Name: Nvidia Kyber / Rubin Ultra and NVL576 on display — 576 GPUs in a single rack with Vera CPUs
Uploaded: 2026-02-07T00:14:04.000Z
Duration: 5 min 43 s
Description: Thousands of YouTube videos with English-Chinese subtitles! Now you can learn to understand native speakers, expand your vocabulary, and improve your pronunciation...

At the heart of Nvidia's latest innovation is the upcoming NVL 576 Rack, a single cabinet housing 576 Rubin Ultra GPUs paired with variseries CPUs.

All of this is built on the new Kyber rack architecture, representing a fundamentally new scale of computing density, power, cooling and system design.

This isn't just faster hardware, it's the kind of machine data centers build when they're thinking AI factory, not just AI server.

The Kyber rack generation replaces the previous Oberon style racks, moving away from conventional horizontally mounted trays.

Instead, Kyber uses vertical blade modules and tight networking midplanes, optimizing space and performance.

Each of the four pods inside the rack takes in 18 blades, with each blade loading up to eight Rubin Ultra GPUs and accompanying VARICPU's.

This configuration adds up to an astonishing 576 GPUs in a single rack, providing unprecedented computing power.

High speed NVLink and custom midplanes enable each GPU to communicate with many others with very low latency.

This makes the rack behave more like one huge compute plane than dozens of separate boxes, a truly game changing approach.

With 576 high end GPUs in one rack, you no longer need dozens of racks to build a large model training or inference system.

The NVL576 aims for exaflop class compute in one footprint, a massive scale in a single compact unit.

This rack is planned to draw up to 600 kilowatts, far above current rack power budgets, necessitating a power and cooling rethink.

Nvidia is shifting to 800 VDC power delivery with reducing copper volume and improving efficiency a critical advancement.

Full liquid cooling solutions are integrated with the entire system, designed to handle the extreme thermal loads.

Vera CPUs are tightly paired with Rubin GPUs, meaning the rack is built not just for throughput but for reasoning class tasks and agentic AI.

We're not just ramping up raw flops, we're building for next gen AI workloads including large context inference.

With 576 GPUs all in one rack, latency and bandwidth between GPUs are paramount, requiring a networking and interconnect redesign.

NVLink switchblades, high bandwidth midplanes and optimized intra rack communication are critical to prevent bottlenecks for large language models.

Multimodal AI or Agentix systems, the NBL576 rack offers unprecedented capability.

Teams can train or serve models at a scale they use to build across many racks or even data centers, all within a single unit.

Inference latency drops, context windows grow, and model size and complexity can increase without splitting across many systems.

However, this advanced technology comes with major infrastructure demands requiring significant upgrades up to 600kW means facilities need upgraded electrical supply, high voltage distribution lines and special bus bars.

At that density, water cooling or immersion cooling is no longer optional.

Traditional air cooled systems won't suffice.

The rack footprint isn't just the cabinet.

Sidecars for power conversion, chillers and monitoring gear must be considered.

Deploying such racks takes planning, capital, skilled operations and time to integrate, representing a big leap for many organizations.

Enterprises aiming to build large scale AI factories now have a target rack of compute, consolidating into fewer ultra dense racks.

Developers of generative AI, multimodal reasoning systems and real time inference pipelines gain a new lever.

More local compute per rack, fewer network hops and more scale are now achievable.

Accelerating innovation Data center design becomes even more critical with power, cooling, networking and maintenance all evolving.

For chip makers and infrastructure vendors, this rack signals a shift.

The bottleneck is no longer just more gpu.

It's how you distribute power, how you cool and how you interconnect.

If you're building something that only needs a few GPUs or moderate inference loads, this type of rack may be overkill.

But if you're planning AI at a large scale, you should consider what it means when one rack can do what formerly required dozens.

GPU capability matters, but so does power architecture, interconnect, cooling and software model design needs to adapt.

With such dense racks, you can push models that need high memory bandwidth and large context windows.

Planning ahead is crucial if you're in procurement, operations or strategy.

Power upgrades, cooling systems and rack level design must be aligned with future racks like the NVL576.

In summary, the Kyber, Rubin Ultra and NBL576 rack is a preview of the next frontier of AI infrastructure.

It takes the concept of cluster and shrinks it into a single rack with ultra dense computer, bespoke power cooling and high bandwidth interconnect.

While planned for 2027, it sets a target for what AI factory hardware will look like, signaling how infrastructure and compute scale will evolve together.