Provider record · updated 2026-09-10
CoreWeave — GPU pricing & deliverability
6 tracked offerings · grade mix A:0 · B:5 · C:0 · D:0 · U:1. CoreWeave prices all GPUs as fixed-size instances (per node), not per GPU; per-GPU figures above are derived by division and noted. HPC interconnect docs (fetched) state two RDMA fabrics — NVIDIA Quantum-2 and Quantum-X800 InfiniBand — in non-blocking fat-tree architecture, plus Spectrum-X RoCE on GB300 instances (800 Gbps options for GB300). B200 release notes give the most explicit per-instance networking (8x ConnectX-7 400G NDR + BlueField-3). GB200 NVL72 instances: 4 GPUs each, deployed only as full 18-node racks, 400 Gb/s NDR IB, Quantum-2. GB300 NVL72, HGX B300 on-demand, and RTX PRO 6000 (standard) are 'Contact sales'. Reserved/committed pricing is not published as a rate card (marketing claims of up to ~60% off exist in secondary sources only).
Tracked offerings
| GPU | SKU | $/GPU-hr | Terms | Grade | Conf | Fabric (as published) | Src |
|---|---|---|---|---|---|---|---|
| H100-SXM | NVIDIA HGX H100 (gd-8xh100ib-i128) — Spot North America, Europe | $2.46 | spot | B | publ | InfiniBand (same instance as on-demand H100) | ↗vendor |
| A100 | NVIDIA A100 (gd-8xa100-i128) North America, Europe | $2.70 | on-demand | U | publ | not published | ↗vendor |
| H100-SXM | NVIDIA HGX H100 (gd-8xh100ib-i128) North America, Europe | $6.16 | on-demand | B | publ | InfiniBand (instance ID carries 'ib'; fabric generation per HPC interconnect docs is Quantum-2/Quantum-X800 non-blocking fat-tree, but per-node bandwidth for H100 not stated on fetched pages) | ↗vendor |
| H200-SXM | NVIDIA HGX H200 (gd-8xh200ib-i128) North America, Europe | $6.31 | on-demand | B | publ | InfiniBand (instance ID carries 'ib'; CoreWeave HPC interconnect docs state Quantum-2/Quantum-X800 non-blocking fat-tree fabrics, but per-node bandwidth for H200 not stated on fetched pages) | ↗vendor |
| B200 | NVIDIA HGX B200 (8x) North America, Europe | $8.60 | on-demand | B | publ | 8x NVIDIA ConnectX-7 InfiniBand HCAs, 400G NDR non-blocking NVIDIA Quantum-2 fabric, plus BlueField-3 DPU (per B200 release notes) | ↗vendor |
| GB200-NVL72 | NVIDIA GB200 NVL72 (gb200-4x) North America, Europe | $10.50 | on-demand · min: docs state NVL72 must be deployed as full racks of 18 nodes | B | publ | 400 Gb/s NDR InfiniBand, NVIDIA Quantum-2 fabric | ↗vendor |
Grades measure what sellers publish — deliverability of the written claim, not measured performance. No fleet holds a measured (Verified) badge yet; that tier arrives with the verification harness. Source classes: API = provider's live pricing API · vendor = provider's published page · mkt API = marketplace API · 3rd-party = aggregator (provider publishes no price on its own pages) · quote = quote-gated or "starting at", not a walk-up price. Quote-class prices are shown in tables but excluded from headline ranges, spreads, and fair-price bands.
Terms & trust (as published, 2026-09-14)
| Uptime SLA | >=99.9% monthly uptime for instances deployed across multiple regions; >=99% for single-region/single-instance deployments (Terms of Service). Separate AI Object Storage SLA: >=99.9% monthly uptime (cwobject.com endpoint only, not cwlota.com). — measured on: Instance level: downtime = loss of external connectivity or persistent disk access for the Single Instance (single region) or all applicable running instances (multi-region), measured in minutes per month (partial minutes under one minute excluded). Storage SLA measured via error percentage (InternalError/ServiceUnavailable) over five-minute intervals. |
| SLA remedy | credits — tiered financial credits on the monthly bill. Compute multi-region: 10% (99-99.9%), 25% (95-98.999%), 50% (<94.999%); single region: 10% (95-99%), 25% (90-94.999%), 50% (<89.999%); capped at 50% of the month's amount due. Storage: 10%/25%/100% of affected zone's monthly bill. Credits apply to future use only (no refunds), must be claimed within 30 days, stated as sole and exclusive remedy. |
| Capacity guarantee | Yes, published in capacity-plans docs: 'Reserved Instances (RI) provide a guaranteed capacity reservation. During your term commitment, the reserved instances are available in your chosen region regardless of platform-wide demand.' Flex Reservations 'provide guaranteed peak capacity while reducing the cost of idle capacity.' On-Demand explicitly has 'no capacity guarantees' and 'may not be available' during peak demand. |
| Egress fees | Free — pricing page lists 'Data transfer between CoreWeave and internet', 'Data transfer within CoreWeave', and 'Egress and IOPS' all as Free. Zero Egress Migration (0EM) program covers egress costs from AWS/Azure/GCP when migrating in; AI Object Storage has no egress fees. |
| Billing granularity | Not explicitly published as per-second/minute/hour. Rates displayed per-hour; capacity-plans docs describe hourly charge calculation with 30-second usage-attribution intervals; ToS states monthly billing in arrears (invoices <$20,000 auto-charged to credit card, >=$20,000 via ACH/wire Net 15). |
| Self-serve signup | no |
| Security certifications | Trust page (coreweave.com/trust) displays SOC 2, ISO 27001, ISO 27017, ISO 27018 badges; other pages state SOC 2 Type II for Bare Metal and CoreWeave Kubernetes Service. GDPR/CCPA described only as 'designed to support customer compliance', not certified. No HIPAA claim found. Detailed reports under NDA via trust.coreweave.com (Conveyor). |
Notes: Sales-led onboarding: docs state 'When CoreWeave approves your organization, the CoreWeave Sales team sends an invitation email' — no public self-serve signup, though credit-card billing exists post-approval. SLA excludes pre-GA features, customer-caused issues, third-party problems, quota downtime, and pre-announced maintenance; maintenance policy also permits unplanned maintenance without notice. Reserved/committed pricing (up to 60% off on-demand) requires contacting sales.
Sources: ↗ ↗ ↗ ↗ ↗ ↗ ↗ ↗ · Fields marked "not published" are exactly that — the seller does not state them anywhere we could find. Absence is a finding.
Reliability record
CoreWeave publishes no machine-readable incident history (status page: https://status.coreweave.com). We track every provider's public status feed weekly; a seller you cannot audit for downtime is a seller whose SLA you cannot price. Absence is a finding.
What CoreWeave does not publish
- max_contiguous_gpus (any SKU)
- per-node InfiniBand bandwidth for H100/H200/A100 instances
- intra-node NVLink/NVSwitch topology per instance
- reserved/committed rate card
- NVL72 NVLink domain size (72 GPUs not stated on fetched pages)
Unpublished fields are recorded as unknown and grade conservatively. Sellers improve their grades by disclosing — that is the point.
Weekly context on every provider
The Delivered Compute Report tracks how these prices and grades move.
Subscribe free