Enterprise - Last verified June 2026

Enterprise GPU cloud (1,000+ employees, 256+ GPUs)

At enterprise scale the conversation moves from rate cards to multi-year reserved capacity contracts, InfiniBand-first networking, MLOps governance, and the question of whether to keep training in cloud or rack your own. Monthly spend at this band lands between $1M and $40M+.

Vendor shortlists

CoreWeave - HGX H100 / H200 / GB200 clusters at scale with InfiniBand-first networking.
Oracle (OCI) - bare-metal H100 / H200 / B200 SuperClusters with RDMA networking.
AWS + EFA + Capacity Blocks for ML when the rest of the estate is on AWS and the integration value justifies the premium.
Azure - ND H100 v5 SuperPOD with Reserved Instances for Microsoft estates.
Google Cloud - A3 H100 + TPU v5p / v5e as a parallel accelerator path.
Crusoe - H100 / H200 / B200 with sustainability positioning.

Reservation strategy

Multi-year reservations are the procurement default at this band. CoreWeave and Crusoe routinely transact 3-year HGX clusters; Oracle sells 1-3 year bare-metal SuperCluster contracts; AWS Savings Plans and Reserved Instances both have 1 and 3 year terms. The deepest discounts assume committed-spend volume and a payment-up-front posture. Vendor account engineering and SRE coverage are negotiated in the same contract, not separately.

Worked example

Acme Foundation Co. (illustrative example, not a real company) is reserving 512x H100 SXM on a 2-year CoreWeave contract. At a negotiated reservation rate (specialist-cloud reserved-cluster rates start at $3.29 per GPU-hour, enterprise discounts apply) the cluster transacts as a multi-tens-of-millions per year line item before storage, egress, and managed-service uplift. Hyperscaler EA equivalents typically come in higher on raw GPU cost; the integration depth is the variable.

Last verified June 2026.