At CES 2026, NVIDIA unveils Rubin AI platform, open models and AV push
Jensen Huang outlined a full‑stack roadmap spanning data centre platforms, open model families, desktop AI and gaming updates.
NVIDIA opened CES 2026 in Las Vegas with a broad slate of AI announcements led by Rubin, a new six‑chip platform that the company says is now in full production.
Founder and CEO Jensen Huang framed the reveal as part of a shift to accelerated computing across industries, alongside new open model families, tools for robotics and autonomous driving, and gaming updates.
Rubin aims to cut AI costs and integrate the stack
According to the company, Rubin is an “extreme codesigned” platform built from the data centre outward. It combines Rubin GPUs rated at 50 petaflops of NVFP4 inference, Vera CPUs for data movement and agentic workloads, NVLink 6 for scale‑up, Spectrum‑X Ethernet Photonics for scale‑out, ConnectX‑9 SuperNICs and BlueField‑4 DPUs. NVIDIA said the architecture is intended to remove system bottlenecks and deliver AI tokens at roughly one‑tenth the cost of its previous platform.
Huang also highlighted an AI‑native storage tier, the NVIDIA Inference Context Memory Storage Platform, which the company claims can lift long‑context inference throughput and efficiency by up to five times on key metrics, pointing to the growing need to serve larger context windows reliably.
Open models and the push into physical AI
NVIDIA said its open models, trained on its own supercomputers, now span six domains, including Clara for healthcare, Earth‑2 for climate science, Nemotron for reasoning and multimodal AI, Cosmos for robotics and simulation, GR00T for embodied intelligence and Alpamayo for autonomous driving. The company positioned these as building blocks that enterprises can evaluate, guardrail and deploy.
On stage, Huang showcased Cosmos for generating realistic video, synthesising multi‑camera driving scenarios and modelling edge cases, and introduced Alpamayo as an open portfolio of reasoning vision‑language‑action models and simulation blueprints intended to enable level‑4 capable autonomy. He said the first passenger car featuring Alpamayo, built on the NVIDIA Drive stack, will arrive in the new Mercedes‑Benz CLA in the United States.
Personal AI and desktop supercomputing
Beyond data centres, the company demonstrated a personalised AI agent running locally on the DGX Spark desktop system and embodied through a Reachy Mini robot, underscoring its pitch that agentic AI will become a hands‑on tool for developers and teams. NVIDIA said DGX Spark delivers up to 2.6 times performance for large models, with support for new image models and planned NVIDIA AI Enterprise availability.
Gaming updates: DLSS 4.5 and displays
For gamers, NVIDIA announced DLSS 4.5, which adds Dynamic Multi Frame Generation, a new 6X mode and an updated transformer for DLSS Super Resolution. The company also spotlighted G‑SYNC Pulsar monitors, positioned as delivering tear‑free visuals with very high perceived motion clarity, while partners rolled out fresh GeForce RTX laptops, desktops and monitors at CES.
How will Rubin change AI infrastructure costs?
The company’s argument is that codesign across compute, networking, storage and software improves utilisation and throughput, so fewer components are needed to deliver a given number of tokens, cutting cost per token.
Rubin’s confidential‑by‑design approach and an AI‑native context memory tier, together with next‑generation interconnects, are meant to keep large‑context inference fast and predictable for developers shipping production workloads.


