# NVIDIA's Vera Rubin NVL72 coming to Lambda's Superintelligence Cloud

January 9, 2026• 3 min read

At Lambda, we build supercomputers that enable AI teams to deliver next-generation, frontier models. Today, we’re announcing the next evolution of our infrastructure: the [NVIDIA Vera Rubin platform](https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/), which will serve as the core building block for Lambda Superclusters.

The Vera Rubin NVL72 rack operates as a single, massive GPU and, when networked into a Supercluster, unlocks scale and efficiency previously out of reach. This enables teams to map models, validate performance, and run at production scale.

**Why this matters**  
Engineers designing and training the next-generation reasoning and agentic models need two critical resources: performance and efficiency. The NVIDIA Vera Rubin architecture delivers both. Vera Rubin NVL72 is a single unified rack with a 72-GPU [NVIDIA NVLink](https://www.nvidia.com/en-us/data-center/nvlink/) domain. It reduces cross-die communication overhead, expands the scale-up memory pool, enables model-parallel training and inference to behave more like single-node runs, and delivers up to 10x higher token throughput per watt and costs one-tenth as much per million tokens as NVIDIA Blackwell.

**What teams can do:**  
- Train and iterate on 10T+ parameter models and production-scale Mixture of Experts (MoE) with significantly lower communication overhead, shortening the time to convergence.  
- Serve real-time, high-throughput inference and long-context reasoning with rack-local HBM4 and high intra-rack bandwidth, which keep KV caches and model state on ultra-fast memory.  
- Run reinforcement learning (RL) actor-learner loops and HPC collectives with predictable, low latency on racks that behave like one NVLink domain, simplifying mapping, scheduling, and synchronization.

**The building block: NVIDIA Vera Rubin NVL72**  
Vera Rubin NVL72 is a full-stack, rack-scale AI supercomputer purpose-built for massive-scale AI.

**Core hardware:**  
- **GPUs:** 72 NVIDIA Rubin GPUs within a single NVLink domain  
- **CPUs:** 36 NVIDIA Vera CPUs integrated into Vera Rubin Superchips  
- **DPUs:** 18 NVIDIA BlueField-4 DPUs that enable secure, software-defined infrastructure  
- **Scale-up:** NVIDIA NVLink 6 Switch trays that create a high-bandwidth scale-up fabric  
- **Scale-out:** [NVIDIA Spectrum-X Ethernet](https://www.nvidia.com/en-us/networking/spectrumx/) or [NVIDIA Quantum-X800 InfiniBand](https://www.nvidia.com/en-us/networking/products/infiniband/quantum-x800/) scaling to 100,000+ GPUs

**Key technical specifications:**  
- **Memory:** Up to 20.7 TBHBM4 per rack with bandwidth targets up to 22 TB/s per GPU  
- **Fabric:** NVIDIA NVLink 6 providing 3.6 TB/s GPU-to-GPU connectivity within the rack, backed by [NVIDIA NVLink Switch](https://www.nvidia.com/en-us/data-center/nvlink/) scale-up fabric  
- **Networking:** External connectivity and DPU offload via [NVIDIA BlueField-4](https://www.nvidia.com/en-us/networking/products/data-processing-unit/) 800G DPUs and [NVIDIA ConnectX-9](https://www.nvidia.com/en-us/networking/products/ethernet/supernic/) 1600G SuperNICs for inter-rack scale-out  
- **Cooling:** 100% direct-to-chip (D2C) liquid-cooled compute and switch trays

**Lambda Supercluster**  
Lambda builds Superclusters — single-tenant, high-density platforms engineered to run mission-critical AI at scale. NVIDIA Vera Rubin NVL72 further enhances Supercluster value:
- **Intra-rack scale-up:** The Vera Rubin NVL72 NVLink 6 design creates a domain where all 72 GPUs access memory at near-local speeds, reducing communication overhead for model-parallel training and inference and making MoE routing far cheaper in latency and bandwidth.
- **Inter-rack scale-out:** Using ConnectX-9 1600G SuperNICs and BlueField-4 800G DPU fabrics, we bind thousands of Vera Rubin NVL72 racks into a non-blocking fabric that delivers linear scaling for very large foundation models.
- **Operational abstractions:** Lambda’s control plane treats Vera Rubin NVL72 racks as unified building blocks. Engineers can provision multi-rack Superclusters with the same simplicity as a single instance and run them with reservation workflows and our managed orchestration.

**Facilities and cooling**  
These high-density racks require purpose-built data centers with full liquid cooling and robust power delivery. Lambda’s Supercluster facilities are engineered to support sustained high-power operation, featuring redundant plumbing and comprehensive monitoring of power and thermal performance. We provide 24/7 co-engineering and operations support to keep workloads running at peak efficiency.

**Availability and access**  
We’re working closely with NVIDIA to deploy the Rubin Platform in Lambda Superclusters. Production availability is planned for the second half of 2026.