Nvidia’s CES 2026 message: the data centre is the product

On Monday, 5 January, Jensen Huang, Nvidia’s CEO opened the show with a message that wasn’t really about a faster GPU. It was about a new unit of competition: the rack-scale “AI factory”.

Named Vera Rubin, the chip making Giant calls it: “six new chips, one AI supercomputer.” Nvidia says Rubin can deliver major reductions in the cost of running AI, not just improvements in raw performance.

The company is claiming up to 10× lower inference token cost versus Blackwell, and says it can train mixture-of-experts models with 4× fewer GPUs. The company also claims that the Rubin GPU itself delivers around 5× the AI training performance of Blackwell.

Nvidia’s own materials describe Rubin as a six-part stack: Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-6 Ethernet switch.

If the “AI factory” is what customers buy, then Nvidia gets paid not only for GPUs, but for the network fabric, the DPUs, and the software that makes the whole thing behave like one machine.

In other words: Nvidia is pushing the industry toward a future where “switching away” isn’t just swapping accelerators. It’s rewiring the data centre.

Nvidia is trying to stop being benchmarked as a component supplier, and start being bought as infrastructure.

Nvidia is heavily promoting the Vera Rubin NVL72 rack design — the kind of configuration cloud providers can scale as a repeatable building block. News reports describe it as 72 GPUs and 36 CPUs in a rack-scale system, designed to be assembled into larger “pods” that can extend past 1,000 Rubin chips.

Nvidia says Rubin-based products will start rolling out via partners in the second half of 2026. Huang also stated Rubin is in “full production.” Wired notes the phrase is not tightly defined and reads partly as a signal: reassure the market after prior generation issues and emphasise that Nvidia believes Rubin is on schedule.

That matters because the AI infrastructure boom is now being financed like industrial expansion. People are building power, cooling, and capacity years ahead. Confidence signals move budgets.

One of the more important, less headline-friendly details: Nvidia is using Rubin to push deeper into regulated enterprise buyers

News reports highlights third-generation confidential computing and Nvidia positioning Rubin as a “rack-scale trusted computing platform.”

This is Nvidia’s bet that the next phase of AI adoption won’t only be frontier labs. It will also be banks, insurers, governments, and hospitals — buyers who care about auditability, isolation, and multi-tenant security at scale.

Rubin also comes with messaging around networking upgrades and new infrastructure layers that sit underneath inference.

The Register’s technical reporting focuses on the rack design and the system-level features, including how Nvidia is extending confidential computing beyond the traditional CPU boundary.

If you zoom out, this is Nvidia trying to defend against a real threat: unbundling. If rivals can split the stack — someone else provides networking, someone else provides inference accelerators, someone else provides orchestration — Nvidia’s pricing power weakens. Rubin is designed to keep the bundle intact.

“Physical AI” is the second act — and it’s meant to widen demand

Huang didn’t only talk about data centres. He also pushed a narrative Nvidia wants to own: “physical AI” — models that can perceive and act in the real world.

Axios reports Nvidia framed this as a “ChatGPT moment for physical AI,” with robotaxis and autonomy positioned as early adopters, and Nvidia talking about testing a robotaxi service by 2027.

Nvidia’s own CES blog post says it announced open models and tools for robotics and autonomous driving alongside Rubin.

There’s a business logic here. If robotics and autonomy become credible markets, they become a new source of compute demand — and Nvidia wants that demand to be trained, simulated, and deployed on Nvidia-first tooling.

The Groq deal explains the subtext: inference is where the war is moving

Two weeks before CES, Groq announced a non-exclusive inference technology licensing agreement with Nvidia, and said Groq’s founder Jonathan Ross and president Sunny Madra would join Nvidia to help scale the licensed technology.

CES 2026 makes one thing clear: AI is moving from a breakthrough story to an operating model story.

Nvidia’s Rubin launch is a bid to keep control as customers become more sophisticated and as competitors try to create escape hatches. It’s also a bet that the constraint won’t just be chips. It will be power, deployment speed, security, and total cost of ownership — the unglamorous parts of building an AI economy.

And that’s why Nvidia walked into CES and talked like a utility company.