NVIDIA broadens its CES 2026 agenda beyond GPUs
At CES 2026 in Las Vegas, NVIDIA presented itself less as a graphics-card maker and more as an end-to-end infrastructure company for artificial intelligence, high-performance computing, autonomous vehicles, and robotics. The announcements, delivered by founder and CEO Jensen Huang at the Fontainebleau Las Vegas, spanned data center platforms, networking and storage, “open” model families, and production-ready automotive software.
Rather than leaning on a single consumer product cycle, the company emphasized building blocks aimed at developers, enterprises, and ecosystem partners—components that underpin training, inference, simulation, and deployment of large-scale AI systems.
Rubin platform: “extreme codesign” for next-gen AI workloads
The centerpiece was the Rubin platform, which NVIDIA described as being built with “extreme codesign”—designing major compute and networking components as a unified system. Rubin combines the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet Switch.
The company positioned Rubin for demanding AI workloads such as mixture-of-experts models, agentic systems, and long-context reasoning. NVIDIA said Rubin can reduce per-token inference cost by up to 10x and, for some training workloads, require fewer GPUs than its Blackwell generation—claims that, if realized in customer deployments, could materially shift total cost of ownership for frontier-scale AI.
DGX SuperPOD and deskside systems target both clusters and local AI
To operationalize Rubin at scale, NVIDIA introduced the DGX SuperPOD as a reference design for Rubin-era deployments across enterprise and research environments. The company framed it as a “default” architecture integrating compute, networking, and software for training, inference, and long-context tasks, again citing up to 10x lower token costs compared with Blackwell-based configurations.
At the other end of the spectrum, NVIDIA also highlighted smaller form-factor systems aimed at running large models locally: DGX Spark and DGX Station. Marketed as “deskside AI supercomputers,” the systems are designed to run open-source and frontier models on-premises and then burst to the cloud when needed. NVIDIA said DGX Spark can handle 100B-parameter models, while DGX Station targets up to 1T-parameter models.
BlueField-4 and “Inference Context Memory Storage” for long-context AI
A recurring theme across the keynote was the growing strain that long-context and multi-agent AI places on memory and data movement. NVIDIA introduced Inference Context Memory Storage, a platform powered by BlueField-4, designed to manage context data that does not fit efficiently in traditional GPU memory.
The company said the approach increases effective memory capacity and enables context sharing across large clusters, with claims of up to 5x more tokens per second and 5x better power efficiency compared with traditional storage methods. BlueField-4 is expected in the second half of 2026, with partners including Dell, HPE, IBM, Nutanix, Pure Storage, Supermicro, VAST Data, and WEKA.
Enterprise AI Factory adds security and Kubernetes isolation
NVIDIA also updated its Enterprise AI Factory validated design, adding integrations aimed at improving security posture and accelerating infrastructure deployment. New and expanded integrations include Armis, Check Point, F5, Fortinet, Palo Alto Networks, Rafay, Red Hat OpenShift, Spectro Cloud (PaletteAI), and Trend Micro.
The company said the additions bring features such as telemetry through NVIDIA DOCA Argus and stronger workload isolation in Kubernetes environments—capabilities increasingly demanded by regulated industries adopting generative AI at scale.
Open resources: Nemotron, Cosmos, Alpamayo, GR00T, Clara
In a move to expand developer adoption, NVIDIA announced a slate of open resources—models, datasets, and training tools—organized into technology families: Nemotron (AI agents), Cosmos (physical AI), Alpamayo (autonomous driving), Isaac GR00T (robotics), and Clara (biomedical).
The company also cited the scale of its contributions, including 10 trillion language tokens, 500,000 robotics paths, 455,000 protein structures, and 100TB of vehicle sensor data—figures intended to underscore the breadth of its ecosystem push beyond hardware.
Autonomous driving: Alpamayo, DRIVE Hyperion, and Mercedes production plans
For automotive, NVIDIA introduced the Alpamayo family, combining AI models, simulation tools, and datasets focused on rare and complex scenarios. Components include Alpamayo 1, AlpaSim, and Physical AI Open Datasets. The company described Alpamayo as a reasoning model intended to advance development toward level 4 autonomy, with early collaborators including JLR, Lucid, Uber, and Berkeley DeepDrive.
NVIDIA also expanded its DRIVE Hyperion ecosystem, citing partnerships with suppliers and sensor companies including Aeva, Bosch, and Sony. The architecture pairs compute and sensors and, according to the company, can be configured with two DRIVE AGX Thor chips delivering more than 2,000 TFLOPS for sensor processing and real-time workloads.
In one of the most concrete production updates, NVIDIA said its DRIVE AV software will debut in the all-new Mercedes-Benz CLA, starting with an “enhanced level 2” driver-assistance system expected on U.S. roads by the end of the year. The company described a “dual-stack” approach: an end-to-end AI driving stack paired with a classical safety stack built on NVIDIA Halos for redundancy.
Industrial and creator updates: Siemens expansion and RTX AI tools
Beyond vehicles, Siemens and NVIDIA announced an expanded partnership to deepen AI in industrial workflows, including digital twins and physical AI. NVIDIA will provide AI infrastructure, simulation libraries, and models, while Siemens said it will commit “hundreds” of industrial AI experts alongside its hardware and software capabilities.
On the PC side, NVIDIA showcased local AI video generation tools using LTX-2, claiming output of up to 20 seconds of 4K video with built-in audio and multi-keyframe support. The company also highlighted performance work with ComfyUI, plus support for NVFP4 and NVFP8 formats on RTX 50 Series GPUs to reduce VRAM use and improve speed.
For gaming, NVIDIA introduced DLSS 4.5, including a “6X” mode that can generate additional frames per rendered frame, and said G-SYNC Pulsar monitors are now available with over 1,000Hz effective motion clarity and G-SYNC Ambient Adaptive Technology.
Related
