# Gigawatt-Scale AI Factories: NVIDIA GB300 NVL72 System on Lambda Cloud

October 22, 2025• 4 min read

The path to superintelligence depends on infrastructure capable of sustaining trillion-parameter models and reasoning workloads at scale. That’s why Lambda is building gigawatt-scale [AI factories on NVIDIA GB300 NVL72 systems](https://www.nvidia.com/en-us/data-center/gb300-nvl72/) as the compute backbone for the next generation of training and inference.

Last week, [the first NVIDIA GB300 NVL72 systems](https://www.businesswire.com/news/home/20250923779565/en/Lambda-and-ECL-Bring-the-First-Hydrogen-Powered-NVIDIA-GB300-NVL72-Systems-Online) were stood up in Lambda’s high-density liquid-cooled datacenters. Each rack integrates 72 NVIDIA Blackwell Ultra GPUs and 36 [NVIDIA Grace CPUs](https://www.nvidia.com/en-us/data-center/grace-cpu/), featuring 37 TB of fast memory and 130 TB/s of NVIDIA NVLink Switch bandwidth, turbocharging frontier-scale research and enterprise AI deployment.

**Hydrogen-powered NVIDIA GB300 NVL72 racks**

## NVIDIA GB300 NVL72: What’s New

Compared to NVIDIA GB200 NVL72, NVIDIA GB300 NVL72 introduces significant architectural gains:

- **+50% HBM3e capacity (20TB per rack):** Supports trillion-parameter models with larger checkpoints, higher batch sizes, and extended context.
- **+1.5× higher dense FP4 performance & 2× faster attention operations:** increases inference efficiency and utilization for reasoning-heavy workloads.

These improvements translate directly into faster training cycles and more efficient inference at scale.

**NVIDIA GB300 NVL72 before InfiniBand XDR scale-out networking**

## Built for reasoning workloads

NVIDIA GB300 NVL72 is designed around the demands of large-scale inference. Reasoning workloads can require up to 100× more compute per query compared to oneshot inference workloads. NVIDIA GB300 NVL72’s expanded HBM3e memory and FP4 efficiency keep those workloads running at full speed.

Rack-scale NVLink5 provides 1.8 TB/s of bandwidth per GPU, connecting them into a single high-speed fabric for model parallelism with NVLink Switch. For multi-rack clusters, NVIDIA Quantum-X800 InfiniBand and ConnectX-8 SuperNICs deliver 800 Gb/s per GPU, reducing communication overhead during distributed training and inference.

Direct-to-chip liquid cooling ensures these systems run at peak utilization without thermal throttling, making it possible to deploy NVIDIA GB300 NVL72 at the density required for gigawatt-scale AI factories.

**NVIDIA GB300 NVL72 power shelves and out-of-band Spectrum switches**

## Why it matters:

For superintelligence and enterprise AI, NVIDIA GB300 NVL72 introduces the second-generation Transformer Engine with dynamic range management and fine-grain scaling techniques that optimize inference efficiency. With 1.5× more memory and up to 50% higher FP4 performance vs. NVIDIA GB200 NVL72. NVIDIA GB300 NVL72 is purpose-built to drive the next leap in trillion-parameter models and reasoning workloads.

NVIDIA GB300 NVL72 clusters deployed by Lambda combine compute, storage, orchestration, and observability into a single system. Each rack provides 3.84 TB of NVMe cache per GPU (276 TB per NVL72), configurable parallel file storage for high-throughput data access, and optional managed orchestration with Kubernetes or Slurm. Unified observability is delivered through Prometheus and Grafana for real-time metrics and alerting.

**NVIDIA GB300 NVL72 rack-scale systems**

## Lambda: The Superintelligence Cloud

Lambda Private Cloud delivers dedicated, bare-metal NVIDIA GPU clusters with low-latency networking and high-throughput interconnects, all in physically isolated data centers. Each rack is housed in a high-density, liquid-cooled facility engineered to maximize performance and ensure efficiency at the gigawatt scale.

Here’s what sets Lambda apart:

- **Data centers purpose-built for AI**  
  Optimized compute, storage, and networking in liquid-cooled data centers for scalable, efficient performance.
- **Single-tenant clusters with observability**  
  Secure, dedicated clusters with optional real-time monitoring, metrics, and reporting for production-grade operations.
- **Co-engineering excellence & embedded support**  
  Hands-on partnership with Lambda’s engineers, infrastructure specialists, and ML experts to reduce bottlenecks and accelerate results.

## The future of AI factories starts here.

With NVIDIA GB300 NVL72, Lambda enables the world’s most ambitious superintelligence labs and enterprises to train and serve next-generation models faster, more efficiently, and at gigawatt scale.