun stresses controllability. For physical AI, she says, a world model is not just a generator of realistic interactive worlds; it must allow the developer to control what is generated. That separates world models for self-driving from systems built to create “pretty movies” or “cool video games.” In Waabi’s case, the world model, simulator, and autonomy system were designed from the beginning as a physical AI platform for multiple use cases.

World models also change the economics of testing, according to Urtasun. They can allow companies to run training and testing in parallel in the cloud, bypassing “many years” or even “centuries” of real-world experimentation. If the world model is efficient, it can also reduce the engineering spend that would otherwise accumulate from large teams manually integrating and validating systems over time. For Waabi, that translates into faster time to market, better safety, and a better understanding of the system’s safety.

Both CEOs describe safety-critical data as a central bottleneck. Kendall says robotics and self-driving do not have the equivalent of internet-scale text for chatbots. The hardest problem is getting data, especially safety-critical data, and proving the system safe. He describes the field as an arms race between learning a driving policy and learning a simulator: if one side is solved, it helps solve the other.

Alex Kendall and Raquel Urtasun both resist the simplified version of the camera-versus-lidar fight, though they land in different places.

Wayve wants to be sensor-agnostic. Kendall says the company supports cameras, radar, and lidar because different products will benefit from different configurations. A camera-only system may be appropriate for some products; robotaxis may be better served by camera, radar, and lidar. The key, in his view, is not ideological purity around one sensor type but whether the system can understand what each sensor architecture can and cannot see.

The world model is central to that argument. If a vehicle’s sensors cannot observe part of a scene, the system will not be able to predict the future in the same way. Kendall says that makes the limits of a sensor architecture learnable. In fog, a camera-only vehicle may struggle; a vehicle with radar may do better. Wayve can simulate not just rain or snow, but a specific vehicle form factor, sensor array, and weather condition.

There is still a minimum bar for hands-off, eyes-off, or driverless safety. Kendall says a camera-only system can in principle reach all levels if it is good enough, but radar or other modalities may make it faster and more efficient to get there. In current robotaxi work with partners, he says camera, radar, and lidar are typically preferred. But he draws a line between today’s mass-market automotive-grade sensors and the bespoke, expensive spinning lidar units associated with earlier self-driving programs.

Urtasun frames the same area as a tradeoff among safety, bill of materials cost, and the cost of creating high-definition maps. If a company can build HD maps efficiently and robustly, she says, the question should not be whether maps are philosophically scalable. It should be whether they add safety. Waabi’s answer is yes: if the maps can be generated in a scalable way using AI, then they should be used as an additional safety layer.

At the same time, Urtasun says the self-driving vehicle must be able to react and drive if the maps are wrong or out of date. HD maps are not a crutch that makes the system helpless without them; they are an extra layer of safety when they can be built and maintained efficiently.

This is one of the sharper technical differences between the two companies’ public positions. Wayve emphasizes generalized driving across vehicles and geographies, with support for whichever sensors a product needs. Waabi emphasizes that maps can be part of a scalable safety architecture if AI makes them cheap and robust enough. Neither speaker treats a single sensor or mapping choice as sufficient by itself.

Alex Kendall’s most consequential business point is that consumer vehicles dwarf robotaxis in scale. He says there are fewer than 10,000 robotaxis in the world today, while roughly 100 million vehicles are produced annually, including 50 million to 60 million consumer cars. Advanced driver-assistance systems are already present in some form in about 15% of the market, but he characterizes much of that as rudimentary highway lane-keeping or similar functionality. Outside China, he says, the “full self-driving experience” is mostly Tesla and still a small fraction of the market.

From that base, Kendall predicts the market moves “from like nothing to everything” over the next few years. Luxury manufacturers are already adding the compute and sensing required for more advanced systems, including Nvidia or Qualcomm GPUs and surround sensing. Volume manufacturers are beginning to follow. Kendall says Nissan plans to bring this kind of technology to vehicles from financial year 2027.

His steady-state view is that every vehicle will eventually be capable of this. If eyes-off driving is available through a low monthly subscription, he expects it to become a standard expectation. He goes further: a manufacturer selling a car without such technology may see demand “fall off a cliff.” Wilhelm suggests that the lowest-end vehicles might be exceptions, but Kendall pushes back. Active braking systems are already a regulatory requirement, he says, and over time autonomy will be so important for road safety that even basic cars will need it.

