Jensen Huang Says Nvidia’s New Vera Rubin Chips Are in ‘Full Production’

Summary

Nvidia CEO Jensen Huang announced at CES that the company’s next-generation AI platform, Vera Rubin, is “in full production” and is expected to begin reaching customers later this year. Nvidia claims Rubin will cut the cost of running AI models to roughly one-tenth of its current Blackwell systems and can train some large models using about one-quarter the number of chips Blackwell requires. Early partners include Microsoft and CoreWeave, with Microsoft data centres in Georgia and Wisconsin slated to deploy thousands of Rubin chips. The Rubin platform is a tightly integrated system of six chips — including a Rubin GPU and a Vera CPU — built on TSMC’s 3nm process and using advanced bandwidth memory and sixth-generation interconnect and switching tech.

The company is also working with Red Hat to ensure enterprise software runs on Rubin. Analysts note that “full production” may mean Nvidia cleared key validation phases and is signalling progress to investors, while broader customer demand and the usual ramping steps still lie ahead. The announcement underscores Nvidia’s shift from GPU vendor to integrated AI system architect, even as hyperscalers pursue custom silicon to hedge their exposure.

Key Points

Jensen Huang said Vera Rubin is “in full production” at CES and will start arriving to customers later this year.
Nvidia claims Rubin will reduce model run costs to about one-tenth of Blackwell and train certain models with ~25% of the chips Blackwell needs.
Rubin is a multi-chip platform (six chips) with a Rubin GPU and Vera CPU, built on TSMC 3nm and using advanced bandwidth memory and new interconnects.
Early partners include Microsoft and CoreWeave; Microsoft data centres in Georgia and Wisconsin will deploy thousands of Rubin chips.
Nvidia is collaborating with Red Hat to broaden enterprise software support for Rubin-based systems.
“Full production” likely signals cleared validation milestones, but chip production typically ramps from low volume to full scale over time.
Past issues (Blackwell overheating when racked) remind the market production can be complex; Nvidia aims to reassure investors and customers about Rubin’s timeline.
The launch strengthens Nvidia’s integrated-platform position, though hyperscalers (eg OpenAI with Broadcom) continue investing in custom silicon as a hedge.

Context and Relevance

This matters because Rubin’s cost and efficiency claims, if realised, could make advanced AI far cheaper to operate — lowering barriers for large-scale AI deployments and reinforcing Nvidia’s dominance across compute, memory and networking for AI. For cloud providers, AI startups, and enterprises running generative models, Rubin could reshape procurement decisions and competitive dynamics in data centres and chip design. The story sits at the intersection of hardware innovation, supply-chain realities (TSMC 3nm), and the strategic battle between integrated platforms and bespoke in-house silicon.

Author style

Punchy. This article is significant — it isn’t just another product update: it’s Nvidia trying to lock in customers with a cheaper, more integrated AI stack. If Rubin delivers, it’s a big deal for AI economics and who controls the foundation of modern AI services.

Why should I read this?

Short version — read it if you care who’ll be powering the next wave of AI. Rubin could massively cut running and training costs, tilt data-centre procurement back toward Nvidia, and make it harder for rivals or in-house chips to compete. It’s the sort of industry move that changes cost maths and strategy, so worth a quick skim if you follow AI, cloud or semiconductors.