NVIDIA Unveils Rubin at CES 2026: Cheaper Chips and Cars That Reason

Jensen Huang once again did what he does best: turn an electronics trade show into a showcase for the direction he wants to steer the entire industry. At the opening of CES 2026, held at the Fontainebleau in Las Vegas, NVIDIA's CEO unveiled Rubin, the successor to the Blackwell architecture, and Alpamayo, a family of open models for autonomous driving. The thread connecting both announcements is a thesis Huang has repeated for years: traditional computing is being replaced by accelerated computing and artificial intelligence.

"About $10 trillion of the last decade of computing is now being modernized toward this new way of computing," Huang said. The figure is the company's own estimate of the value of the world's IT infrastructure that, according to his narrative, will need to be rebuilt around GPUs. It's worth reading it for what it is: the size of the market NVIDIA aims to capture, not a neutral forecast.

Rubin: six chips designed as a single piece

The announcement with the most technical weight is Rubin, named after American astronomer Vera Rubin and already, according to the company, in full-scale production. NVIDIA describes it as its first "extreme-codesigned" AI platform, a term of its own coinage meaning that its six components were designed together to function as a single system, rather than assembling parts conceived separately.

Those six components are:

• Rubin GPU, delivering 50 petaflops of inference in NVFP4 format (a low-precision numerical representation that allows models to run faster while using less memory).
• Vera CPU, geared toward data movement and agent processing.
• NVLink 6 for internal server interconnection.
• Spectrum-X Ethernet Photonics for the network linking racks together.
• ConnectX-9 SuperNICs.
• BlueField-4 DPU, processors dedicated to offloading networking and storage tasks.

The logic behind designing everything together is to avoid bottlenecks. When a model is trained or run across thousands of chips at once, performance isn't set by the fastest GPU but by the weakest link in the chain: the network, storage, memory. Huang argues that integrating every layer drastically cuts the cost of training and inference.

The figure he wants the industry to remember is this: Rubin would generate tokens at a tenth of the cost of the previous platform. A token is the smallest unit a model processes (roughly a syllable or word fragment), and its cost determines how expensive it is to run an AI service at scale. If the promise holds up in production, it would significantly cut the cost of deploying large models, currently one of the main economic barriers for companies.

As a complement, NVIDIA introduced an "AI-native" storage system designed to speed up inference with long contexts —cases where the model must recall lengthy conversations or documents— with, according to its own measurements, 5x multipliers in tokens per second, performance per dollar, and energy efficiency.

A strategic pivot: NVIDIA as a model builder

The most interesting move from a business standpoint isn't in the hardware, but in Huang's insistence on presenting the company as a "frontier AI builder." NVIDIA has spent years making money selling the picks and shovels of the gold rush; now it wants to mine too.

The company laid out six families of open models, trained on its own supercomputers and released publicly: Clara for medicine, Earth-2 for climate science, Nemotron for reasoning and multimodality, Cosmos for robotics and simulation, GR00T for embodied intelligence in robots, and Alpamayo for autonomous driving.

"We build them completely open so that every company, every industry, and every country can be part of this AI revolution," Huang said. The strategy has an obvious commercial reading: the more open models circulate optimized for NVIDIA hardware, the harder it becomes to leave its ecosystem. Open source, here, is also a retention tool.

Cars that reason and explain their decisions

The announcement with the most tangible implications for the general public is Alpamayo, an open portfolio of models, simulation blueprints, and data for developing Level 4 autonomous driving (vehicles capable of driving themselves in defined environments without human intervention).

Its centerpiece is Alpamayo R1, which NVIDIA describes as the first open "vision-language-action" (VLA) model for driving. The difference from conventional systems is conceptual: beyond translating what the cameras see into actions on the steering wheel, brake, and accelerator, the model reasons about the action it's about to take. It's the promise of explainable autonomous driving, able to justify why it brakes or changes lanes—relevant both for user trust and for future regulatory audits.

The leap from lab to road already has a date and a license plate: the first passenger car with Alpamayo, built on NVIDIA's DRIVE platform, will be the new Mercedes-Benz CLA, with "AI-defined" driving arriving in the United States this year. The model recently earned five stars in EuroNCAP safety tests.

Rounding out the lineup is AlpaSim, an open simulation environment for testing autonomous vehicles in high-fidelity scenarios before taking them onto the street. The idea of training in synthetic virtual worlds —recreating extreme cases that would be dangerous or impossible to trigger in real life— is today one of the pillars of the industry.

From supercomputer to desktop

Huang spent part of the presentation bringing the message down to a personal level. He showed an AI agent running locally on the DGX Spark, an AI-focused desktop computer, embodied in a small Reachy Mini robot using Hugging Face models. The point was to illustrate that running agents locally, without relying on the cloud, is now trivial.

Among the companies integrating its technology, NVIDIA cited Palantir, ServiceNow, Snowflake, CrowdStrike, and NetApp. "The agentic system is the interface," Huang summed up, pointing to a future where people interact with software through agents rather than menus and buttons.

There was also room for the business that gave rise to the company in the first place: video games. NVIDIA announced DLSS 4.5, featuring a new frame-generation mode and more than 250 games compatible with its DLSS 4 technology, along with G-SYNC Pulsar monitors and the arrival of GeForce NOW on Linux and Amazon Fire TV.

What it all means

The underlying message of CES 2026 is one of continuity and ambition. NVIDIA no longer just sells chips; it sells full platforms —from the data center to the robot and the car— and wants a presence at every layer of the AI value chain, including building the models themselves.

The performance and cost figures come from the company itself and will need to be confirmed by real-world deployments and independent benchmarks; that's the usual caution warranted with any manufacturer's announcement. But two bets deserve close attention for their practical consequences. The first is cheaper tokens: if Rubin delivers, the economics of running AI services change for many mid-sized companies. The second is explainable autonomous driving: the Mercedes CLA with Alpamayo will be a real-world test case, on real roads with regulators watching, to see whether a car that "reasons" about its decisions offers anything beyond marketing.