Waabi’s $1B raise puts a hard question to the AV industry: do you really need separate stacks for trucks and robotaxis?

Waabi has raised $1 billion in a Series C and signed a deal with Uber to deploy robotaxis on the ride-hailing platform.

The funding is big. The underlying bet is bigger.

Waabi wants one autonomous driving stack to span long-haul trucks and passenger cars.

That’s an aggressive claim in a sector that has spent years finding new ways to make autonomy look narrower, slower, and more expensive than expected. Plenty of companies have pitched general-purpose self-driving. Very few have had to prove it across vehicle classes with different dynamics, sensor layouts, duty cycles, and regulatory headaches.

The round gives Waabi time to try. It also gives the company unusually strong strategic backing. Uber joined the round alongside NVentures, Volvo Group Venture Capital, Porsche Automobil Holding SE, BlackRock, and BDC Capital’s Thrive Venture Fund. Total funding is now about $1.28 billion.

The investor mix matters. Uber gives Waabi a distribution channel. Volvo gives it a path to factory-integrated vehicle platforms. Nvidia’s venture arm signals that compute is still part of the story, even if Waabi likes to frame its approach as less brute-force than a lot of AV rivals.

The robotaxi move matters

Waabi has mostly been positioned as an autonomous trucking company. It runs pilots in Texas with a human safety driver and has a partnership with Volvo on purpose-built autonomous trucks. A fully driverless highway launch was once targeted for late 2025 and has slipped to “the next few quarters,” according to founder and CEO Raquel Urtasun.

That slip matters. Getting to a real driverless deployment still looks harder than the company’s messaging sometimes implies.

Now Waabi is moving into robotaxis through Uber. That gives it something many AV startups never had: access to riders without having to build its own app, dispatch layer, or consumer operations. Uber, meanwhile, keeps building a multi-partner AV strategy instead of tying itself to one autonomy stack.

The timing works for Uber too. It has just launched Uber AV Labs, a new effort to collect driving data for AV partners using instrumented Uber vehicles. For a company that leans hard on simulation, that kind of targeted data is valuable. You don’t need infinite logs. You need the right logs from the right ugly situations.

Night merges in heavy rain. Occluded left turns. Dense pickup zones with impatient pedestrians and double-parked cars. Those cases help tune sensor models, stress the planner, and expose where behavior still falls apart.

Waabi’s core bet: simulation can replace a lot of fleet grind

Waabi has been consistent on the technical side. It argues that its closed-loop simulator, Waabi World, can take over much more of the training, testing, and validation work than traditional AV programs have managed.

That sounds abstract until you compare it with the old model.

A lot of AV development has depended on expensive road fleets, giant annotation pipelines, hand-picked corner cases, and open-loop evaluation on recorded logs. Open-loop testing is useful, but it has a clear limit. The model never has to deal with the consequences of its own actions. It predicts against past data instead of driving in a system where a bad decision changes the next state.

Closed-loop simulation fixes that specific problem. The stack makes decisions inside the simulator, and those decisions shape what happens next. If the planner takes a bad line into an unprotected turn, the world changes with it. If the prediction module misreads a cyclist, the rest of the system has to live with that mistake.

That matters because autonomous systems are very sensitive to covariate shift. Small policy errors stack up fast. A planner trained on clean logs can look decent in evaluation and still behave badly once its own choices push it off the training distribution.

Waabi says Waabi World builds digital twins from real data, models cameras, lidar, and radar in real time, and generates scenarios automatically to stress the stack. The appeal is obvious enough. Fewer people labeling data. Smaller road fleets. Faster iteration. Lower burn.

But the catch is brutal. The simulator has to be good enough to trust.

The sim-to-real gap is still the hard part

A simulator-first program stands or falls on fidelity. That includes geometry, traffic flow, agent behavior, weather, lighting, and sensor noise. If the camera model handles glare badly, or lidar degradation in rain is off, you end up with a policy that looks polished in simulation and shaky on the road.

Human behavior is just as hard to model. Real drivers hesitate, bluff, wave people through, break norms when it suits them, and do irrational things in batches. A robotaxi in city traffic has to read that mess better than a highway truck does.

That’s one reason the single-stack claim deserves a hard look. Trucks and robotaxis do share core autonomy machinery. Perception still has to identify lanes, objects, lights, and drivable space. Prediction still has to estimate what nearby agents will do. Planning still has to choose maneuvers under constraints. Control still has to turn those maneuvers into steering, braking, and throttle.