The consumer pricing model remains unsettled. Kendall says the consumer will likely pay the manufacturer, which will pass economics through to Wayve. Automakers are exploring several approaches: bundling autonomy into the car, treating it like a seatbelt-style expected feature, charging a one-time fee, charging a subscription, or offering a free trial followed by a subscription. He expects the industry to move toward subscriptions because performance will improve over time through over-the-air updates, and because L3 and L4 systems create ongoing insurance costs for the manufacturer or operator.

The liability structure depends on autonomy level. In a properly implemented hands-off system, Kendall says the driver should remain liable. In eyes-off or driverless systems, liability shifts to the manufacturer or operator, with insured and contracted liability flowing through the broader ecosystem. Wayve does not plan to build its own insurance product.

If autonomy becomes a recurring software layer on tens of millions of vehicles, Wayve’s model would be a high-margin software business across a broad vehicle base rather than a fleet operation constrained by local deployment. Kendall says Wayve has enough capital to reach deployment, free-cash-flow positivity, and “escape velocity,” though he does not rule out future raises to accelerate or expand into other verticals.

Alex Kendall treats validation as engineering work, not cleanup after launch. To move from hands-off to eyes-off or driverless, Wayve needs vehicles with the right infrastructure, scaled AI performance, and validation activities that prove the system is safer than a safe and competent human driver before launch. Regulatory sign-off follows from that work.

Kendall says regulators have in some cases moved ahead of product readiness. In the U.S., he says some states allow deployment and some do not, but a market exists. Outside the U.S., he says the UN recently put in place a legal pathway for L3 and L4 driving that covers basically every country except the U.S. and China. He also says Wayve co-chairs the industry committee for UN autonomy regulations. The claim is a pathway, not a declaration that global regulation is finished or uniform.

His liability answer is conditional. It depends on autonomy level, regulatory environment, and the commercial contracts among the companies that bring the product together. At a high level, he says, a hands-off system leaves liability with the driver if implemented correctly. For eyes-off or driverless systems, the manufacturer or operator remains liable, with insured and contracted liability flowing through parts of the ecosystem.

Raquel Urtasun makes the related point from Waabi’s side through the company’s insistence on OEM platforms. Waabi does not plan to become an OEM or rely on retrofitting as its path to market. Urtasun says partnering with manufacturers that have spent a century building vehicles is the right path to safe self-driving. For robotaxis, Waabi expects Uber to provide the market component, an OEM to provide the redundant vehicle platform, and Waabi to provide the autonomy technology.

The commercial product includes the hardware platform, the safety case, the regulatory pathway, the liability allocation, and the contracting structure that determines who operates the vehicle and who bears risk when no human driver is responsible.

Raquel Urtasun says Waabi began in trucking and is now expanding into robotaxis, but she does not describe the company as a trucking-only autonomy stack. From day one, Waabi’s world model, simulator, and autonomy system were designed as a physical AI platform for multiple use cases.

The key point is that the company does not need to fork the stack for trucks and robotaxis. Urtasun says it is “the same brain” and the same simulator and world model for both. That does not mean the system ignores vehicle differences. An autonomy system must know whether it is driving an 80,000-pound truck or a robotaxi; the driving style cannot be identical. But the core capabilities — perceiving and understanding the world in 4D, reasoning, and acting — are common.

This is also Waabi’s answer to specialization. Human drivers do not swap brains when they move between vehicles, and Urtasun argues that self-driving systems should not require separate technology stacks for every vehicle type. Specialization may make sense for more distinct skills, but within driving, Waabi wants additive learning across programs. Trucking should accelerate robotaxis, and robotaxis should accelerate trucking.

Commercially, Waabi is further along in trucking. Since 2023, Urtasun says the company has been conducting commercial operations with major shippers and carriers in North America. Waabi has a partnership with Uber Freight for “billions of miles” of deployment on the Uber Freight network and operates a double-digit fleet of self-driving trucks. Its commercialization path runs through the OEM. When Wilhelm asks where Waabi is today, Urtasun says public comments from its partner indicated launch was “quarters away,” and that 2027 “would be hundreds of trucks.” She presents that as a way to understand the commercialization path, not as a standalone forecast detached from partner statements.

Waabi’s critique of some trucking deployment models is product-focused. The industry, Urtasun says, often pursued hub-to-hub autonomy: trucks drive autonomously between highway-adjacent hubs, while humans handle the start and end of each trip. That approach simplified the autonomy problem because surface streets are harder. But she argues it created a product customers do not want. The drayage on both ends can cost roughly $0.60 to $0.80 per mile depending on length of haul, which can break the economics. Waabi instead invested in generalized surface-street capability so trucks can go to the customer’s door.

