COREWEAVE · TECH
CoreWeave: enterprise GPU cloud USA + UK + Spain, H100 from USD 4.25/h, NVIDIA premium partner
CoreWeave is a US GPU cloud focused on enterprise workloads. Sites in US East/West, UK and Spain. H100 from USD 4.25/h, H200 from USD 4.50/h, large clusters with InfiniBand.
Researched & fact-checked by: DuneDive LLC · As of: 2026-05
What is CoreWeave?
CoreWeave is a US cloud platform focused on GPU compute for AI and visual-effects workloads, founded in 2017 in Roseland, New Jersey. As of May 2026 CoreWeave is one of the fastest-growing cloud providers on the market – the company went public on Nasdaq in March 2025 and has since been among the largest new public tech IPOs of recent years. CoreWeave operates its own data centres in the USA (several sites on the East and West coasts plus Texas and Illinois), in the United Kingdom (London and Aberdeen) and in Spain (Barcelona). Further EU sites are announced.
The product portfolio is built for large enterprise GPU workloads. Multi-GPU nodes with full InfiniBand topology (NVLink + InfiniBand), up to very large cluster configurations (256+ H100 with shared network). Bare-metal GPU servers as dedicated hardware. CoreWeave Kubernetes Service (CKS) as managed Kubernetes with GPU pod support. Object Storage, Block Storage, networking. CoreWeave AI Object Storage as a specialised storage layer with high throughput for training datasets.
In May 2026 CoreWeave runs one of the largest H100 and H200 inventories outside the hyperscalers. NVIDIA premium partner status means preferred access to Blackwell and H200 hardware. Microsoft, OpenAI, NVIDIA and other major players use CoreWeave as external GPU capacity – the contract with OpenAI over USD 11.9 billion (signed in May 2025) is one of the largest GPU cloud contracts in history.
The pricing model is built for enterprise contracts. List prices exist but are negotiable depending on configuration and contract term. H100-80GB SXM from USD 4.25/h on-demand, H200-141GB from USD 4.50-7/h depending on configuration, B200-180GB from USD 7-10/h. Reserved contracts for multi-year commitments markedly cheaper. Standard hourly prices are higher than Lambda Labs or RunPod, but for multi-node training the better InfiniBand topology and cluster size offset the premium through higher throughput.
Why it matters
Three points make CoreWeave the standard choice for serious AI workloads: largest multi-GPU clusters outside the hyperscalers, EU presence with UK and Spain, and NVIDIA premium partner status for preferred hardware access.
Largest multi-GPU clusters: CoreWeave offers multi-node training clusters up to 256+ H100 with full InfiniBand topology and shared network fabric. For a 70B or 405B parameter model pretraining, cluster size and network bandwidth between GPUs are the decisive performance factor. CoreWeave is here among the few providers (alongside Lambda Labs, AWS, Azure, GCP) that offer this size class.
EU presence: London (UK), Aberdeen (UK) and Barcelona (Spain) are established CoreWeave sites. UK is no longer EU after Brexit but is recognised in the nFADP annex as a country with adequate data protection. Spain is clearly EU. These EU sites make CoreWeave interesting for EU AI research projects that do not get hyperscaler capacity or do not want to pay hyperscaler prices.
NVIDIA premium partner: CoreWeave typically gets H100, H200 and Blackwell GPUs before other providers. Anyone needing first-customer access to new NVIDIA hardware (B200, B300, GB200) is often 6-12 months ahead of RunPod or Lambda Labs at CoreWeave. For research groups and state-of-the-art training workloads that is a differentiator.
AI Object Storage: CoreWeave has a specialised storage layer for training datasets, with throughput optimisation and caching logic for repeated reads across many GPU nodes. With very large datasets (10+ TB) storage throughput is often the bottleneck, not the GPU compute itself.
Regulatory position: CoreWeave Inc. is a US Delaware corporation, subject to US law and the CLOUD Act. A US subpoena can directly compel CoreWeave to surrender data – even for workloads in the UK or Spain. This must be documented in the TIA. For EU AI Act high-risk AI with maximum compliance demand, CoreWeave is acceptable only with additional technical measures (customer-managed encryption keys, separate key management). For open-weight model pretraining without personal data the constellation is pragmatic.
