by CoreWeave, Inc.

CoreWeave pricing, plans and limits

GPU cloud for AI training and inference on bare-metal NVIDIA nodes.

Best for
ML infrastructure leads running multi-node training
Not for
Solo developers who need one GPU for a weekend

Should you use CoreWeave?

Built for teams that train or serve models on whole GPU clusters and want Kubernetes or Slurm already working. It replaces a hyperscaler GPU quota with a specialist that sells most capacity as 8-GPU servers. Skip it for single-GPU experiments, and note that no Asia-Pacific zone is listed in its docs yet.

Operational depth: burn-in testing on idle nodes, GPU straggler detection and a fleet lifecycle controller that CoreWeave told SemiAnalysis brings up around 10,000 GPUs a week.

Operational depth and early access to new NVIDIA generations are hard to copy, while heavy debt and a sales-led model limit who can buy and how fast capacity arrives.

Also for

  • Platform engineers who already run Kubernetes or Slurm
  • ML research teams training models on multi-node GPU clusters
Show 3 more
  • AI platform teams running production inference on Kubernetes
  • Quantitative trading and research firms renting dedicated capacity
  • Enterprise teams that standardize on Slurm and want it managed on Kubernetes

Also not for

  • Teams that need a VPS, a domain or a managed database

Main trade-off

Capacity is mostly sold as whole servers through a sales conversation, so small teams and single-GPU users are poorly served.

CoreWeave pricing

Server plans

CoreWeave server plans
PlanGPUCPURAMStoragePrice
Intel Ice Lake 6342 (General Purpose)Cloud · Entry CPU node. Spot per hour in North America: 2.29. · North America · Europe—96 vCPU384 GB19200 GB$3.36/hr
Intel Emerald Rapids 8562Y+ (General Purpose, High Memory)Cloud · High-memory CPU node. Spot per hour in North America: 3.27. · North America · Europe—64 vCPU512 GB7680 GB$5.31/hr
AMD Genoa 9274F (High Performance)Cloud · CPU node for data prep, orchestration and supporting services. Spot per hour in North America: 3.50. · North America · Europe—96 vCPU768 GB7680 GB$6.42/hr
GH200 instanceGPU · Single-GPU inference and smaller jobs. Price is per instance-hour. · North America · Europe1× NVIDIA GH200 Grace Hopper, 96 GB VRAM72 vCPU480 GB7680 GB$6.50/hr
AMD Genoa 9654 (High Core)Cloud · High-core-count CPU node for parallel preprocessing. Spot per hour in North America: 4.23. · North America · Europe—384 vCPU1536 GB61440 GB$7.54/hr
AMD Genoa 9454 (General Purpose)Cloud · General-purpose CPU node. Spot per hour in North America: 3.32. · North America · Europe—192 vCPU1536 GB7680 GB$7.78/hr
AMD Turin 9655P (General Purpose, High Memory)Cloud · High-memory general-purpose CPU node. Spot per hour in North America: 4.62. · North America · Europe—192 vCPU1536 GB7680 GB$8.18/hr
L40 serverGPU · Lower-cost inference and rendering. Price is per 8-GPU server-hour. Spot per hour: 6.27 North America. · North America · Europe8× NVIDIA L40, 48 GB VRAM128 vCPU1024 GB7680 GB$10/hr
L40S serverGPU · Inference, rendering and fine-tuning. Price is per 8-GPU server-hour. Spot per hour: 7.88 North America. · North America · Europe8× NVIDIA L40S, 48 GB VRAM128 vCPU1024 GB7680 GB$18/hr
RTX PRO 6000 Blackwell server (High Memory)GPU · Inference, rendering and fine-tuning. Price is per 8-GPU server-hour. Spot per hour: 11.09 North America, 11.01 Europe. · North America · Europe8× NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB VRAM128 vCPU1024 GB7680 GB$20/hr
A100 serverGPU · Established training and inference workloads. Price is per 8-GPU server-hour. Spot per hour: 9.65 North America, 9.51 Europe. · North America · Europe8× NVIDIA A100, 80 GB VRAM128 vCPU2048 GB7680 GB$21.60/hr
GB200 NVL72 instanceGPU · Large-model training and inference on NVL72 racks. Price is per 4-GPU instance-hour. · North America · Europe4× NVIDIA GB200 (Blackwell), 186 GB VRAM144 vCPU960 GB30720 GB$42/hr
HGX H100 serverGPU · Training and inference. Price is per 8-GPU server-hour. Spot per hour: 19.71 North America, 19.51 Europe. · North America · Europe8× NVIDIA H100, 80 GB VRAM128 vCPU2048 GB61440 GB$49.24/hr
HGX H200 serverGPU · Training and memory-heavy inference. Price is per 8-GPU server-hour. Spot per hour: 20.93 North America, 20.64 Europe. · North America · Europe8× NVIDIA H200, 141 GB VRAM128 vCPU2048 GB61440 GB$50.44/hr
HGX B200 serverGPU · Large-scale training and inference. Price is per 8-GPU server-hour. Spot per hour: 34.11 North America, 34.87 Europe. · North America · Europe8× NVIDIA B200 (Blackwell), 180 GB VRAM128 vCPU2048 GB61440 GB$68.80/hr
GB300 NVL72 instanceGPU · Largest-model training and inference on NVL72 racks. Quote only: contact sales. · North America · Europe4× NVIDIA GB300 (Blackwell Ultra), 279 GB VRAM144 vCPU960 GB61440 GB—
HGX B300 serverGPU · Frontier training. On-demand quote only: contact sales. Spot per hour: 35.84 North America, 36.70 Europe. · North America · Europe8× NVIDIA B300 (Blackwell Ultra), 270 GB VRAM192 vCPU4096 GB61440 GB—
RTX PRO 6000 Blackwell server (Standard Memory)GPU · Inference and rendering. On-demand quote only: contact sales. Spot per hour: 9.56. · North America · Europe8× NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB VRAM128 vCPU512 GB7680 GB—
Minimum commitment
None for on-demand and Spot; reserved capacity is a committed contract
Billing
Pay as you go · Hourly · Committed spend
How the pricing works, in full

Prices are per node-hour, identical in North America and Europe for on-demand GPUs (opened 2026-10-03). 8-GPU servers: HGX H100 $49.24, HGX H200 $50.44, HGX B200 $68.80, A100 $21.60, L40S $18.00, L40 $10.00, RTX PRO 6000 Blackwell (High Memory) $20.00. The GB200 NVL72 instance (4 GPUs) is $42.00 and the single-GPU GH200 is $6.50. GB300 NVL72, HGX B300 and RTX PRO 6000 (Standard Memory) are quote-only. CPU nodes run $3.36 to $8.18 per hour. The monthly "from" and "to" figures in the page header are the cheapest CPU node and the HGX B200 server run for a 730-hour month. AI Object Storage is $0.06 (Hot), $0.03 (Warm), $0.015 (Cold) and $0.0125 (Archive, limited availability) per GB-month; Distributed File Storage is $0.070. Spot rates, per-GPU inference pricing for inference-platform customers and reserved discounts sit on top of these. See GPU compute services for other providers and NVIDIA for the chips behind the nodes.

Bandwidth pricing

No charge for ingress, egress or transfer between CoreWeave regions; internet data transfer, VPC and NAT gateway are free. Public IPv4 is $4.00 per IP per month.

Costs that are easy to miss

  • Public IPv4 addresses are billed per address per month.
  • Direct Connect ports are priced per month and start at four figures.
  • SUNK has no fee of its own, but the CPU nodes that run it are billed.
  • GB300, B300 and RTX PRO 6000 Standard Memory nodes need a sales quote before you can budget.

CoreWeave regions and compliance

Available regions

  • US East
  • US Central
  • US West
  • Canada East (Ontario)
  • EU North (Sweden, Norway, Denmark)
  • EU South (Spain)
  • EU West (United Kingdom)

Compliance certifications

  • SOC 2
  • ISO 27001
  • ISO 27017
  • ISO 27018

As stated by the provider. Ask for the current report before you rely on one.

Managed services

  • CoreWeave Kubernetes Service (CKS)
  • SUNK (Slurm on Kubernetes)
  • AI Object Storage
  • Distributed File Storage
  • Direct Connect
  • Ingress Service
  • Mission Control
  • CoreWeave Observe
  • Managed Inference

Access methods

  • Web console
  • API
  • Dashboard

Networking

  • Load balancerSupported
  • Private networksSupported
  • CDNNot supported

Storage services

  • Object storageSupported
  • S3 compatibleSupported

What CoreWeave does

CoreWeave operates GPU cloud infrastructure across 59 availability zones in the United States, Canada and Europe, renting bare-metal NVIDIA nodes from L40 inference cards to rack-scale Blackwell systems. Compute is delivered through managed Kubernetes (CKS) or Slurm on Kubernetes (SUNK), with InfiniBand networking and AI-tuned object storage built around the same fleet.

  • Bare-metal GPU nodes

    Nodes run without a hypervisor and range from single-GPU Grace Hopper machines to rack-scale NVL72 instances and 8-GPU HGX servers across the Hopper and Blackwell generations.

  • DPU-enforced tenant isolation

    A BlueField DPU on every node handles VPC networking, NAT and tenant separation on-chip, so a node can change hands between customers without a hypervisor in the path.

  • Two cluster fabrics

    HPC Interconnect offers NVIDIA Quantum InfiniBand and Spectrum-X RoCE with GPUDirect RDMA for GPU-to-GPU traffic inside a cluster.

  • Slurm on Kubernetes (SUNK)

    Self-Service SUNK lets a team declare a Slurm cluster as a single Kubernetes resource on CKS, with the control plane and login pods run for it.

Show all 6 features
  • AI Object Storage with LOTA

    An S3-compatible object store, generally available since March 2025, paired with a node-local accelerator (LOTA) that caches data on the GPU node.

  • Capacity Finder

    A console tool and, since September 2026, a beta API that rank availability zones by Spot capacity before you create a Spot node pool.

Strengths and limits

Strengths

  • Pricing is legible: the rate card is public for most nodes and carries no data-transfer line items, so a bill is nodes plus storage plus IP addresses.
  • Newest silicon shows up early; Vera Rubin rack-scale instances opened in limited availability on 30 September 2026, though access is by request.
  • Zones come in two kinds, shared General Access and single-tenant Dedicated Access, so very large customers can take space that no other tenant shares.

Limits

  • Entry is not self-serve for the infrastructure: sales approval comes first, and GB300, B300 and RTX PRO 6000 Standard Memory nodes carry no public price.
  • Of the zones listed in the docs, 21 are General Access and the other 38 are reserved for select customers; in Europe only the two Spanish zones are General Access, and no Asia-Pacific zone is listed yet.
  • Counterparty risk: SemiAnalysis noted roughly $35 billion of debt in September 2026 and that the Core Scientific merger fell through, which matters if you sign a multi-year commitment.
  • Managed inference was still maturing when SemiAnalysis tested it in September 2026, so raw compute is the safer purchase today.
  • The catalogue is GPU and AI only. There is no VPS, domain, DNS, CDN or managed database product.

CoreWeave alternatives

ServiceRegionsPricingComplianceChecked
CoreWeave This pageUS · Canada · EuropePay as you go · HourlySOC 2 · ISO 270013 Oct
Crusoe CloudPick Crusoe if list price per GPU-hour matters most; pick CoreWeave if long multi-node training runs depend on cluster reliability.US · Europe · CanadaPay as you go · HourlySOC 219 Oct
BasetenPick Baseten if you want managed model serving with per-minute GPU billing; pick CoreWeave if you want to rent whole nodes and run your own scheduler.US · Europe · Asia PacificFree tier · Pay as you go · HourlySOC 2 · HIPAA · GDPR29 Oct
AtlasCloudPick AtlasCloud if you want pay-as-you-go GPUs with one API to many hosted models; pick CoreWeave if you need cluster-scale capacity under a sales-led contract.Europe · Asia Pacific · USPay as you go · HourlySOC 2 · HIPAA23 Oct

CoreWeave facts

Provider founded
2017
Headquarters
Livingston, New Jersey, USA
Total funding
Public company (Nasdaq: CRWV since March 2025)
Team size
2,189 employees at 31 December 2025
Open source
No

What changed

  • Release Vera Rubin NVL72 instances enter limited availability source
  • New feature Capacity Finder API enters beta source
  • Integration Cross-cloud tools added: Interconnect, SUNK Anywhere, LOTA Cross-Cloud source

Questions about CoreWeave

What are CoreWeave's pricing plans in 2026?

CoreWeave bills by the hour per node, and North America and Europe share the same on-demand list rates. An 8-GPU H100 server is $49.24 an hour, an H200 server is $50.44, and a B200 server is $68.80. Spot rates are lower on most nodes (H100 $19.71 in North America) and reserved commitments cut on-demand rates by up to 60 percent. Data transfer, NAT gateways and VPCs are free, while a public IP is $4.00 a month and a 10G Direct Connect port starts at $1,250 a month.

Is CoreWeave free to use?

No. The GPU cloud has no free tier or trial credit on its pricing page, and most nodes are priced as whole 8-GPU servers. CoreWeave Forge, a separate line of experiment tracking, tracing and evaluation tools, has a free plan, a Pro plan from $60 a month with a 30-day trial, and free access for academic research.

What are the best alternatives to CoreWeave?

Crusoe suits teams that want H100 and H200 capacity at a lower list price with free egress. Baseten suits teams that want managed model serving instead of renting nodes. AtlasCloud suits teams that want pay-as-you-go GPUs plus one API to many hosted models. SemiAnalysis also places Nebius beside CoreWeave in the Platinum tier of ClusterMAX 3.0.

CoreWeave or Crusoe: which should you pick?

Crusoe is cheaper on list price: its H100 is $3.90 per GPU-hour on its pricing page, against about $6.16 on CoreWeave's rate card, so CoreWeave costs roughly 58 percent more. SemiAnalysis rated CoreWeave Platinum in September 2026 and moved Crusoe down to Bronze, which is the argument for paying more when a failed node would ruin a long training run. Crusoe fits cost-led inference or smaller clusters, and CoreWeave fits multi-node frontier training.

What does it take to start using CoreWeave?

Access starts with CoreWeave's sales team: once your organization is approved you receive an invitation email, and the first invited user becomes the administrator. From the Cloud Console you create an API token and kubeconfig, create a CKS cluster and a node pool, then work with kubectl and Helm (or SSH for SUNK). Quotas are visible in the console, and extra capacity is requested through your account executive.

About CoreWeave

CoreWeave, Inc. is a GPU cloud provider founded in New Jersey in 2017 and listed on Nasdaq as CRWV since March 2025. Its annual report shows 2,189 employees at the end of 2025 and revenue of $5.1 billion for that year, up from $1.9 billion in 2024; second-quarter 2026 revenue was $2.58 billion against a revenue backlog of about $104 billion. The customer list explains the scale: CoreWeave announced a multi-year agreement with Anthropic, SemiAnalysis names Meta, OpenAI, Microsoft and NVIDIA among its largest customers, and Mistral AI trains models on it.

Read the full background

What you rent is a whole machine. Nodes are bare metal, from a single-GPU Grace Hopper machine up to rack-scale Blackwell systems and 8-GPU HGX servers, and an NVIDIA BlueField DPU on each node enforces tenant isolation in place of a hypervisor. You reach the hardware through CoreWeave Kubernetes Service (CKS) or through SUNK, a Slurm scheduler that runs on Kubernetes. On independent evidence, SemiAnalysis kept CoreWeave in the Platinum tier of its ClusterMAX 3.0 ranking in September 2026 and described its clusters as the yardstick it uses to grade lower-rated providers.

This is the premium end of the market, not the cheap end. In November 2025 SemiAnalysis wrote that, in examples it knew of, CoreWeave charged roughly ten to fifteen percent more per GPU-hour for managed Slurm or Kubernetes clusters than competitors such as Nebius, Crusoe and Lambda. Teams chasing a lower hourly rate can start with Crusoe, then collect a second quote from AtlasCloud. Anyone who would rather deploy models than rent nodes can look at an inference platform such as Baseten.

Buying is sales-led. After CoreWeave approves an organization it emails an invitation, the newest Blackwell Ultra nodes show Contact sales instead of a price, and reserved contracts are quoted to discount on-demand rates by up to 60 percent. Vera Rubin instances entered limited availability in September 2026 in a single Texas zone. Named customers in the Q2 2026 results include Databricks, Runway and Grammarly, and the same release lists CoreWeave Interconnect, private fiber into other clouds that starts with Google Cloud.

Managed inference is the newest and least settled part. When SemiAnalysis tested it in September 2026, public endpoints did not yet serve Kimi K3 or GLM-5.3, and per-token access ran through the Weights & Biases portal or through OpenRouter.

Sources and guides

More AI Infrastructure Services on Hokai