Nvidia just pulled the wraps off its next big AI platform — and it’s gunning straight past Blackwell.
Announced by CEO Jensen Huang on stage at CES on January 5, 2026, the new Rubin computing architecture is now in full production and scheduled to ramp in the second half of the year.
“Vera Rubin is designed to address this fundamental challenge that we have: The amount of computation necessary for AI is skyrocketing,” Huang told the audience. “Today, I can tell you that Vera Rubin is in full production.”
Named after astronomer Vera Florence Cooper Rubin, the platform is Nvidia’s latest response to the runaway compute demands of modern AI — from giant foundation models to “agentic” systems that run long, multi-step workflows.
Rubin replaces Blackwell, keeps Nvidia on top
Rubin is the successor to Blackwell, which itself followed the Hopper and Lovelace architectures. That upgrade cadence is a big part of how Nvidia became the most valuable corporation in the world, and Rubin is designed to keep that flywheel spinning.
The new chips are already spoken for by nearly every major cloud provider. Nvidia highlighted partnerships with:
- Anthropic
- OpenAI
- Amazon Web Services (AWS)
On the high‑performance computing side, Rubin systems will power:
- HPE’s Blue Lion supercomputer
- The upcoming Doudna supercomputer at Lawrence Berkeley National Lab
In other words: if you’re training or serving cutting‑edge models at scale, Rubin is likely somewhere in your future bill of materials.
Six‑chip architecture built for AI bottlenecks
Rubin isn’t just a new GPU. Nvidia is pitching it as a six‑chip architecture designed to work in concert and attack the real-world bottlenecks that show up in large AI clusters.
Key components include:
- Rubin GPU – the central accelerator for training and inference
- BlueField enhancements – targeting storage and data movement bottlenecks
- NVLink improvements – higher‑bandwidth interconnect between chips
- Vera CPU – a new CPU line aimed at agentic reasoning workloads
That focus on data movement and memory is deliberate. Modern AI models don’t just need flops; they need fast access to gigantic working sets of data, especially key‑value (KV) caches used to condense inputs over long contexts.
Nvidia’s senior director of AI infrastructure solutions, Dion Harris, pointed directly at those cache demands when explaining a new storage tier in Rubin.
“As you start to enable new types of workflows, like agentic AI or long-term tasks, that puts a lot of stress and requirements on your KV cache,” Harris told reporters, referring to the memory system used by AI models. “So we’ve introduced a new tier of storage that connects externally to the compute device, which allows you to scale your storage pool much more efficiently.”
That external tier is meant to let operators grow memory and cache capacity without constantly buying more GPUs just for their onboard RAM.
3.5x faster training, 5x faster inference
Under the hood, Rubin is about raw performance and efficiency.
According to Nvidia’s own tests, Rubin delivers:
- 3.5× faster than Blackwell on model training tasks
- 5× faster on inference
- Up to 50 petaflops of compute
- 8× more inference compute per watt
Those numbers won’t be independently verifiable until customers start publishing benchmarks, but on paper they’re a major jump — especially on efficiency, which is now as critical as top‑line speed for data center operators running into power and cooling limits.
Chasing a $3–4 trillion AI build‑out
Rubin lands in the middle of a global AI infrastructure arms race. AI labs and cloud platforms have been scrambling not just for Nvidia chips, but for the power, cooling, and real estate to run them at scale.
On an earnings call in October 2025, Huang estimated that $3 trillion to $4 trillion will be spent on AI infrastructure over the next five years. Rubin is clearly built to capture as much of that spend as possible.
With Blackwell still rolling out and Rubin now entering production, Nvidia is effectively overlapping its own product cycles — ensuring that, for hyperscalers and leading AI labs, the “latest Nvidia platform” is always just a purchase order away.
The big open questions now:
- How fast Rubin systems will actually ship in volume
- How much of the performance uplift translates to real‑world workloads
- Whether rivals in custom silicon and open accelerators can chip away at Nvidia’s lead
For the moment, though, Rubin looks like exactly what Nvidia wanted it to be: a bigger, faster, more efficient AI engine dropped into a market that can’t seem to get enough of them.
