🚀⚙️ Nebius locks in NVIDIA Vera Rubin NVL72 to power agentic and reasoning AI at scale, starting H2 2026

Nebius has confirmed it will integrate NVIDIA Vera Rubin NVL72 into its AI cloud platform starting in the second half of 2026, becoming one of the first AI cloud providers to deploy Rubin-class systems for production workloads.

This is not a routine hardware upgrade.

It’s a strategic positioning move around how next-generation AI will actually run.

Nebius plans to deploy Vera Rubin NVL72 across its U.S. and European data centers, integrating the platform directly into Nebius AI Cloud and its enterprise inference and post-training platform, Nebius Token Factory.

The target workload is explicit:

agentic AI and reasoning-heavy systems, not just model pretraining.

That distinction matters.

Agentic and reasoning AI place very different demands on infrastructure than traditional training runs:

persistent inference,

high token throughput,

tight latency constraints,

and continuous multi-tenant operation.

Vera Rubin NVL72 is designed precisely for that regime.

As a reminder of what Rubin represents:

it’s a rack-scale AI supercomputer, not a single accelerator.

The platform combines Vera CPUs, Rubin GPUs, sixth-generation NVLink switching, ConnectX-9 networking, BlueField-4 DPUs, and Spectrum-X 102.4T CPO into a tightly coupled system built for large-scale AI factories.

According to Jensen Huang, Vera Rubin has already entered full production, positioning it as the successor to Grace Blackwell for the next phase of AI infrastructure buildout.

From a performance and economics standpoint, the shift is material:

Rubin GPUs deliver multi-fold gains in inference throughput,

enable large Mixture-of-Experts models to be trained with a fraction of the GPUs,

and dramatically reduce per-token cost at scale.

Just as importantly, Rubin introduces rack-level confidential computing, making it viable for regulated, privacy-sensitive, and enterprise-grade deployments.

Nebius is also explicit that Rubin will complement, not replace, its existing capacity built on NVIDIA GB200 NVL72 and Grace Blackwell Ultra NVL72.

That signals a multi-architecture strategy where workloads are matched to the most cost-efficient and performance-appropriate platform, rather than forcing everything onto a single generation of hardware.

The timing is critical.

H2 2026 may sound distant, but for data center power, networking, cooling, regulatory approvals, and rack-scale integration, this is early.

By the time agentic AI becomes a mainstream production requirement, Nebius intends to already be deployed, tested, and operational.

The broader takeaway is simple:

the competitive frontier in AI is shifting from model access to system-level delivery of sustained reasoning and agentic compute.

Models can be copied.

Weights will leak.

Algorithms converge.

But rack-scale infrastructure, deployed early and delivered reliably across regions, does not.

If agentic AI becomes the dominant workload over the next cycle, the question isn’t who has the best demo.

It’s who already has the infrastructure live when demand shows up.

📮 Ongoing analysis on AI infrastructure, compute economics, and NVIDIA’s platform evolution — focused on the assets that determine scale before the market prices them in.

#Nebius #NVIDIA #VeraRubin #NVL72 #AIInfrastructure #AgenticAI #ReasoningAI #CloudComputing #DataCenters