Section A · Origin & scale

Infrastructure

CoreWeave's physical footprint is what separates it from competitors. Multiple large datacenters, tens of thousands of latest-generation GPUs, NVIDIA-reference networking. The capital intensity is real and the moat is in the inability of newcomers to replicate this overnight.

Datacenter footprint

CoreWeave operates from a portfolio of datacenter sites across multiple North American regions. The mix combines:

  • Owned datacenters built or fully renovated for GPU compute.
  • Co-location relationships with established datacenter operators where rapid expansion is needed.
  • Custom power and cooling configurations to support the very-high-density GPU racks current-generation training requires.

Total capacity has grown rapidly through 2023-2025. Specific MW figures shift as buildouts complete; the multi-hundred-MW order of magnitude is the right mental model. New regional sites have been added in 2024-2025 to meet customer commitments.

GPU inventory

The GPU fleet is heavily concentrated in current-generation datacenter cards:

  • H100: The bulk of the deployed fleet during 2023-2024 ramp.
  • H200: Significant deployment through 2024-2025.
  • B100/B200 (Blackwell): Early and substantial allocation; CoreWeave was among the first to deploy Blackwell at meaningful scale through 2024-2025.
  • GB200 NVL72: NVIDIA's rack-scale Blackwell systems are being deployed for the largest training clusters.
  • Some A100 legacy supply remains.

CoreWeave generally does not list consumer cards. The customer base doesn't want them; the operational profile doesn't justify them.

NVIDIA reference designs

CoreWeave deploys NVIDIA reference designs for cluster architecture — DGX SuperPOD-style configurations or close equivalents. The reference designs specify:

  • GPU-to-GPU networking topology (NVLink within nodes, InfiniBand between nodes).
  • Storage architecture (parallel filesystems for large dataset access).
  • Software stack (NVIDIA's Base Command, MagnumIO, etc.).
  • Operations practices (failure-domain configuration, monitoring, etc.).

The reference-design adherence matters: customers training frontier models want validated configurations that match what works at NVIDIA's own benchmarking. CoreWeave clusters deliver near-reference performance, which is why they capture the marquee training deals.

Network fabric

The network fabric is one of CoreWeave's biggest operational investments. Inside a cluster:

  • NVLink + NVSwitch: GPU-to-GPU within a server, providing hundreds of GB/s of bandwidth.
  • InfiniBand HDR / NDR: Between servers within a rack and between racks. NDR is 400 Gb/s per port; large clusters have many ports per server.
  • NVIDIA's Quantum-2 switches: The InfiniBand switching layer.
  • Ethernet for control plane and external traffic.

This network is what enables multi-thousand-GPU distributed training. Without it, you can't run the all-reduce operations that backpropagation requires. Marketplace competitors (Vast, RunPod Community) can't deliver this fabric.

Power and cooling

Modern GPUs draw enormous power — an H100 server with 8 GPUs and supporting infrastructure consumes 10+ kW continuously. A rack of multiple such servers exceeds the per-rack power density most legacy datacenters were built for.

CoreWeave invests in:

  • High-density power distribution.
  • Liquid cooling for Blackwell-generation hardware (Blackwell's thermal profile pushes traditional air cooling past its limits at scale).
  • Direct chip cooling and rear-door heat exchangers in some configurations.
  • Power sourcing relationships that can deliver tens of MW per site reliably.

This power-and-cooling capability is itself a moat. Building a new datacenter capable of supporting modern AI training densities takes 18-24 months minimum.

Expansion strategy

CoreWeave's expansion through 2024-2026 has been customer-led: capacity is being built to meet specific signed commitments, not on speculation. This pattern reduces capacity risk (won't build empty datacenters) but creates execution risk (customers expect delivery on aggressive timelines).

The expansion includes:

  • New North American sites.
  • International expansion exploration (Europe, etc.).
  • Continued GPU procurement to fill committed allocations.
  • Operational hiring to staff new sites.

Takeaway

CoreWeave's infrastructure is the substance behind the company's strategic claims. The fleet, network fabric, and power/cooling capability are years of capital and execution work that can't be quickly replicated by newcomers. The next chapter looks at the customer side: who's actually committing to multi-year deals at this scale.