Facing its most serious competition in years, Nvidia used CES 2026 to unveil the Vera Rubin platform — a six-chip system it frames as a complete AI supercomputer — alongside a wave of new open models for reasoning, robotics and autonomous driving.
From GPU vendor to full-stack AI supplier
At the CES show in Las Vegas, Nvidia introduced the Rubin platform, the successor to its widely used Blackwell architecture. The central message, echoed by analysts, was that competitive advantage in AI is no longer just about the GPU but about an integrated system spanning compute, networking and software.
The Rubin platform brings together six chips: the Vera CPU, the Rubin GPU, the NVLink 6 switch, the ConnectX-9 SuperNIC, the BlueField-4 DPU and the Spectrum-6 Ethernet switch. Using Nvidia’s NVLink interconnect and transformer-acceleration technologies, it is designed to scale agentic AI, advanced reasoning and mixture-of-experts models beyond what Blackwell offered. Nvidia positioned the platform not only for its own use but as something third parties can build full-stack products around.
A push into open models
Alongside the hardware, Nvidia expanded its open-model line-up. The Nemotron 3 family, aimed at building and deploying multi-agent systems, includes speech models for real-time, low-latency recognition in captioning and voice applications, plus new embedding and reranking vision-language models for retrieval-augmented generation (RAG), along with datasets, training resources and blueprints. Nvidia also detailed its Cosmos world foundation models — used to generate synthetic training data and support physical-AI applications such as humanoid robots — and Alpamayo, a family of open vision-language-action models targeted at autonomous vehicles. The broader idea of models that build an internal representation of the physical world is explored in this overview of world models.
What analysts made of it
Observers saw a deliberate strategy of specialisation. Mark Becque, an analyst at Omdia, noted that releasing narrowly specialised open models is an unusual approach, but one that can help customers adopt them faster. Bradley Shimin of Futurum tied it to a 2026 trend toward faster deployment of task-specific rather than general models, pointing to Nvidia’s focus on domains such as health care, autonomous vehicles and specific enterprise use cases — an effort, in his framing, to be not just the leading frontier-model builder but the leading builder of applied intelligence. The specialisation contrasts Nvidia with rivals including AMD, Intel and Qualcomm, which analysts said are moving in a similar direction but have not yet matched Nvidia’s breadth.
Limitations and what to watch
The announcements are a statement of direction as much as a finished offering. Even with open weights and recipes, enterprise adoption of open models remains limited, and analysts noted that companies still lean more on proprietary models. There is also a strategic tension: the more Nvidia integrates chips, networking, models and infrastructure, the harder it becomes for enterprises to avoid deepening their reliance on a single vendor. Performance and availability claims come largely from Nvidia’s own launch materials and will need independent benchmarking, and product timelines at events like CES can shift. Buyers should weigh lock-in, real-world support for open models, and verified performance before committing.
