AI at Scale: Switch Hosts CoreWeave’s Landmark NVIDIA GB300 Deployment

In the relentless race to scale artificial intelligence (AI), infrastructure must evolve as quickly as the models it supports. That evolution is now on full display at Switch, where one of the industry’s most advanced data center designs, known as the EVO AI Factory, is enabling a breakthrough deployment from AI cloud provider CoreWeave.

Switch’s latest innovation, founded on nearly two decades of thermal engineering leadership, is hosting the first-ever cloud deployment of the NVIDIA GB300 NVL72 platform. This milestone marks a new chapter in high-performance AI computing, made possible through collaboration between Switch, CoreWeave, Dell Technologies, and Vertiv.

A Legacy of Density, Reimagined for AI

Switch has long been synonymous with high-density, high-efficiency design. From the original SuperNAP’s patented T-SCIF containment to the current multi-campus network of Tier 5 data centers, company founder Rob Roy’s vision has continually pushed the envelope on what’s possible in data center cooling.

Now, with the debut of Rob Roy’s EVO AI Factories, Switch is setting a new benchmark. The hybrid air-and-liquid cooled environment is purpose-built to support up to 2 megawatts per rack, aligning closely with NVIDIA’s DGX and MGX roadmaps. This advanced architecture provides the thermal and electrical foundation needed for generative AI and accelerated computing at hyperscale.

Marc Ganzi, CEO of DigitalBridge, which acquired Switch in 2022, described the accomplishment succinctly:

“Switch has moved past concept and chatter to pure execution as the leading deployment partner to CoreWeave and NVIDIA with our innovative and patented EVO cooling system,” Ganzi said. “Truly remarkable and testimony to the team and Rob Roy’s vision.”

CoreWeave’s Leap to the GB300 NVL72

The latest advancement comes from CoreWeave, which last week became the first AI cloud provider to stand up the NVIDIA GB300 NVL72, a platform representing a quantum leap in inference performance, throughput efficiency, and GPU interconnectivity.

“This platform and CoreWeave’s ability to quickly bring it to market via its AI cloud platform are a major enabler for AI-driven filmmaking to become a reality for our industry,” said Naeem Talukdar, CEO of generative video startup Moonvalley, one of CoreWeave’s customers.

CoreWeave’s deployment of the GB300 is built atop a cloud-native, HPC-grade infrastructure stack. From its proprietary Rack LifeCycle Controller (RLCC) and “Cabinet Wrangler” dashboard, to integrations with Weights & Biases for real-time observability, every element is engineered to maximize the output of large-scale AI workloads.

It’s a continuation of CoreWeave’s aggressive roadmap: from pioneering early HGX H100 deployments, to supporting GB200 NVL72 systems, and now, pushing ahead with the GB300 NVL72, all within Switch’s AI Factory footprint.

What’s Inside the GB300 NVL72?

The GB300 NVL72 system is a rack-scale behemoth. Housed in a liquid-cooled, high-density configuration, the system integrates:

• 72 NVIDIA Blackwell Ultra GPUs: Featuring 1.5x the compute of standard Blackwell GPUs, optimized for AI reasoning and agentic inference.
• 36 NVIDIA Grace CPUs and 18 BlueField-3 DPUs: Enabling efficient CPU-GPU memory sharing and high-speed, secure cloud networking.
• 21 TB of HBM3e memory per rack: Allowing for much larger batch sizes and model parameters.
• 130 TB/s aggregate NVLink bandwidth: Supporting ultra-fast GPU-to-GPU communication within the rack.
• 800 Gb/s per-GPU InfiniBand connectivity: Powered by Quantum-X800 switches and ConnectX-8 SuperNICs, this ensures RDMA performance at massive scale.
• Secure multi-tenant infrastructure via NVIDIA DOCA: Delivering tenant isolation and line-speed data movement across network and storage fabric.

The platform’s performance is frankly staggering. According to CoreWeave, the GB300 delivers:

• Up to 10x better user responsiveness.
• 5x throughput per watt over previous Hopper-based systems.
• 50x higher inference output for reasoning models.

These metrics matter. AI development is increasingly bottlenecked not by innovation, but by the infrastructure’s ability to train and deploy trillion-parameter models quickly and reliably. GB300 NVL72 is the hardware answer to that bottleneck, and Switch’s EVO design is the thermal and electrical envelope that makes it viable.

Engineering at the Edge of AI

It's important to remember that CoreWeave’s platform is about more than just the chips. It’s about the systems built around them.

Through Kubernetes-native services (CKS), Slurm on Kubernetes (SUNK), real-time telemetry, and AI-integrated observability, the company has created an environment where performance tuning, uptime, and development cycles are tightly coupled. By streaming cluster health events into AI model monitoring tools, customers can trace every training hiccup to a physical cause, and then resolve it immediately.

This full-stack integration is how CoreWeave squeezed 2.86x better per-GPU inference performance from GB200 NVL72 compared to H100 systems earlier this year. That same approach is now being applied to GB300, with results that are poised to redefine inference efficiency.

A Collaborative Future for AI Infrastructure

The success of this new rollout is not attributable to any one player. CoreWeave, Dell, Vertiv, and Switch have together demonstrated what’s possible when AI-native infrastructure meets deep, integrated collaboration.

Speed to deployment, long a pain point in hyperscale infrastructure, is now a competitive differentiator.

And as generative AI, simulation, and autonomous systems continue to evolve, this new ecosystem of AI Factories, GB-class systems, and AI-optimized cloud services may become the new normal.

The Bottom Line

The future of AI is now actively being built, rack by rack, inside Switch’s EVO AI Factories.

By combining cutting-edge thermal engineering, liquid cooling, and rack-scale power with AI-specialized cloud services from CoreWeave, the industry is witnessing a new phase of infrastructure maturity.

For data center builders, it’s now not just about keeping up with AI: it’s about enabling it to accelerate.

As AI deployments grow in both ambition and complexity, facilities like these will be the proving grounds; not just for models, but for the data center architectures that support them. With this “industry-first” GB300 deployment, CoreWeave and Switch are already running that future today.

At Data Center Frontier, we talk the industry talk and walk the industry walk. In that spirit, DCF Staff members may occasionally use AI tools to assist with content. Elements of this article were created with help from OpenAI's GPT4.

Keep pace with the fast-moving world of data centers and cloud computing by connecting with Data Center Frontier on LinkedIn, following us on X/Twitter and Facebook, as well as on BlueSky, and signing up for our weekly newsletters using the form below.