That point explains why Urtasun does not treat earlier on-road operations as the only measure of progress. Waabi, she says, took longer to go on-road because it invested in foundational technology first. The payoff, in her view, is a better product when commercialization begins.

Waabi and Wayve both want to be technology providers, not vehicle manufacturers or fleet owners. But Raquel Urtasun is more explicit about the pricing unit.

Waabi’s model is “Driver-as-a-Service” for both trucking and robotaxis. The company does not plan to own and operate trucks or robotaxis. In robotaxis, Uber provides the marketplace. An OEM will provide the redundant vehicle platform into which Waabi integrates, though the OEM for the Uber robotaxi partnership has not been announced. In trucking, the payer may vary: it could be the OEM, or it could be a direct customer such as a shipper, depending on the operating structure. Volvo Autonomous Solutions, Urtasun notes, plans to operate some self-driving vehicles through a Transportation-as-a-Service business unit.

For Waabi, the business model remains the same regardless of who pays directly. The fee is primarily per-mile, making it recurring and usage-based. Urtasun says blended structures are possible, but the central component is per-mile pricing. Alex Wilhelm notes that this aligns incentives: the more the system is used and the more value the customer gets, the more Waabi earns. Urtasun agrees that it keeps everyone “on the same page.”

This model depends on Waabi remaining capital efficient. Urtasun says the company raised more than $1 billion in its latest round not because it needed to spend the money on rolling hardware, but because the capital gives it stability, lets it make long-term investments, and allows it to pursue trucking leadership and robotaxis without compromise. She calls a billion dollars “infinite money for us” because of Waabi’s capital efficiency.

The raise is also strategic in a market where delays can occur across adoption, scaling, or partner timelines. Urtasun says the capital makes Waabi robust to ecosystem delays and removes quarterly pressure to prove short-term progress in ways that could compromise the safe path to market. The company will remain a technology provider and does not believe retrofitting or becoming an OEM is the right or safe path.

Alex Kendall makes a parallel point from Wayve’s side. Wayve has more than $2 billion in capital and believes it has what it needs to deploy, reach free-cash-flow positivity, and give automakers confidence in decade-long relationships. The company’s capital is being used to integrate, validate, and deploy across OEM programs rather than to build a proprietary car fleet.

Raquel Urtasun gives a broader explanation for why self-driving may be entering a deployment phase now. Several “tectonic plates,” she says, are converging at the same time.

First, the hardware and OEM platforms are ready. After a decade of investment by trucking and passenger-car manufacturers, redundant platforms are now becoming available for safe, scalable products. Second, regulatory frameworks are evolving to enable deployment. Third, consumers appear more willing to use robotaxis. Urtasun points to Waymo’s deployments as evidence that people can learn to trust the technology when they experience it. Alex Wilhelm adds that he has been surprised by how quickly ordinary consumers have adopted Waymo.

Trucking has a different demand profile. Urtasun calls it a “no-brainer” because of driver shortages, the cost of human drivers, and persistent safety issues. If the product solves customer pain points, she says, the case for adoption is clear.

The fourth convergence is AI itself. Urtasun says next-generation AI changes the product companies can build, especially compared with earlier AV 1.0 systems that could deploy but struggled to scale beyond small operational design domains. With stronger reasoning and generalization from few examples, she argues, companies can address the long tail more quickly and expand geographically and across use cases.

Alex Kendall makes a similar but more technically cautious analogy to large language model scaling. He says Wayve’s remaining AI performance gap should follow a predictable curve involving data, compute, and algorithmic innovation. But he also emphasizes the differences from LLMs. Self-driving requires real-time embodied inference onboard a vehicle, under tighter compute constraints, with safety-critical outputs, larger-dimensional sensor data, and uncertainty-aware decision-making. A chatbot can hallucinate; a self-driving car cannot.

Self-driving is not waiting on a single missing invention. The claimed shift is toward execution across AI performance, redundant hardware, OEM integration, validation, regulation, consumer trust, and business model.

One of the sharpest disagreements in the market is implicit rather than personal: whether advanced driver-assistance systems naturally evolve into driverless autonomy. Raquel Urtasun says that assumption has misled the industry.

For a decade, she says, many people believed autonomy would progress linearly: L2 first, then L3, then L4. It seems intuitive because each level appears to add one increment of capability. But Urtasun says her career and Waymo’s experience convinced her that this is not necessarily the fastest path and may not be a path at all.

Her argument is that L2 and L4 are different safety problems. A performant L2 product can be evaluated in terms of how well it drives and how rarely it needs intervention, because the human remains responsible. That is not an adequate L4 metric. In L4, there is no human fallback. The system must be built natively for that safety case.

