Section C · Operations

Infrastructure & GPU Mix

Crusoe's GPU fleet has scaled rapidly through 2024-2026. The infrastructure focus mirrors the larger frontier-AI competitors — current-generation cards, InfiniBand fabric, liquid cooling.

GPU fleet

Crusoe's GPU inventory is heavily concentrated in current-generation datacenter hardware:

  • H100/H200: The deployed bulk during 2023-2025 ramp.
  • Blackwell (B100/B200/GB200): Aggressive deployment through 2024-2026, particularly tied to Stargate-scale capacity.
  • Some legacy A100. Smaller share of the fleet.

Consumer cards are not part of the product. The customer base doesn't want them, and the energy-arbitrage thesis at scale doesn't work for low-power-per-dollar hardware.

Network fabric

Within a datacenter cluster:

  • InfiniBand-based GPU-to-GPU networking, supporting multi-node training.
  • NVIDIA-reference network topology in the larger consolidated sites.
  • High-throughput storage networking for parallel filesystem access.

The network architecture supports the multi-thousand-GPU training cluster sizes that frontier customers need. This is one of the operational capabilities that distinguishes Crusoe from smaller energy-arbitrage plays.

Cooling and power infrastructure

Modern GPUs (especially Blackwell-generation) push thermal management:

  • Liquid cooling is increasingly standard for new buildouts.
  • Direct-chip cooling and rear-door heat exchangers in some configurations.
  • High-density power distribution to support 100+ kW per rack.
  • Redundant power infrastructure where customer SLAs require it.

The energy-led siting strategy adds operational complexity: behind-the-meter sites require backup power for grid-disconnect events, careful integration with the host energy source, and robust monitoring of the unusual power profile.

Scaling rate

The pace of capacity additions through 2024-2026 is among the fastest in the industry. Drivers:

  • Stargate-aligned buildouts at multi-hundred-MW scale.
  • Additional sites supporting growing customer base.
  • Generation rollovers (H200, Blackwell) overlaying existing capacity.

Execution at this pace is non-trivial. Crusoe's organizational growth has had to keep up.

Takeaway

Crusoe's hardware footprint is firmly in the frontier-tier of neoclouds. The energy-led strategy provides cost basis advantage; the infrastructure quality supports the largest training workloads. The next chapter looks at the ESG narrative — the story Crusoe tells about emissions and the criticism that follows.