CoreWeave Pricing (2026): GPU Prices per Hour and Storage
CoreWeave has no free tier and no self-serve plans: you pay per node per hour, or sign a committed contract. On-demand GPUs are sold as whole nodes. In North America an 8x H100 node costs $49.24 an hour ($6.16 per GPU), 8x H200 $50.44 and 8x B200 $68.80, and a GB200 NVL72 instance (4 GPUs) costs $42. Spot capacity is roughly 40% to 60% cheaper where offered (8x H100 $19.71). Reserved contracts cut up to 60% off on-demand prices.
Storage costs $0.015 to $0.06 per GB a month for object storage, and data transfer to the internet is free. Prices are from CoreWeave's public pricing page in US dollars, checked on 25 September 2026.
Prices checked 25 September 2026 on CoreWeave's official pricing page. Prices are in US dollars unless stated and can change at any time.
CoreWeave plans and prices
On-demand GPU instances
- 8x H200 $50.44, 8x B200 $68.80, 8x A100 $21.60, 8x L40S $18.00 per hour
- 8x RTX PRO 6000 (high memory) $20.00; 1x GH200 $6.50
- Up to 61.44 TB local storage per H100/H200/B200 node
- Same on-demand prices in Europe
- Sold per node, usually 8 GPUs
- B300, GB300 and some configurations through sales only
GB200 / GB300 NVL72
- Each instance is 2 Grace Blackwell Superchips (4 GPUs)
- 186 GB VRAM per GPU on GB200, 279 GB on GB300
- No spot pricing for NVL72 systems
Spot GPU instances
- 8x H200 $20.93, 8x B200 $34.11, 8x B300 $35.84, 8x L40 $6.27
- Europe spot prices differ slightly (8x H100 $19.51)
- Governed by CoreWeave's spot terms of service
- Not available for every GPU type or region
Reserved capacity
- Dedicated capacity for the contract term
- Access to configurations not sold on demand
- Terms and exact discounts through sales
Inference single-GPU pricing
- Pay for single GPUs instead of whole nodes
- GB200 $10.50 per GPU-hour
- Only for CoreWeave inference platform customers, through your account executive
CPU instances
- AMD Genoa and Turin, Intel Emerald Rapids options
- Up to 384 vCPUs and 1,536 GB RAM per instance
- Some configurations are contact-sales or spot only
Storage
- AI Object Storage: Hot $0.06, Warm $0.03, Cold $0.015
- No ingress, egress or IOPS fees
- LOTA caching on every node
- Archive tier ($0.0125) in limited availability through CoreWeave
Networking and managed services
- Free internet egress, NAT gateway, VPC and internal transfer
- Free CKS control plane and SUNK (excluding CPU it uses)
- Direct Connect from $1,250/month (10G dedicated)
- Bring-your-own IP costs $4 one-time per IP
- 100G and 400G Direct Connect cost $12,500 and $50,000 a month
Is there a free plan?
CoreWeave has no free plan, free trial credit or self-serve signup. New organisations are approved by CoreWeave's sales team, which then sends an invitation to the platform. The smallest on-demand GPU purchase on the public price list is a whole node, such as 8x L40 at $10 an hour, or a single GH200 at $6.50 an hour. If you want to experiment cheaply first, RunPod or Modal ($30 of free compute a month) are better starting points.
What you will actually pay
Worked examples using the published prices above.
| Scenario | Cost | How we got there |
|---|---|---|
| One-day fine-tune on one 8x H100 node | $1,181.76 on demand or $473.04 on spot | 24 hours x $49.24 = $1,181.76. Spot: 24 x $19.71 = $473.04, if spot capacity is available and the job checkpoints. |
| Two-week training run on 64 H100s (8 nodes) | $132,357.12 on demand; about $52,942.85 at the maximum 60% commitment discount | 8 nodes x 336 hours x $49.24 = $132,357.12. At 60% off: $132,357.12 x 0.4 = $52,942.85. Actual discounts depend on your contract; 60% is CoreWeave's stated maximum. |
| Always-on 8x H200 inference node for 30 days | $36,316.80/month on demand | 720 hours x $50.44 = $36,316.80. Inference platform customers can instead pay per GPU ($6.31 per H200-hour) for fewer GPUs. |
| Storing a 50 TB training dataset for a month | $3,072 hot, $1,536 warm or $768 cold | 50 TB = 51,200 GB (binary). Hot: 51,200 x $0.06 = $3,072. Warm: x $0.03 = $1,536. Cold: x $0.015 = $768. No egress fees to read it back out. |
CoreWeave pricing compared with rivals
| Tool | Paid plans from | Free plan | Note |
|---|---|---|---|
| CoreWeave | From $10/hour (8x L40); 8x H100 $49.24/hour ($6.16 per GPU) | No; accounts are approved by CoreWeave's sales team | The tool on this page |
| Nebius | H100 $3.85/GPU-hour ($4.50 from 1 Oct 2026) | No | Also Platinum-rated; per-GPU pricing, preemptible from $0.79 and up to 35% off reserved. |
| Lambda | H100 SXM $3.99/GPU-hour (8x); $4.29 (1x) | No | Self-serve instances; 1-Click Clusters $5.54-$6.16 per H100-hour. Plus sales tax or VAT. |
| RunPod | H100 SXM $2.69/hour (Community), $3.49 (Secure) | No | Per-second billing for single GPUs, no egress fees; rated below Bronze for clusters. |
| Together AI | H100 $3.99/GPU-hour (on-demand cluster) | No | Reserved from $3.19, preemptible $1.99; also sells inference and fine-tuning APIs. |
| AWS | $55.04/hour per 8x H100 (p5.48xlarge, US East) | No | About $6.88 per GPU-hour on demand; egress fees apply. |
| Modal | $0/month + usage; H100 about $3.95/hour | Yes, $30 of free compute a month | Serverless GPUs billed per second from Python code; CPU and memory billed on top. |
On-demand CoreWeave is not cheap. At $6.16 per H100-hour it costs more than Nebius, Lambda, Together AI and RunPod, and only slightly less than AWS's $6.88. If you only need a few GPUs for a few days, you will pay less elsewhere.
The value is in reserved capacity and reliability. Big customers buy on contracts with up to 60% off, which can bring H100s well below the neocloud on-demand prices, and they get the most reliable large clusters SemiAnalysis tests. On a multi-week run on hundreds of GPUs, avoiding one lost day of training can be worth more than the hourly gap. Free egress also removes a cost that adds up quickly on AWS. See CoreWeave alternatives for cheaper options.
CoreWeave GPU prices (North America, September 2026)
Per instance per hour. Per-GPU figures are the instance price divided by the GPU count.
| GPU | GPUs per instance | VRAM per GPU | On-demand | Per GPU-hour | Spot |
|---|---|---|---|---|---|
| GB300 NVL72 | 4 | 279 GB | Contact sales | - | N/A |
| GB200 NVL72 | 4 | 186 GB | $42.00 | $10.50 | N/A |
| HGX B300 | 8 | 270 GB | Contact sales | - | $35.84 |
| HGX B200 | 8 | 180 GB | $68.80 | $8.60 | $34.11 |
| HGX H200 | 8 | 141 GB | $50.44 | $6.31 | $20.93 |
| HGX H100 | 8 | 80 GB | $49.24 | $6.16 | $19.71 |
| A100 | 8 | 80 GB | $21.60 | $2.70 | $9.65 |
| RTX PRO 6000 (high memory) | 8 | 96 GB | $20.00 | $2.50 | $11.09 |
| L40S | 8 | 48 GB | $18.00 | $2.25 | $7.88 |
| L40 | 8 | 48 GB | $10.00 | $1.25 | $6.27 |
| GH200 | 1 | 96 GB | $6.50 | $6.50 | N/A |
Europe has the same on-demand prices with slightly different spot prices (for example 8x H100 $19.51). Single-GPU pricing at the per-GPU rates is only offered to CoreWeave inference platform customers.
Storage and networking prices
| Item | Price |
|---|---|
| AI Object Storage, Hot | $0.06/GB/month |
| AI Object Storage, Warm | $0.03/GB/month |
| AI Object Storage, Cold | $0.015/GB/month |
| AI Object Storage, Archive (limited) | $0.0125/GB/month |
| Distributed File Storage | $0.07/GB/month |
| Egress and IOPS | Free |
| Internet data transfer, NAT gateway, VPC | Free |
| Public IP address | $4 per IP a month |
| Direct Connect, 10G | $1,250 (dedicated) or $1,500 (virtual) a month |
| Direct Connect, 100G / 400G | $12,500 / $50,000 a month |
| CKS control plane and SUNK | Free (SUNK's CPU use billed) |
CoreWeave measures storage in binary gigabytes (1 GB = 2^30 bytes).
On-demand vs spot vs reserved
- On-demand: list price, no commitment, available for most GPU types. Best for short bursts.
- Spot: roughly 40% to 60% cheaper where offered (8x H100 $19.71 vs $49.24, 8x A100 $9.65 vs $21.60), but CoreWeave can reclaim the capacity. Use it only with frequent checkpoints.
- Reserved: committed contracts with up to 60% off on-demand, which is how most large customers buy. Ask for the discount at different terms and volumes, and check whether storage and CPU nodes are included.
- Price your job per GPU-hour, not per node: divide the node price by 8 before comparing with per-GPU clouds.
- Run evaluation sweeps and checkpointed training on spot nodes, which cost 60% less than on-demand for H100s.
- Keep hot training data in AI Object Storage and move finished checkpoints to cold storage at a quarter of the price.
- Use CoreWeave's free egress to keep a copy of key checkpoints with another provider, so a contract ending never traps your work.
- Before signing a reserved contract, benchmark one real training step on the exact node type you will commit to.
Jargon explained
- Node
- One physical server in a cluster. On CoreWeave most GPU nodes hold 8 GPUs, and you rent the whole node.
- Spot instance
- Spare capacity sold at a discount that the provider can take back at short notice. Good for jobs that save progress often.
- Reserved capacity
- GPUs you commit to use for a fixed term in exchange for a lower price than on-demand.
- Slurm
- A job scheduler used on supercomputers. Researchers submit training jobs to it and it assigns them to free GPUs.
- Egress
- Data sent out of a cloud to the internet or another provider. CoreWeave does not charge for it; many clouds do.
Frequently asked questions
How much does an H100 cost on CoreWeave?
As of 25 September 2026, an 8x H100 node costs $49.24 an hour on demand in North America, which is $6.16 per GPU. Spot costs $19.71 per node, and reserved contracts can cut up to 60%.
Does CoreWeave have a free tier?
No. There is no free tier or trial credit on the pricing page, and new organisations are approved by sales before getting access.
Can I rent a single GPU on CoreWeave?
Only a GH200, which comes as a single-GPU instance at $6.50 an hour, or single GPUs through the inference platform via your account executive. Other on-demand GPUs are sold as 8-GPU nodes.
Does CoreWeave charge for egress or storage transfer?
No. Internet data transfer, internal transfer, NAT gateway and object-storage egress are free. You pay for storage by the GB and $4 a month per public IP.
How big are CoreWeave's commitment discounts?
CoreWeave says committed usage gets up to 60% off on-demand prices. Exact discounts depend on term, volume and hardware, and are agreed with sales.
Is CoreWeave cheaper than AWS?
On demand, slightly: an 8x H100 node is $49.24 an hour against $55.04 for AWS p5.48xlarge in US East, and CoreWeave charges no egress. Both are well above neoclouds such as Nebius, Lambda and RunPod.
Sources
Every fact on this page comes from public information. Vendor figures are labelled as vendor claims.
- CoreWeave cloud pricing (CoreWeave)
- CoreWeave platform documentation (CoreWeave)
- CoreWeave reports strong second quarter 2026 results (CoreWeave)
- CoreWeave becomes the only provider to earn SemiAnalysis' Platinum ClusterMAX rating three consecutive times (CoreWeave)
- ClusterMAX 3.0: the industry standard GPU cloud rating system (SemiAnalysis)
- CoreWeave completes acquisition of Weights & Biases (CoreWeave)
- CoreWeave reviews (Trustpilot)
- Nebius AI Cloud prices (Nebius)
- Lambda pricing (Lambda)
- Runpod GPU cloud pricing (Runpod)
- Together AI pricing (Together AI)
- Modal pricing (Modal)
- Amazon EC2 on-demand pricing (Amazon Web Services)
- Amazon EC2 P5 instances (Amazon Web Services)