Nvidia introduces the Vera Rubin platform as its next major AI leap

Nvidia uses CES to signal its next era of AI infrastructure

Nvidia walked onto the CES stage with a message that was bigger than a single GPU. Nvidia positions the Vera Rubin platform as a rack-scale ‘AI factory’ that integrates compute, networking, DPUs, and storage to simplify deployment and lower operating cost, a claim the company explicitly makes in its CES announcement.

The platform’s stated goals are faster training, lower inference token cost, and smoother scale-up for MoE systems, outcomes Nvidia says Rubin enables, though real-world gains will depend on model type and deployment choices.

Vera Rubin is positioned as a leap beyond Blackwell

Nvidia argues Rubin reflects a shift in which inference economics matter as much as raw training throughput, and the company highlights lower cost-per-token for large production workloads as its commercial focus.

Even if real-world results vary by workload, the direction is clear: efficiency per dollar is now the headline metric, not just raw speed.

Extreme codesign is the core idea behind the platform

Instead of treating parts like add-ons, Rubin is built around tight integration across six chips. Nvidia pairs a new CPU with a new GPU, then designs the interconnect, NIC, DPU, and Ethernet switching around that duo.

In practice, this means fewer bottlenecks, fewer wasted watts, and less tuning pain when scaling from one server to an entire data hall. The platform story is the product.

The latest NVLink interconnect

Nvidia’s NVLink-6 interconnect increases per-GPU bandwidth (Nvidia quotes 3.6 TB/s per GPU) and supports rack topologies like the NVL72 to reduce cross-GPU communication overhead for large MoE training scenarios.

The takeaway is that Rubin is not only about faster math. It is about moving tensors and parameters fast enough that expensive GPUs are not waiting around.

The Vera CPU is designed for AI factories

Nvidia’s Vera CPU is described as a power-efficient host built for the realities of AI data centers. The company emphasizes custom cores, Arm compatibility, and high-bandwidth chip-to-chip links to the GPU.

That matters because modern AI stacks are increasingly CPU-limited in orchestration, preprocessing, and data movement. If the CPU cannot keep up, the best GPU looks average. Rubin tries to close that gap.

The Rubin GPU targets inference with a new Transformer Engine

Rubin’s GPU pitch leans hard into inference economics. Nvidia points to a third-generation Transformer Engine and support for very low-precision formats aimed at delivering high throughput per watt.

That is precisely where the market is going as AI moves from demos to daily production. If your model serves millions of queries, a slight cost reduction per token can become a significant operating savings over the course of a year.

Confidential computing is becoming a first-class feature

As models and data become more valuable, security can no longer be an afterthought. Nvidia says Rubin brings rack-scale confidential computing across CPU, GPU, and interconnect domains.

The practical outcome is that regulated industries and cautious enterprises can keep proprietary training and inference workloads safer, even in multi-tenant environments. The bigger story is that true security features now shape where AI can be deployed.

Reliability and serviceability are treated as performance multipliers

Rubin also leans into uptime. Nvidia discusses health checks, fault tolerance, and proactive maintenance through an updated RAS approach, along with a more modular rack design that supports faster assembly and easier servicing.

That may sound boring until you run thousands of GPUs. At that scale, small reductions in downtime and repair time translate into a significant performance win, as idle hardware represents wasted capital.

A new storage layer for long context AI reasoning

One of the more forward-looking ideas is inference context memory storage, designed to share and reuse key value cache data across infrastructure. This is aimed at multi-turn, agentic workloads where context is expensive and persistent.

By moving context management into an AI native storage layer powered by the DPU, Nvidia is signaling that the next bottleneck is memory and state, not just compute.

BlueField and ASTRA push toward secure multi-tenant AI factories

Rubin’s BlueField DPU is positioned as a control and security anchor for bare metal and multi-tenant deployments. Nvidia’s ASTRA concept aims to establish a trusted control point for provisioning, isolating, and operating large-scale environments without compromising performance.

If you are a cloud provider or an enterprise running shared clusters, this is the part that determines whether Rubin is easy to operate or a nightmare.

Spectrum Ethernet and photonics are about scaling without burning the grid

Networking is where AI factories either scale gracefully or hit a wall. Nvidia’s Spectrum Ethernet and Spectrum X photonics systems are pitched as higher efficiency, more resilient fabrics with better performance per watt.

The emphasis on co-packaged optics is a tell: power and signal integrity are now strategic constraints. Rubin is trying to make networking feel like an accelerator, not a tax on every training step.

Rubin arrives in multiple system shapes to fit real deployments

Nvidia frames Rubin as both rack-scale and server-scale. The Vera Rubin NVL72 rack-scale system combines 72 Rubin GPUs and 36 Vera CPUs, utilizing NVLink, NICs, DPUs, and switching, to form a unified AI factory.

The HGX platform targets more traditional server designs that still want NVLink benefits. The point is flexibility. Cloud builders, enterprises, and labs do not buy the same shape, but they all want the same platform advantages.

For a sense of how demand is already shaping Nvidia’s strategy, it’s worth a look at why Blackwell chips are selling fast even as analysts warn about the risks of heavy customer concentration.

The ecosystem rollout shows how fast Nvidia wants to move

Nvidia says parts of the Rubin stack are already entering production and that partner systems are expected to ship in the second half of 2026; these are vendor timelines and may vary by partner and region.

Server makers and software partners are lining up to ship tuned stacks, because nobody wants to integrate this alone. The real story is cadence. Nvidia is forcing the market to plan around its roadmap, and its rivals must match its tempo or risk losing market share.

If you’re curious about how Nvidia is pairing that rapid rollout with privacy promises, it’s worth taking a quick look at the company’s new AI tool, designed to track data while keeping it fully private.

What do you think about Nvidia introducing the Vera Rubin platform as its next major AI leap? Please share your thoughts and drop a comment.

This slideshow was made with AI assistance and human editing.

Don’t forget to follow us for more exclusive content on MSN.

Read More From This Brand:

• Nvidia’s 50,000-GPU deal with Samsung could mark a new AI milestone
• NVIDIA partners with Groq in strategic AI tech licensing agreement
• Nvidia creates a location tracking system to fight chip smuggling