Nscale activates first NVIDIA Vera Rubin NVL72 AI racks
Nscale has activated its first NVIDIA Vera Rubin NVL72 racks, each with 72 GPUs and 36 CPUs. The global rollout tests factory-built networking and NVIDIA's vendor-reported 10x token-efficiency claim.

Why it matters
- November 18, 2026: Nscale activated its first NVIDIA Vera Rubin NVL72 rack-scale systems.
- Each NVL72 rack contains 72 Rubin GPUs and 36 Vera CPUs.
- Nscale says it is the first to deploy Dell's fully integrated, liquid-cooled L11 networking rack.
- The Dell IR9148 platform includes eight NVIDIA SN6600-LD switches.
- NVIDIA claims up to 10 times more tokens per megawatt than GB200 NVL72, but did not measure the claim on Nscale systems.
By James Calloway
Nscale activated its first NVIDIA Vera Rubin NVL72 rack-scale AI systems on November 18, 2026; each rack contains 72 Rubin GPUs and 36 Vera CPUs. The global rollout will continue over the coming months, according to Nscale, while Dell PowerEdge XE9812 systems add capacity for customers training and serving large models.
Nscale's announcement did not disclose how many racks came online, where they operate, which customers use them or what pricing applies. Those omissions prevent an assessment of the rollout's commercial scale.
What exactly is inside each Vera Rubin rack?
Each NVL72 rack contains memory and high-speed connections designed to operate as one rack-scale supercomputer. In plain terms, Nscale is selling access to a coordinated block of accelerators, not a collection of independent servers.
Nscale's companion technical post identifies five workload categories:
- Large distributed training, which spreads model training across many GPUs.
- Long-context reasoning, which handles long inputs and multistep outputs.
- Reinforcement learning, which trains systems through reward signals.
- Agentic inference, which runs AI systems that perform multistep tasks.
- Low-latency services processing high volumes of tokens, the text units models process.
Nscale says this mix makes the network and operating model central to the deployment. Training runs and production inference both require many accelerators to exchange work without stalling on communication or operations.
The practical takeaway is that the hardware specification is clear, but the rollout's size and customer uptake remain undisclosed.
What does Dell's factory-integrated rack change?
Nscale says it is the first in the industry to deploy Dell's fully integrated, liquid-cooled L11 networking rack. Built on Dell's IR9148 platform, it contains eight NVIDIA SN6600-LD switches, shuffle boxes, power shelves, an integrated rack controller and an in-rack coolant distribution unit.
The rack combines network equipment, power, controls and cooling in one installation. The announcement names the shuffle boxes without defining their function, so their exact role cannot be assessed from the source.
Dell integrates and tests the populated rack at the factory, which eliminates onsite network-rack assembly. “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,” said Ihab Tarazi, CTO of Compute and Networking in Dell Technologies’ Infrastructure Solutions Group.
That statement comes from a Dell executive whose systems the deployment uses. It supports the factory-assembly claim, but the announcement provides no independent timeline or comparison showing how much faster factories finish the work.
The practical takeaway is that factory integration reduces onsite work, but Nscale has not quantified the deployment-time advantage.
How do NVLink and Ethernet connect the system?
Dell's system overview describes the XE9812 as the Vera Rubin successor to the XE9712. It uses direct liquid cooling, a rack-scale design and NVLink-interconnected GPUs, according to Dell.
Within each rack, NVLink supplies the scale-up fabric, meaning tightly coupled GPU-to-GPU communication. Between racks, Nscale uses a RoCE scale-out network, meaning high-throughput, low-latency links across a larger cluster.
RoCE stands for RDMA over Converged Ethernet, a standards-based method for moving distributed AI traffic with low latency. Nscale says its Ethernet fabric also incorporates high-performance switching, congestion management, telemetry and automation.
The stated goal is to keep GPUs working on parallel jobs instead of waiting on memory, networking or operational bottlenecks. Nscale did not publish measured latency, bandwidth or utilization for the first live racks, so the architecture remains an unverified operating claim in the announcement.
The practical takeaway is that the network has a clear division of labor, but its real performance remains unmeasured.
How will Nscale manage a global rack fleet?
Nscale puts each rack into its cloud operating model as soon as it comes online. Fleet Operations handles:
- Enrollment, configuration, validation and burn-in, meaning initial testing before service.
- Network integration, maintenance, repair and retirement.
A separate Control Center applies software, firmware, network and security baselines before validated systems enter service through an automated process. Nscale also links infrastructure telemetry across compute, networking, storage and orchestration with automated responses.
Nscale's Radar API, a programming interface for operational data, exposes availability, maintenance and repair status, device state, capacity and other events to customers. According to Nscale, those controls give customers a common view of capacity while providing a consistent process for bringing additional systems into service.
Nidhi Chappell, Nscale’s president of AI infrastructure, said the company expects “the rollout to operate as one production environment under the same fleet automation and observability as the rest of its cloud.” Nscale says customers can contact the company to place orders or discuss deployment timing.
The announcement does not specify burn-in duration, fleet size, maintenance intervals or service-level targets. Those omissions make Nscale's operating promises difficult to compare with those of other AI cloud providers.
The practical takeaway is that Nscale has described the control plane, but not its measurable service guarantees.
What does NVIDIA's 10-times claim prove?
On NVIDIA's product page, the vendor reports up to 10 times more tokens per megawatt than GB200 NVL72 for its stated Kimi-K2 Thinking configuration. That configuration uses 32K input and 8K output sequence lengths, and NVIDIA says the inference performance is subject to change.
That is a vendor comparison, not a result measured on Nscale's systems. The announcement calls token throughput within a fixed power footprint a central economic consideration for large-scale inference.
The comparison does not establish training performance, application-level latency or total cost for Nscale customers. Nor does the announcement disclose rack count, deployment sites, customer names, pricing, measured power draw or utilization. The efficiency figure therefore cannot support a capacity or price estimate for Nscale's service.
The practical takeaway is that NVIDIA's figure supports a product claim, not a verified measure of Nscale's performance.
Can Nscale scale the rollout cleanly?
The deployment joins a new GPU generation to less visible systems work: factory-integrated liquid-cooled networking, NVLink inside racks, Ethernet scale-out and shared fleet controls. Nscale's announcement frames the next operational test as whether the company can repeat that package as additional systems arrive without fragmenting its global production environment.
Nscale has put the first systems online and supplied a detailed architecture. Its next disclosures should settle the harder commercial questions: how many racks will run, where, for whom and at what measured efficiency.
Until then, the rollout remains a deployment milestone rather than evidence of a repeatable global service.
Source: unite.ai
More from James Calloway
Show full bio
News editor covering people, launches and day-to-day practice in artificial intelligence and machine learning at AI In Context.
791 articles