Price-wise CoreWeave sits at the upper end of the GPU cloud market. H100 hours are more expensive than at RunPod or Lambda Labs, but on larger clusters with high InfiniBand demand the premium is compensated by better throughput. For single-GPU or 8-GPU workloads CoreWeave is over-priced; from 32+ GPU multi-node onward the calculation often tilts in favour of CoreWeave against competitors.
How it works
Ordering: through the portal coreweave.com followed by sales engagement, or directly via CLI and API access for existing customers. In May 2026 CoreWeave operates predominantly in the enterprise sales model with account manager, cluster architecture workshop, contract negotiation and defined onboarding process. Self-service is limited; standard workloads need 1-2 weeks of onboarding.
Sample prices May 2026 (list prices): H100-80GB SXM on-demand USD 4.25/h. H100-80GB NVL: USD 4.49/h. H200-141GB SXM: USD 4.50-6.99/h depending on cluster size and contract terms. B200-180GB: USD 6.99-10/h. A100-80GB SXM: USD 2.39/h. Multi-GPU clusters with larger nodes and InfiniBand: tiered prices with volume rebates from 32+ GPUs. Object Storage AI from USD 0.025/GB/month. Block storage from USD 0.05/GB/month. Egress traffic from USD 0.04/GB.
Reserved contracts: 1-year reserved for H100 SXM from USD 3.25/h, 3-year from USD 2.49/h. For enterprise customers with multi-million USD volume, additional individual rebates are negotiable.
Kubernetes (CKS): CoreWeave Kubernetes Service as managed Kubernetes with GPU pod scheduling, Slurm integration for batch workloads, integrated monitoring. Control plane from USD 200/month for standard clusters, worker nodes billed as GPU instances.
Network: each GPU node has InfiniBand connectivity to other nodes in the cluster. Bandwidth between nodes depending on configuration 800 Gbit/s NDR InfiniBand. Multi-cluster networking between sites via dedicated backbone network. Anti-DDoS protection standard.
Storage: CoreWeave AI Object Storage with high throughput for training datasets. Block storage as persistent volume mount. External object-storage integration (AWS S3, Google Cloud Storage) via cross-cloud gateway. Network volumes with POSIX filesystem semantics for multi-pod sharing.
Contract details: enterprise model with individual contract design. DPA per GDPR Art. 28 available for EU customers, CLOUD Act applicability must be documented in the TIA. Minimum contract term typically 1-3 years for reserved, on-demand contracts monthly.
Migration: a typical migration path from Lambda Labs or AWS to CoreWeave takes 2-6 weeks depending on cluster size. CoreWeave offers migration services with its own architecture team and performance tuning.
CoreWeave setup for multi-node training in 5 steps
- 01Submit a sales inquiry via coreweave.com, describe the workload profile and cluster requirement (number of GPUs, model, estimated duration), agree on an architecture workshop.
- 02Contract negotiation: reserved vs. on-demand, cluster size, EU site (UK or Spain), DPA clauses, CLOUD Act mitigation by customer-managed encryption.
- 03Define cluster configuration: H100 SXM or H200 SXM, InfiniBand topology (NDR 800 Gbit/s), storage layer (AI Object Storage for datasets, block storage for checkpoints).
- 04Onboarding phase 1-2 weeks: cluster provisioning, Slurm scheduler setup, data upload via cross-cloud gateway, performance validation with a small test job.
- 05Production: multi-node training with Slurm submission, checkpoint logic every 30 minutes, monitoring in the CoreWeave portal, at the end data export and cluster release to reduce cost.
When to use CoreWeave
CoreWeave is the right choice when (a) very large multi-GPU clusters (32+ H100/H200 with full InfiniBand topology) are needed, (b) preferred access to the newest NVIDIA hardware (B200, B300, GB200) is wanted, or (c) EU AI workloads in UK or Spain (instead of US-only) are required. Concrete cases: AI lab with 70B model pretraining over 100+ GPUs over several weeks. EU research consortium needing an EU data location and the newest Blackwell hardware. Enterprise AI initiative with a multi-million USD compute budget.
For multi-node training with InfiniBand topology, CoreWeave is in May 2026 one of the economically most attractive providers. The combination of cluster size, network topology and NVIDIA hardware access is hard to beat in the market.
For EU AI research needing a UK or Spain location, CoreWeave is one of few GPU cloud providers with genuine EU presence and simultaneous large multi-GPU cluster availability. Scaleway Paris is the alternative with a clear EU position but smaller cluster sizes.
For enterprise customers with long-term GPU compute contracts (3-5 years, multi-million USD volume) CoreWeave is competitive against AWS, Azure and GCP through individual contract negotiation and tailored cluster configuration.
When not to use
For small to medium workloads (single GPU, 8-GPU clusters, short experiments) CoreWeave is over-priced. RunPod, Lambda Labs or Scaleway are price-attractive for these cases. CoreWeave only makes economic sense from 32+ GPU multi-node or multi-million USD contracts.
Anyone needing a self-service portal with credit card and wanting to start a pod in 30 seconds is in the wrong place at CoreWeave. The enterprise sales model with 1-2 weeks of onboarding and a prior architecture workshop does not fit agile small teams. For these cases RunPod or Lambda Labs fits better.
For workloads with strict Swiss data residency CoreWeave is not appropriate – no CH region in May 2026. For EU data, UK (with nFADP annex status) or Spain (clearly EU) is possible, but the US parent and CLOUD Act applicability must be documented in the TIA. For mandates under professional secrecy (Art. 321 SCC) Exoscale CH or self-hosted Hetzner GPU is the clean alternative.
For hobby and pro AI workloads without an enterprise budget CoreWeave is not affordable. At H100 from USD 4.25/h versus USD 2.49 at Lambda Labs, the choice is clear.
General caveat: CoreWeave capacity is in May 2026 often reserved for large contracts (OpenAI, Microsoft, NVIDIA). Standard enterprise customers get capacity but with lead time. Anyone needing to scale up at short notice often has faster availability at RunPod or Vast.ai (with all their trade-offs).
Trade-offs
STRENGTHS
- Largest multi-GPU clusters on the market outside the hyperscalers (256+ H100 with InfiniBand)
- EU presence in UK (London, Aberdeen) and Spain (Barcelona)
- NVIDIA premium partner status for preferred access to H200, B200, B300
- Specialised AI Object Storage with high throughput for large training datasets
WEAKNESSES
- US Delaware corporation under CLOUD Act, TIA documentation mandatory for all workloads
- H100 from USD 4.25/h on-demand – Lambda Labs and RunPod are more attractive per hour
- Enterprise sales model with 1-2 weeks of onboarding, no self-service with credit card
- Economically viable only from 32+ GPU multi-node clusters, over-priced for smaller workloads
FAQ
Does CoreWeave have a CH region?
As of May 2026 no. CoreWeave sites are US East/West, UK (London, Aberdeen) and Spain (Barcelona). Further EU sites are announced but no CH region. For workloads with strict Swiss data residency Exoscale Zurich (A100 GPU in CH) is the natural choice. For EU data residency on multi-GPU training CoreWeave UK or Spain is possible, with a documented TIA due to US parent and CLOUD Act applicability.
What does a 32x H100 cluster cost over three months?
On-demand H100 SXM at USD 4.25/h × 32 × 24h × 90d = USD 293,000. With a 3-month reserved contract and individual volume discount typically USD 200,000-230,000 for 90 days. Compared to Lambda Labs: USD 2.99/h × 32 × 24h × 90d = USD 207,000 on-demand. Lambda Labs is more attractive per hour, CoreWeave scores with larger cluster availability and better InfiniBand topology across the whole cluster.
How does CoreWeave NVIDIA premium partner status work?
CoreWeave is among the largest NVIDIA customers worldwide and holds premium partner status, which means preferred access to new hardware. When NVIDIA ships H200 or B200, CoreWeave clusters are typically available 6-12 months before other GPU cloud providers. For research groups and state-of-the-art training workloads, that is a competitive advantage – anyone able to use the newest hardware for 3-5x better efficiency has a lead at CoreWeave. For standard workloads running on H100 or A100, the status is less relevant.
Is CoreWeave appropriate for EU AI Act high-risk AI?
Conditionally. The EU sites (UK, Spain) enable EU data residency. The ISO 27001 certification is in place. The US parent and CLOUD Act applicability must, however, be addressed in the TIA and in the EU AI Act conformity dossier. For high-risk AI with especially sensitive data OVHcloud (SecNumCloud) or Scaleway Paris (French SA without US parent) is the regulatorily cleaner choice. CoreWeave remains attractive when cluster size and hardware generation justify the premium and the additional compliance work.