A four-billion-parameter world model, running in real time, on the smallest Jetson boards, fast enough for a robot's control loop. That's Cosmos 3 Edge, and NVIDIA's own line about it is the tell.
It delivers "higher speed and accuracy than similarly sized open models on the same Jetson hardware."
Read that phrase twice. The benchmark isn't Cosmos versus the field. It's Cosmos versus other models — evaluated while already sitting on NVIDIA silicon. The comparison is scoped so that the only variable is which model you run, never which platform you run it on.
The comparison is scoped so that the only variable is which model you run, never which platform you run it on.
That framing is the whole story.
Last week the thread was NVIDIA anchoring physical AI to its own stack — Cosmos, GR00T, Isaac, all codesigned in the cloud. This week the same architecture reaches the edge, with new Jetson Thor modules and the Vera Rubin production ramp closing the loop from training rack to robot.
Start with the seductive part, because it is genuinely impressive.
The Jetson T3000 delivers up to 865 FP4 TFLOPS from a Blackwell GPU with 1,536 CUDA cores, at up to 70 watts. Blackwell-class inference in a module you can bolt to a robot arm.
The marquee names are already onboard. Jetson AGX Thor is, per NVIDIA, powering robots at 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi and Techman — with a long tail of hardware integrators like Aetina, Advantech and Seeed building Thor-based systems around them.
Not cloud demos. Real robots, real control loops, on-device. That's the pitch, and it lands.
Now trace the single thread.
Cosmos 3, per NVIDIA's Cosmos Lab, uses "one backbone to solve all the different tasks" — world understanding, prediction, simulation and action, across humanoids, grippers and self-driving cars. Ming-Yu Liu's framing: "Every embodiment speaks a different language. Our solution is to build a common vocabulary."
That common vocabulary is the point. It is the same backbone whether the model is training in a data center or inferring on a 70-watt board.
Cosmos 3 Edge is explicitly "its newest addition" — the same family, shrunk. Not a separate edge product line, but a compression of the cloud model onto the smallest Jetson.
The training end of that thread is Vera Rubin. Cosmos-Dreams, the closed-loop simulator collection announced alongside, reportedly lets developers "verify the accuracy of their models before deploying them in a real fleet" — synthetic scenarios generated on NVIDIA GPUs, validating policies destined for NVIDIA edge silicon.
And the policies themselves are GR00T, post-trained with NVIDIA TAO — "Post-Train NVIDIA Cosmos 3 In a Day with NVIDIA TAO Agent Skills," as the release puts it.
So follow the developer's actual workflow. You simulate in Cosmos-Dreams. You train GR00T policies on Vera Rubin. You post-train with TAO. You compress to Cosmos 3 Edge. You deploy on Jetson Thor.
Every layer is NVIDIA. Every layer is codesigned to interoperate only with the layers above and below it, all riding CUDA.
Ask the question that last week's coverage left open: can a competing simulator or accelerator let a developer revalidate a GR00T-trained policy without starting from scratch?
For the edge tier, the answer is now visible in NVIDIA's own materials. There is no revalidation path off the stack, because the stack was never assembled from parts — it was engineered as one system.
That phrasing is not mine. It's NVIDIA's, describing Spectrum-6 as "inherent to NVIDIA's vertically integrated, horizontally open approach: codesigning silicon, systems and software." Seven chips and five rack trays in Vera Rubin, "engineered as a single system rather than assembled from separate off-the-shelf products."
Read that as an engineering boast and it's a triumph. Read it as a dependency architecture and it's a warning.
Because "engineered as a single system" means the interfaces between layers are not stable, documented, swappable contracts. They are internal seams. When your simulator, your training cluster, your model and your edge silicon are codesigned together, pulling out one layer doesn't cost you that layer — it costs you the validation of the whole chain.
A robot policy is only as trustworthy as its last validation run. Swap the accelerator, and every guarantee you established in simulation and training is now unverified against new hardware behavior. Swap the world model, and your post-training assumptions no longer hold.
The vertical integration is not a side effect of good engineering. It is the product. The lock-in is what you're buying, and it isn't priced into the humanoid-robotics hype.
Two exposures fall out of NVIDIA's own numbers.
The first is power economics. CoreWeave's first Vera Rubin benchmark on DeepSeek-R1: 10x more throughput per megawatt than Grace Blackwell NVL72. Against GB200 NVL72, Vera Rubin NVL72 is pitched at up to 10x more tokens per megawatt and one-tenth the cost per million tokens.
Throughput-per-megawatt is a real constraint — agentic systems, NVIDIA notes, can consume up to 15x more tokens than traditional applications, and AI factories are power-bound. So the efficiency gain is genuine.
But watch what it does. When the binding constraint is watts, the vendor with the best tokens-per-watt doesn't just win the deal — they set the terms of every deal. Efficiency becomes leverage. The better NVIDIA's performance-per-watt, the more expensive it becomes, in raw compute, for anyone to run an alternative stack at competitive cost.
The throughput-per-megawatt pitch is a dependency multiplier dressed as a sustainability win.
The second exposure is geopolitical, and it sits in Fort Worth.
Wistron's D1 plant — 324,000 square feet, a $700 million commitment — is building the Vera Rubin Superchip on U.S. soil, scaling to tens of thousands of boards per month. Part of NVIDIA's pledge to manufacture up to $500 billion in AI platforms in America.
Huang's frame: "in the future these AI factories will be a fundamental infrastructure of society… Electricity comes in … and the tokens get generated in the factories."
If tokens are infrastructure like roads and electricity, then the question of who owns the toll booth is not academic. The re-shoring is urgent precisely because the supply chain is singular — 350+ factory sites in 30 countries, but all feeding one architecture.
Here is the tension NVIDIA's materials never resolve.
The T3000, per NVIDIA, isn't expected to ship until Q1 2027. The full edge tier — Nemotron, Cosmos 3, GR00T, Jetson agent skills on the new modules — is a 2027 promise, even as AGX Thor reportedly ships today running GR00T at, per Aetina, up to 2,070 FP4 TFLOPS in a 130-watt envelope.
So the lock-in is being committed to before the hardware that completes it exists. Developers building on Thor today are betting on a roadmap, wiring their robots to a stack whose edge silicon is more than a year out.
That's the shape of the thing. A physical-AI industry consolidating onto one simulator, one policy framework, one software layer and one silicon family — cloud to edge, training rack to control loop — and doing it at the exact moment it scales.
The word for a component that everything depends on and nothing can replace is not "platform."
It's single point of failure.
The demos are real. The efficiency is real. The robots move. And every one of them, from the Vera Rubin rack that trained the policy to the 70-watt board running it, speaks only one vocabulary now — and there is no dictionary out.