NVIDIA ships Jetson Thor, the on-robot brain half the industry was waiting for.

NVIDIA announced general availability of Jetson AGX Thor, its Blackwell-based on-robot computer, with a developer kit starting at $3,499. The headline numbers — up to 2,070 FP4 teraflops in a roughly 130-watt envelope, about 7.5x the AI compute of the previous Jetson Orin — matter less than what they enable: running large generative and vision-language-action models locally, on the robot, in real time, rather than piping everything to the cloud. The named early adopters read like a roster of the category's leaders: Figure, Agility, Boston Dynamics, Amazon Robotics, Caterpillar and Meta among them.

The strategic story here is standardization. Just as a common x86 platform let PC makers stop designing silicon and start competing on product, a de facto standard humanoid compute module lets robot builders offload one of their hardest problems. For a buyer, that is quietly reassuring: a fleet built on Thor inherits a well-supported software stack, a large developer community and a predictable upgrade path, instead of being hostage to one startup's bespoke electronics. It lowers the risk of stranding a capital investment on an orphaned architecture.

It also deepens NVIDIA's grip on the entire embodied-AI value chain. The company already sells the training GPUs, the Isaac simulation tooling and, through NVentures, equity in the robot makers themselves — FieldAI and Figure both closed rounds with NVIDIA participation in this same window. Thor closes the loop by putting NVIDIA silicon inside the robot too. The counter-signal to would-be competitors is that the ecosystem's center of gravity is consolidating fast, and differentiating on raw compute is becoming a losing game.

The caveat for buyers is that compute is necessary but not sufficient. Thor gives every vendor access to the same horsepower, which means the real differentiation moves up the stack to data, behavior models and reliability in the field — exactly the layer Boston Dynamics and Figure are racing on. A robot is only as good as the policy running on the chip. But as an enabling milestone, Thor's availability removes a genuine bottleneck, and the timing — days after the Atlas LBM demo — was no accident.