Table of Contents
Nvidia Vera Rubin is technically a big improvement — but potential customers are still working on their own hardware
In sum – what we know:
- Official launch: NVIDIA unveiled the Vera Rubin AI platform at CES 2026, named after the astronomer known for discovering dark matter.
- Performance leap: The system promises 5x the computing power of Blackwell and features HBM4 memory bandwith of 22 TB/s.
- Availability: Chips are currently in production on TSMC’s 3nm process, with delivery to major customers set for the second half of 2026.
Nvidia CEO Jensen Huang took advantage of CES 2026 to finally unveil the next-generation of Nvidia AI accelerators, in the form of Vera Rubin. The platform arrives as AI models continue their relentless expansion.
Vera Rubin may feel like an iterative upgrade given its naming scheme — the previous generation was called Rubin. But it’s actually a pretty substantial leap forward. Nvidia is abandoning soldered-down memory in favor of a modular architecture. It’s a direct response to the flexibility and upgrade constraints that plagued earlier generations.
Tech specs
Under the hood, Vera Rubin brings together six tightly coordinated chips, all manufactured on TSMC’s 3nm node. The Vera CPU packs 88 custom Olympus cores — tuned specifically for AI factory environments. The Rubin GPU handles the heavy compute lifting, pairing HBM4 memory with NVIDIA’s latest Transformer Engine.
The interconnect system is perhaps even more ambitious. It’s built on NVLink 6, which pushes GPU-to-GPU bandwidth to 3.6 TB/s, while ConnectX-9 handles high-throughput, low-latency networking. Rounding things out is the BlueField-4 DPU for data processing tasks.
The numbers Nvidia is touting around are hard to ignore. Vera Rubin will apparently deliver five times Blackwell’s compute performance, along with data movement between devices that supposedly exceeds global internet traffic. HBM4 bandwidth hits 22 TB/s — nearly triple what Blackwell managed with HBM3e. And, the Vera CPU more than doubles Grace’s memory bandwidth at 1.2 TB/s. Capacity scales accordingly — offering up to 288GB of HBM4 and 1.5TB of LPDDR5X through SOCAMM modules.
The next big thing?
Nvidia also used the event to reveal Alpamayo, a self-driving system built alongside Mercedes-Benz. The pitch is that it can “think like a human” behind the wheel, handling edge cases that typically confound autonomous systems.
AWS, Google Cloud, Microsoft, and Oracle Cloud Infrastructure have all lined up for early deployments in 2026, which speaks to how the market is reading this launch. Chips are already in production, with initial deliveries expected before year’s end.
Whether Vera Rubin ends up being as pivotal as past Nvidia generations is harder to tell. The competitive landscape has shifted considerably since Blackwell. Hyperscalers have poured resources into custom ASICs and homegrown silicon. Google’s TPUs keep advancing, while Amazon has doubled down on Trainium and Inferentia, and Microsoft now has its own AI accelerators in the mix. These alternatives may trail Nvidia on raw performance, but they offer cost efficiencies and deeper integration with proprietary stacks.
That said, Nvidia has a history of delivering on bold performance claims, and supply constraints across the industry aren’t easing up. Vera Rubin will almost certainly find eager buyers. The real question is whether demand will approach the near-monopoly Nvidia commanded in earlier infrastructure cycles, or whether the market has fragmented enough to reset expectations.
