> At a Glance
> – Nvidia CEO Jensen Huang says the next-gen Vera Rubin AI platform is now in full production
> – Rubin slashes AI running costs to one-tenth of today’s Blackwell system
> – Microsoft and CoreWeave will be first cloud partners to deploy Rubin later this year
> – Why it matters: Cheaper, faster AI could lock customers tighter into Nvidia’s ecosystem
Nvidia used its CES stage to tell investors and customers that Vera Rubin-its upcoming six-chip AI super-platform-remains on track for release in the second half of 2026.
Inside the Rubin Platform
The Rubin family pairs a Rubin GPU with a Vera CPU, both etched on TSMC’s 3-nanometer node and linked by Nvidia’s sixth-generation interconnect. Huang called every element “completely revolutionary and the best of its kind.”
Sunday’s analyst briefing added two headline figures:
- Operating cost per model drops to ~10% of Blackwell levels
- Training can use 75% fewer chips for certain large models
Early Adopters and Supply Chain
Nvidia confirmed that Microsoft and CoreWeave will offer Rubin-powered cloud instances as soon as volume ships. Microsoft’s new Georgia and Wisconsin AI data centers are being designed to house thousands of the chips.
Other partners already sampling early silicon include:
- Red Hat, targeting open-source enterprise stacks for banks, automakers, airlines, and governments
- Select AI labs running unreleased next-gen models
| Metric | Blackwell | Rubin |
|---|---|---|
| Relative running cost | 100% | ~10% |
| Typical training chips | 100 | ~25 |
| Process node | 4N | 3 nm |
What “Full Production” Means
Huang’s phrase “full production” is deliberately vague. For a chip this complex, the phrase usually signals that initial low-volume wafers have passed validation and are now building through TSMC’s fabs, with volume ramp still slated for H2 2026. Analyst Austin Lyons told News Of Fort Worth the announcement counters Wall-Street chatter that Rubin might slip.
Competitive Stakes
The AI boom has cloud providers racing for every new GPU generation. While some-like OpenAI with Broadcom-are designing custom silicon, Lyons argues Nvidia’s tighter integration across compute, networking, memory, and software “is getting harder to displace.”
Key Takeaways
- Vera Rubin silicon is now taping out through TSMC’s 3 nm lines
- Customers can expect 90% lower operating costs and 4× faster training on large models
- Microsoft and CoreWeave first in line for cloud deployments later this year
- Full volume shipments remain scheduled for the second half of 2026
If Rubin hits its performance and price claims, Nvidia’s grip on the AI stack could tighten even as the industry explores home-grown alternatives.
