NVIDIA Vera Rubin Goes Global With 10x Tokens-Per-Megawatt Gains And New Cloud Rollouts

NVIDIA's Vera Rubin platform is moving into full-scale deployment, with production of the NVL72 system ramping up across partners including CoreWeave, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure. Backed by around 300 partners and supported by a supply chain spanning more than 350 factory sites in 30 countries, the company is touting leadership on performance per watt and the lowest token costs for AI workloads.

Early benchmarks from CoreWeave on the DeepSeek-R1 workload report a tenfold increase in throughput per megawatt versus the Grace Blackwell NVL72, underscoring NVIDIA's focus on tokens-per-megawatt as a key metric for power-constrained AI data centres.

NVIDIA attributes the gains to extreme co-design across the stack. The rack-scale system integrates seven chips and five tray types into a single architecture — encompassing elements such as the Vera Rubin NVL72 accelerator tray, the Vera CPU rack, Spectrum-6 switching, and the Vera BlueField-4 — rather than assembling off-the-shelf parts. At the heart of the platform, the NVIDIA Vera CPU introduces a custom Olympus core that delivers twice the single-thread performance, triple the core-to-core bandwidth and 40% lower memory latency compared with competing chiplet designs, aiming to optimise agentic workloads.

Networking is split between scale-up and scale-out. Sixth-generation NVLink reportedly provides more than 2x throughput on complex jobs, with 3x lower latency and 10x higher packet rates than standard Ethernet. For scale-out, Spectrum-X Ethernet combines 102.4T Spectrum-6 switches with 1.6T ConnectX-9 SuperNICs, adaptive routing, congestion control, telemetry and open OS support, enabling up to 1.6x higher RDMA bandwidth versus off-the-shelf Ethernet. Early adopters of Spectrum-6 include CoreWeave, Microsoft, SpaceXAI and Tesla. NVIDIA's photonics switch with co-packaged optics — billed as the first of its kind in volume production — targets 5x lower power and a 10x higher MTBI than pluggable transceivers, with CoreWeave, Lambda and OCI among first users. Spectrum-XGS extends performance across sites, with NVIDIA citing up to 1.9x multi-site throughput. NVLink Fusion opens the platform to third-party XPUs on the NVLink scale-up fabric.

Design changes are also aimed at speed and sustainability. The Vera Rubin NVL72 compute tray is built without cables, fans or hoses, cutting tray assembly from hours to around one minute. A 45°C liquid-cooling inlet enables chiller-free, dry-cooler operation; NVIDIA says higher-temperature dry cooling and closed-loop liquid systems can save millions of gallons of water per megawatt annually in new AI facilities.

In Europe, Vera Rubin underpins an expanded Microsoft–Mistral alliance centred on sovereign AI. A new multibillion-dollar agreement will grow regional AI infrastructure, with Mistral adding GPU capacity based on thousands of the latest Vera Rubin GPUs. The rack-scale system, now ramping to full production and unifying seven new chips in a single design, will power the next generation of Mistral Compute and Microsoft's European AI infrastructure, drawing on tens of thousands of GPUs.

Europe's public sector and regulated industries are seeking AI that operates under local laws and governance, with data control and resilience baked in. NVIDIA notes that agentic systems can consume up to 15x more tokens than traditional AI applications, heightening the need for efficient infrastructure. The companies say Vera Rubin provides a full-stack foundation — compute, networking and software — for Europe's open-model ecosystem, supporting efficient training, inference and production deployment.

On the software side, Mistral Medium 3.5 and OCR 4 are now available in Microsoft Foundry, and Mistral models are integrated into Microsoft Copilot Studio. Through Azure Local and Foundry Local, organisations can run the same models and tooling across public cloud and customer-controlled environments. NVIDIA says that, compared with NVIDIA GB200 NVL72, Vera Rubin NVL72 can deliver up to 10x more tokens per megawatt and one-tenth the cost per million tokens, intended to help meet Europe's goals for innovation, economics and control.

CoreWeave, which co-engineered with NVIDIA to bring up the new generation of accelerated computing, became the first AI cloud to validate Vera Rubin NVL72 in production and publish live performance data. Its DeepSeek-R1 benchmark showed a 10x uplift in tokens per second per megawatt versus Grace Blackwell NVL72. The company has also deployed NVIDIA Spectrum-X Ethernet SN6600-LD as the switching fabric for Vera Rubin NVL72, built on the 102.4 Tb/s Spectrum-6 chip and using liquid-cooled designs to achieve 1.64 Pb/s per rack — double the capacity of previous air-cooled switches — with a fully non-blocking, multi-plane, multi-rail leaf-spine architecture to avoid oversubscription. Within the rack, a 260 TB/s all-to-all NVLink 6 fabric is designed to make the NVL72 behave as a single accelerator, matching the all-to-all communication needs of mixture-of-experts models like DeepSeek-R1.

Google Cloud has introduced its first A5X instances powered by Vera Rubin NVL72. The service is already in use by London-based Ineffable Intelligence, which is developing reinforcement-learning-driven "superlearner" systems that learn via continuous interaction and simulation rather than static datasets — a workload that places extreme demands on compute, memory bandwidth and interconnect latency. "The next era of research requires the next era of hardware," said Lasse Espeholt, cofounder of Ineffable Intelligence. "We feel privileged to work with the teams at NVIDIA and Google Cloud, who were able to grant us early access to Vera Rubin. The support across both teams has been unmatched; we were up and running almost immediately and are already testing infra for our superlearners." NVIDIA positions Vera Rubin NVL72 as purpose-built for agentic training, citing predictable latency, high utilisation and significantly more intelligence per dollar than prior systems. Google says A5X delivers up to 10x lower inference cost per token and 10x higher token throughput per megawatt than the previous generation. The instances pair NVIDIA ConnectX-9 SuperNICs with Google's next-generation Virgo networking, with scale claims of tens of thousands of Rubin GPUs within a single site and nearly a million across multi-site configurations.

Separately, benchmarks from AI cloud platform DeepInfra indicate that the NVIDIA Vera CPU more than doubles orchestration speed and supports more concurrent AI agents compared with other CPUs in production tests. DeepInfra, which processes nearly 5 trillion tokens per week with about 30% driven by agentic systems, reports support for up to 1.6x more concurrent agents at the same quality of service and up to 2.2x faster orchestration versus alternative CPUs. NVIDIA says these results reflect its co-design approach for agentic workloads, aiming to improve utilisation and cost efficiency for cloud providers.

With deployments broadening across hyperscalers and specialised AI clouds, NVIDIA is positioning Vera Rubin as the rack-scale foundation for the next wave of agentic, open and reinforcement-learning systems — prioritising performance per watt and cost per token as AI factories scale.

www.nvidia.com/en-gb/