“You need to either build level four technology or you build level two technology,” she says. That, in her view, separates Waabi from some other end-to-end companies.

This is why Urtasun is cautious when Alex Wilhelm asks whether he will be able to buy a personally owned L4 or L5 self-driving car in the next three years. Experiencing robotaxis at scale within that window, she says, is plausible. Personally owned L4 or L5 vehicles in that timeframe are harder.

Alex Kendall is more bullish on consumer vehicles entering the market soon, including through partners such as Nissan, but he also distinguishes between hands-off and eyes-off or driverless. He does not claim the entire market reaches general-purpose driverless immediately. Instead, he expects improving experiences over time, premium introductions of eyes-off technology, and eventual flow-down across the market. His prediction that demand for cars without autonomy will fall off a cliff is a five-year market view, not a claim that every consumer car becomes L4 in three years.

The practical difference is that Wayve is building across L2, L3, and L4 programs for OEMs that want one partner across the spectrum, while Waabi insists that L4 must be architected as L4 from the beginning. Both companies use end-to-end AI and world models, but they position the path through autonomy levels differently.

Uber appears in both companies’ commercialization stories, but in different forms.

For Alex Kendall, Uber is a partner in Wayve’s planned supervised robotaxi trials, which he says will begin this year in London, Tokyo, and 10 other cities. Uber is one of the investors and partners backing Wayve’s licensing strategy. Wayve’s broader argument is that demand aggregators like Uber and supply-side manufacturers like Nissan are the right channels because they already have customers, vehicles, and operating scale.

For Raquel Urtasun, Uber is central on both freight and robotaxis. Waabi has a partnership with Uber Freight for billions of miles of deployment, and its robotaxi partnership with Uber is larger than Wilhelm initially states. He describes it as “up to 25,000 robotaxis,” but Urtasun corrects him: it is not up to 25,000; it is over 25,000, or a minimum of 25,000. The difference matters because it changes the partnership from a ceiling to a floor.

Waabi has not announced the OEM that will provide the redundant vehicle platform for that robotaxi partnership. Urtasun says there are now a few OEMs with redundant platforms ready, and she describes the ecosystem as excited to partner, but she declines to name the manufacturer.

Wilhelm asks whether Uber has tried to buy Waabi, given Urtasun’s four years at Uber ATG and Waabi’s partnerships across Uber Freight and Uber’s ride-hailing side. Urtasun answers more broadly: since Waabi’s founding, many people have tried to buy the company. Her goal, she says, is to build a physical AI powerhouse that transforms the world.

In both stories, Uber functions less as a full-stack autonomy company and more as a marketplace and deployment partner. That is consistent with both Wayve’s and Waabi’s desire to remain technology providers rather than fleet owners. Uber supplies demand; OEMs supply vehicles; autonomy companies supply the driver.

Alex Kendall says Wayve started with consumer vehicles because he considers them the hardest mobility application. Consumer cars must run on hundreds of dollars of hardware, work everywhere, and handle a wide range of vehicle lines. Nissan alone has 60 car lines, he says. Wayve also began learning in London, which Kendall describes as one of the hardest driving environments. The logic was that solving the hardest scalable automotive problem would force the company to build the most general technology.

Kendall says the stack can transfer to any robotics application with wheels. Wayve has run proofs of concept in sidewalk delivery, trucking, mining, and warehouse logistics. With a small amount of domain data added to the foundation model, he says, the system can learn behaviors in those domains. GAIA, Wayve’s simulator, can also adapt. The driving policy, reinforcement learning stack, and simulation stack transfer with data.

But Kendall separates mobility from manipulation. He says many people are excited about manipulation robotics, but he expects mobility to arrive much sooner. At robotics conferences during his PhD, mobility and manipulation were treated as separate communities; over time, he expects them to converge into the same AI. Today, however, mobility has platforms, sensors, compute, software-defined vehicles, and automotive manufacturing scale. Manipulation still needs platforms at scale and data. Wayve may adapt to manipulation in a few-shot setting using what it learns from mobility, but Kendall expects manipulation to come second.

Raquel Urtasun is more expansive in the final exchange. When Wilhelm suggests that Waabi’s world-model approach could apply inside buildings, factories, delivery bots, and other robotics domains, she adds “humanoids” and “everywhere.” Her stated ambition is to build a physical AI powerhouse, not a trucking or robotaxi point solution.

The strategic bet is that driving supplies the data, validation discipline, and commercial pressure needed to build general physical AI. Automotive and trucking are not side projects for robotics; they are the first markets with enough scale, urgency, and infrastructure to make the technology real.