Image courtesy of Nvidia
Nvidia $NVDA is in Las Vegas sounding a little bit like an infrastructure ministry: Here are the chips, here are the racks, here’s the networking, here’s the software — and by the way, those robots and cars you keep hearing about are supposed to run on all of it.
The through-line in this year’s Consumer Electronics Show (CES) batch is control of the full stack, with a particular fixation on storage and what Nvidia keeps framing as the next bottleneck: agentic AI that needs more context, more memory, more networking, and fewer excuses for why it can’t run in the real world. The pitch is that “AI factories” are now a product category, and Nvidia intends to sell the blueprints, the machines, the operating system, and everything else.
A lot of what Nvidia and CEO Jensen Huang announced Monday afternoon has been floating around for months — Rubin as the post-Blackwell architecture, BlueField-4 as the DPU jump, Nemotron as Nvidia’s “open” model family, Halos as the safety umbrella. What's new is the bundling. Nvidia is turning that roadmap into a single argument: six chips, one platform, plus the networking and “context memory” plumbing to keep long-horizon agents from stalling out.
The headline hardware story is that Rubin is getting carved into shippable building blocks. Nvidia laid out a Rubin “platform” made up of six components — GPUs and CPUs in Rubin and Rubin Ultra flavors, plus NVLink 6 switches and a ConnectX-9 SuperNIC — with performance and cost-per-token claims all built around that full-system co-design and all designed to drive down the cost of intelligence.
On the system side, Nvidia is positioning Vera Rubin NVL72 as the rack-scale workhorse (72 GPUs and 36 CPUs, with exaflops-class FP4 claims), and Rubin Ultra NVL288 as the bigger follow-on (288 GPUs and 144 CPUs). The company is also plugging Rubin into DGX-branded “AI factories,” pairing DGX Rubin NVL72 for training with DGX Rubin NVL8 for inference as a more turnkey, standardized unit of capacity. Nvidia says Rubin-based products will be available from partners in the second half of 2026.
Two infrastructure add-ons are doing a lot of quiet work here. First, Nvidia is leaning hard into networking as a first-class performance feature, touting Spectrum-X $TWTR Ethernet photonics switch systems and attaching “five times” claims around inference performance and power efficiency. Second, the company is trying to make “long context” feel like an infrastructure purchasing decision, unveiling an “inference context memory” storage platform to extend agentic AI context windows. If the subtext of Rubin is “the roadmap is real,” the subtext of the surrounding plumbing is “the next margin pool is everything around the GPU.”
Nvidia keeps describing physical AI as the moment when “agents” stop being chatty and start being competent — meaning they need perception, reasoning, and action in the same loop. That means three building blocks: better simulation, better robot models, and better “mobility” models.
“The ChatGPT moment for robotics is here,” Huang said in a press release, arguing that “models that understand the real world, reason, and plan actions” are opening “entirely new applications.” Automotive is, then, perhaps where Nvidia’s “full stack” argument turns into a credibility test, because it’s the one category where “demo” and “deployment” are separated by regulation, liability, and a decade of bruised optimism.
Nvidia says its Drive AV platform for assisted driving tech is “in production” for the 2026 Mercedes-Benz CLA, which received the highest Euro NCAP safety score in all of 2025. The company says the car has “advanced Level 2 automated driving capabilities” with “point-to-point urban navigation,” including “address-to-address” trips — and frames Hyperion as the compute-and-sensor architecture that adds redundancy for safety. Nvidia says the car will be capable of hands-free driving on U.S. roads by the end of the year.
Then, there’s the broader bet. “We believe physical AI and robotics will eventually be the largest consumer electronics segment in the world,” said Ali Kani, Nvidia’s automotive VP. “Everything that moves will ultimately be fully autonomous, powered by physical AI.” Kani said that Alpamayo, Nvidia’s “family of open source AI models, simulation tools, physical AI datasets” for autonomous driving, is built to accelerate “safe, reasoning-based physical AI development.” The company released 1,700 hours of driving data alongside an open-source simulation framework — and positioned the tools as the starter kit for Level 4 autonomy.
Nvidia is pointing to Isaac GR00T N1.6 as an open reasoning vision-language-action model for robot skills, along with Isaac Lab Arena as an evaluation framework for testing policies at scale. The company also calls out Cosmos Reason 2 as a model aimed at improving physical reasoning, and the broader Cosmos lineup as a way to generate synthetic data for training physical AI. Nvidia is also positioning Jetson T4000 as the edge compute target for robots, paired with the same training-to-deployment pipeline that feeds back into DGX-class infrastructure. Nvidia is trying to make the robot stack feel like the software stack: train in a world it can generate, test in a world it can vary, deploy on hardware it can sell.
Nvidia’s “open models” story is less about joining the open-source movement out of ideological awakening and more about pulling developers into its ecosystem with free samples — packaged to run best on Nvidia infrastructure. “An expansion,” as Nvidia VP of generative AI and software Kari Briski said. The company frames this as a bundle: new Nemotron-3 models (including Llama Nemotron-3 variants in 70B, 34B, and 8B sizes), plus new datasets and tools meant to help teams build domain agents and deploy them through Nvidia’s stack.
“In 2025, Nvidia was the top contributor ... on Hugging Face with 650 open models and 250 open datasets,” Briski said. Essentially: Nvidia wants to be the place you start — even if you don’t stay “open” for long.
The company says it’s releasing Nemotron-CC, a multilingual pretraining corpus of 1.4 trillion tokens across more than 140 languages, positioned as an “open” foundation layer for building and adapting models. It also highlights a “Granary” instruction dataset meant to make models more useful out of the box for enterprise-style tasks. Nvidia is framing Nemotron as a toolkit for the agentic era: models and datasets for safety, RAG, speech, and reasoning.
Zoomed out, Nvidia’s CES message is consistent across all three buckets. The future is the pipeline, and Nvidia wants every bit of it — compute, networking, storage, safety, simulation — to run on something it already sells.