The constraints split quickly after that.

A Class 8 truck has different stopping distance, acceleration, visibility, failure modes, and usually a narrower operating domain if it stays mostly on highways. A robotaxi in urban service deals with denser interactions, lower speeds, curbside chaos, pickups and drop-offs, and far more vulnerable road users.

Waabi’s argument is that the world model and decision logic can stay largely unified while vehicle dynamics and control adapt to the platform. That’s plausible. It’s also one of the few paths to capital efficiency in AV that doesn’t sound delusional. Separate stacks for every vehicle type burn cash and engineering focus at a rate investors have lost patience with.

Waymo spent years on both trucking and robotaxis, then shut down freight to focus on passenger service. Waabi is taking the opposite view.

OEM integration matters more than demos

One of the smarter parts of Waabi’s strategy has less to do with AI architecture than with vehicle integration.

Urtasun has been explicit that Waabi wants a vertically integrated, redundant platform built with the OEM. That means factory-installed sensors, compute, power, cooling, actuation redundancy, and a vehicle designed for autonomy from the start.

It’s not glamorous. It does matter.

Retrofitted stacks are fine for prototypes. They’re much worse for scale. Packaging gets awkward. Thermal margins tighten. Power redundancy turns into a systems problem. Serviceability gets worse. So does cost.

If Waabi can carry that OEM-first model from trucks into robotaxis, it has a better shot at scaling operations than companies still bolting hardware onto existing vehicles and calling it a platform.

For engineers, this part maps cleanly to execution. Factory integration changes failure modes, validation, maintenance windows, boot behavior, and cybersecurity posture. It also changes how safety cases get built against standards like ISO 26262, ISO 21448 for SOTIF, UL 4600, and UNECE WP.29 requirements.

A simulator can produce evidence. Regulators still want traceability between simulation results, on-road testing, software changes, and incident response. That work is tedious and unavoidable.

What technical teams should watch

Closed-loop evaluation matters more than pretty benchmark scores

Waabi is right to focus on closed-loop simulation. In autonomy and robotics, open-loop metrics can flatter a model that won’t survive contact with the real world.

The useful question is whether the system recovers after a mediocre decision and deals with the next five seconds getting worse.

Scenario coverage is a tooling problem

If simulation is the engine, scenario generation is a big part of the machinery. Teams need strong case libraries segmented by road type, weather, city, lighting, and known failure families. Coverage metrics matter as much as pass rates.

That’s where Uber’s data collection effort could help. Good simulation programs don’t always need more data. They need specific slices that close specific gaps.

Unified stacks need disciplined interfaces

A shared architecture across trucks and robotaxis only works if the interfaces stay clean. Perception -> prediction -> planning -> control is easy to sketch and easy to corrupt once one vehicle program starts baking assumptions into the wrong layer.

The reusable layer should be the scene representation and decision framework. Vehicle-specific dynamics should stay vehicle-specific.

Capital efficiency is now a technical claim

For a while, AV companies could burn money under the banner of inevitability. That period is over. If Waabi can show fewer on-road miles, fewer annotators, smaller teams, and lower compute demands without giving up safety, competitors are going to get pressed on cost structure.

That doesn’t make safety cheaper. It does mean efficiency now sits inside the engineering argument, not off to the side in finance slides.

The open question

Waabi looks more credible than most young AV companies because the thesis hangs together. It has enough capital to test it, OEM ties, and now a path into robotaxis without having to build a consumer network from scratch.

That still leaves deployment.

The trucking timeline has slipped. Urban robotaxis are harder than highway freight in several ways. Simulation can compress development. It can also hide bad assumptions if the models underneath it are wrong.

This round doesn’t settle the argument. It makes it sharper.

If Waabi can prove that one stack, a simulation-heavy workflow, and tight OEM integration scale across trucks and robotaxis, a lot of AV orthodoxy starts to look expensive and unnecessary. If it can’t, the industry gets another reminder that generality in autonomy is easy to pitch and hard to ship.

What to watch

The funding number does not prove durable demand. It shows investor appetite and gives the company more room to execute. The real test is whether customers keep using the product after pilots, whether margins survive real workloads, and whether the team can turn technical interest into repeatable revenue.