Hardware & Infrastructure
GPU mix, datacenter footprint, network fabric, and geographic distribution across the cast.
GPU mix
| Company | Hardware focus |
|---|---|
| CoreWeave | H100/H200/B200 datacenter-grade, GB200 NVL72 systems |
| Crusoe | H100/H200/B200 datacenter-grade, GB200 NVL72 emerging |
| Nebius | H100/H200/B200 datacenter-grade |
| Lambda | A100/H100/H200/B200 datacenter-grade |
| Together.AI | H100/H200/B200 (owned and partner) |
| RunPod | Mix: consumer (3090/4090/5090), workstation, datacenter (A100/H100/H200) |
| Hyperbolic | H100/H200 + marketplace mix |
| Vast.AI | Broadest: 3090/4090/5090, A6000, A100, H100, H200 |
| TensorDock | 4090/5090, A100, H100, growing |
Datacenter footprint
- Multi-hundred MW per site or aggregate: CoreWeave, Crusoe (with Stargate buildouts).
- Hundreds of MW aggregate: Nebius, Lambda.
- Tens of MW aggregate or partner-based: Together, RunPod Secure Cloud, Hyperbolic.
- Distributed across many provider sites: Vast, TensorDock, RunPod Community.
Network fabric
- InfiniBand fabric supporting thousands of GPUs: CoreWeave, Crusoe, Nebius. Also Lambda for cluster offerings.
- InfiniBand at smaller cluster sizes: Together, RunPod Secure Cloud.
- Single-node only / no inter-node fabric: Vast, TensorDock, RunPod Community, Hyperbolic marketplace, most consumer-card listings broadly.
Geographic distribution
- US-anchored, expanding international: CoreWeave, Crusoe.
- European-anchored, expanding US: Nebius.
- US + Europe: Lambda, Together, RunPod.
- Global through providers: Vast, TensorDock.
Takeaway
Infrastructure capability tracks roughly with capital intensity. The fabric availability is one of the cleanest predictors of which customer workloads each company can serve. The next chapter compares pricing.