Here’s my read after looking at what the robotaxi field is actually building on in 2026: the interesting story isn’t the cars, it’s that the companies competing hardest against each other are increasingly buying the same toolchain.
NVIDIA has laid out a full-stack robotaxi platform described as a three-computer architecture — DGX systems for training the AI models, Omniverse and Cosmos running on RTX PRO servers for simulation and validation, and DRIVE for the compute that actually rides in the vehicle. Train, test, drive. That’s the whole pitch, and the leading robotaxi companies are adopting it.
Meanwhile the deployment side is moving. Uber is scaling its fleet. Waymo is expanding coverage. Tesla plans to launch its own autonomous robotaxi service. A long list of companies are testing in China. Different business models, different cities, different regulatory headaches — and a shared dependency underneath a lot of it.
Why the middle computer is the one that matters
If you review developer tools for a living, you learn to ignore the flashiest layer and look at where the painful work lives. For autonomy, that’s not the in-car chip and it’s not the training cluster. Everyone has a story about training compute. The hard part is validation.
You cannot road-test your way to confidence on rare events. The interesting failures are the ones that happen once every few million miles: the pedestrian who steps out from behind a delivery van, the construction cone pattern that makes no sense, the four-way stop where every human involved is wrong. Simulation is how you generate those cases on demand and run them ten thousand times with small variations.
Which is why the Omniverse and Cosmos piece is the part of this stack I’d scrutinize first. It’s the layer that decides whether your safety case is a document or an argument. It’s also the layer that’s hardest to swap out later, because your entire regression suite, your scenario library, and your internal definition of “good enough” get built inside it.
The honest lock-in conversation
I’m not going to pretend a vertically integrated stack is a bad deal. It usually isn’t, early on. When training, simulation, and in-vehicle inference come from one vendor, the model you validate is closer to the model you ship, and the tooling between stages is somebody else’s problem. For a team trying to get cars on real streets against a regulatory clock, that’s worth real money.
The tradeoff arrives later, and it’s the same one you get with any full-stack toolkit:
- Pricing power sits with the vendor. When the training cluster, the simulation servers, and the car computer are all line items from one supplier, you have very little negotiating room.
- Your safety artifacts are format-bound. Scenario libraries and validation results don’t port cleanly. Moving stacks means rebuilding trust from close to zero.
- Technical differentiation compresses. If your closest competitor trains on the same hardware, validates in the same simulator, and infers on the same in-car module, your edge has to come from data, operations, and city-by-city execution rather than from the stack itself.
That last point is the one I’d underline. It suggests the robotaxi race gets decided on unglamorous things — fleet utilization, depot logistics, remote assistance staffing, permit relationships, how fast you recover when a vehicle blocks an intersection at rush hour. Uber scaling a fleet and Waymo expanding coverage are operational achievements as much as technical ones.
What I’d want to see before calling it solid
I can’t verify from what’s public how deep each company’s adoption goes. “Adopting a full-stack platform” covers a lot of ground, from buying training compute to running the vendor’s software end to end. Those are very different commitments, and the difference matters for anyone trying to judge who’s actually locked in.
Three questions I’d ask any vendor selling this shape of product:
- Can I run validation against a simulator I didn’t buy from you, and will the results be comparable?
- What happens to my scenario library if I change in-vehicle hardware?
- How much of the “full stack” is genuinely integrated versus co-marketed?
None of that is a knock on the platform. A single supplier making the training-to-validation-to-vehicle path work is a legitimate engineering feat, and the fact that serious operators are standardizing on it tells you the alternative — assembling this yourself — is harder than it looks.
But it does change what a robotaxi company is. Less a builder of autonomy from first principles, more an operator running a very expensive vehicle service on a stack it licenses. That’s a normal maturity pattern for a technology, and it’s usually the point where the money starts working. It’s also the point where you should stop asking who has the best model and start asking who has the best depot.
🕒 Published: