
Facts from CoreWeave’s own pricing and docs pages, September 2026 — 31 read from a quoted sentence on the page we archived.
CoreWeave rents GPUs from $1.25 per GPU-hour on demand, bills hourly usage. CoreWeave does not publish an egress rate or a refund policy.
| On-demand | Spot | Reserved / Bulk Credits | Classic (per-GPU) | |
|---|---|---|---|---|
| From | $6.16/GPU-hr (H100, from a $49.24 8× node) | $19.71/hr for an 8× H100 node | Not published | $0.24 (Quadro RTX 4000) |
| Billing | Monthly in arrears on prior month's usage | Monthly in arrears on prior month's usage | Per the order form | Per hour, components priced separately |
| Uptime SLA | 99.9% across regions, 99% single region | — | 99.9% across regions, 99% single region | 99.9% across regions, 99% single region |
| Tenancy | Whole HGX nodes, 8 GPUs; GH200 sold as one | Same nodes, spot capacity | Reserved compute instances under an order form | Single GPUs with vCPU and RAM billed on top |
| Provisioning | Self-serve in General Access regions; Dedicated Access by arrangement | — | — | — |
| Payment | Invoice on receipt, via payment processor | Invoice on receipt | Order form terms | Invoice on receipt |
As of September 2026, the lowest published on-demand rate is $1.25 per GPU-hour for the NVIDIA L40 48GB. CoreWeave publishes an on-demand rate for 8 GPU models. Billing is hourly usage.
CoreWeave states 99.9% across regions, 99% single region. Service credits: percentage of the affected month's bill, credited to future bills; sole remedy.
CoreWeave does not publish an egress rate, so the cost of moving data out is not something a reader can check before signing up.
CoreWeave lists uS and European data centres; General and Dedicated Access regions.
The contracting entity is CoreWeave, Inc..
Reviewed September 2026, from CoreWeave’s own site. 31 facts, each quoted from the page it was read on.