Networks & InfrastructureNscaleNvidia Vera RubinDellNvl72

Nscale powers up first Nvidia Vera Rubin NVL72 rack

Nscale has powered up its first Nvidia Vera Rubin NVL72 rack on Dell hardware, claiming the industry's first factory-built liquid-cooled network rack as a global rollout begins.

4 min read

Nscale brings Nvidia Vera Rubin NVL72 rack onlineNetworks & Infrastructure
Nscale brings Nvidia Vera Rubin NVL72 rack onlineTama66 / pixabay

Why it matters

  • Nscale brought online its first Nvidia Vera Rubin NVL72 rack, built on a Dell PowerEdge XE9812 system.
  • Nvidia claims Rubin delivers 5x inference and 3.5x training performance versus Blackwell.
  • The NVL72 rack holds 36 CPUs and 72 GPUs, is 100% liquid-cooled, and Nvidia says modular trays cut install time from two hours to five minutes.
  • Nscale claims the industry's first deployment of Dell's liquid-cooled L11 network rack on the IR9148 platform, pairing eight Nvidia SN6600-LD switches.
  • CoreWeave, Oracle and Google Cloud have also brought Vera Rubin chips online; Nscale expects a global rollout in the coming months.

The story

AI cloud operator Nscale has brought its first Nvidia Vera Rubin NVL72 rack-scale system online, claiming an industry first for a factory-built, fully liquid-cooled networking rack supplied by Dell. The company describes the deployment as the start of a "worldwide deployment" over the coming months, though it has not disclosed the data center where the rack now runs.

The compute side is built on a Dell PowerEdge XE9812 system, integrated into Nscale's own cloud platform. The networking claim is the more specific one: Nscale says it is the first operator to deploy Dell's fully integrated, liquid-cooled L11 network rack, built on the Dell IR9148 platform. That rack combines eight Nvidia SN6600-LD switches with shuffle boxes, power shelves, an integrated rack controller and an in-rack coolant distribution unit.

What is actually in the rack?

The Vera Rubin platform comprises six chips: the Vera CPU, the Rubin GPU, the NVLink 6 Switch, the ConnectX-9 SuperNIC, the BlueField-4 DPU and the Spectrum-6 Ethernet Switch. Vera succeeds Nvidia's Grace CPU; Rubin succeeds Blackwell.

Nvidia claims Rubin delivers 5x the inference performance and 3.5x the training performance of Blackwell. Those are vendor figures for the new silicon, not measured results from Nscale's installation — no throughput numbers from the live deployment have been published.

The NVL72 system, Nvidia's rack-scale offering combining 36 CPUs and 72 GPUs, differs from earlier generations in two ways relevant to operations. It is 100 percent liquid-cooled, and it uses cable-free modular tray designs. Nvidia claims the modular trays cut installation times from two hours to five minutes — a 96 percent reduction that, if it holds in production, would compress deployment labor for an operator rolling out racks globally.

The integration pitch

Dell positions the L11 rack as a way to move work from the data center floor to the factory. Ihab Tarizi, CTO of compute and networking in Dell's Infrastructure Solutions Group, said: "Nscale's deployment marks an industry first for a factory-built, fully integrated, liquid-cooled networking rack. Dell PowerRack integrates and validates compute, networking, power, and cooling in the factory, reducing the work required onsite and helping Nscale bring its high-performance compute infrastructure into production faster. It's an important step in delivering complete AI infrastructure at rack scale."

The commercial appeal is speed to revenue: racks that arrive pre-validated can go from loading dock to serving customers faster than kit assembled and cabled on site. What Dell's statement does not address is cost, power density requirements or how the in-rack coolant distribution unit performs under sustained load — details that will matter to any operator weighing the same route.

Who said what about the rollout

Nidhi Chappell, president of AI infrastructure at Nscale, framed the deployment around demand for reasoning and agentic workloads. "We have brought up our first Nvidia Vera Rubin NVL72 racks, and over the coming months we expect to continue a global rollout as a single production environment, managed by the same fleet automation and observability as the rest of our cloud," Chappell said. "Because we operate the full stack from ground to cloud, we control that timeline, and customers get next-generation compute at the scale and efficiency frontier workloads demand."

The "full stack from ground to cloud" claim is Nscale's core differentiator argument: owning facilities through to the orchestration layer means no third party dictates deployment pace. It also means Nscale carries the full capital burden of that pace itself. The company has not disclosed how many racks the global rollout comprises, where they will be sited, or when customers can first run workloads on them.

Where the competition stands

Nscale is not alone in moving to the new generation. CoreWeave, Oracle and Google Cloud have also brought Vera Rubin chips online, putting Nscale in the company of hyperscalers with far larger balance sheets. Its edge, if any, is integration speed on the Dell reference design rather than scale.

Nscale says the global rollout continues over the coming months as a single production environment under one fleet-management layer — a timeline it controls, and one it has yet to bind to a date.

Also reported

Share this article:

« PreviousNext »

More from Rebecca Stone

Rebecca Stone

Show full bio

Correspondent covering technology and product news in Carrier business, 5G and 6G, fiber and broadband at Telecom Gazette.

560 articles