CoreWeave pricing: every GPU node, Sep 2026
On CoreWeave's pricing page, checked 28 September 2026, the HGX B200 is the dearest self-serve GPU node at $68.80 per node-hour ($8.60 per GPU-hour) and the one-GPU GH200 the cheapest at $6.50.
The NVIDIA HGX B200 is CoreWeave’s most expensive GPU node with a published on-demand price: $68.80 per node-hour for eight GPUs, or $8.60 per GPU-hour, on CoreWeave’s pricing page checked 28 September 2026.1 The page lists twelve GPU nodes. Nine carry an on-demand price, and the cheapest is a single GH200 at $6.50 an hour.1 Per GPU, on-demand runs from $1.25 for an L40 to $10.50 for a GB200 NVL72. Spot prices in North America sit 37% to 60% below on-demand where both are listed.1
TL;DR
- You buy CoreWeave GPUs by the node. Nine of the twelve nodes hold eight GPUs, so the per-GPU rate only holds when every GPU in the node is working. CoreWeave sells single-GPU pricing to inference platform customers only.1
- Spot is the largest published discount. An HGX H100 node costs $19.71 an hour on spot against $49.24 on demand. Over a 720-hour month that is $14,191.20 against $35,452.80.1
- Three nodes have no self-serve on-demand price: GB300 NVL72, HGX B300 and RTX PRO 6000 Blackwell Server Edition (Standard Memory). The B300 and the Standard Memory RTX PRO 6000 can still be bought on spot.1
- Committed usage earns “up to 60%” off on-demand, with terms set in a sales conversation. Sixty per cent off the H100 node lands within two cents of its North American spot price.1
- On-demand prices are the same in North America and Europe. Spot prices differ by up to $0.86 per node-hour, and Europe lists no spot price for the L40 or L40S.1
- Egress and data transfer cost nothing. Storage is billed per binary gigabyte per month.1
How this ranking was built
Glamdring Research ranked every GPU node on CoreWeave’s pricing page by one field: On-Demand Price (Per Hour), the hourly price of a whole node, from the North America table.1 The ranking sits in our model infrastructure category. The page was captured on 28 September 2026 at 15:05 Australian Eastern Standard Time. The page shows no last-updated date. The page lists 12 GPU nodes in North America and the same 12 in Europe. Nine North American nodes carry an on-demand price and are ranked. Three show “Contact sales” and sit unranked at the foot of the table.1 Prices appear with a $ sign. The captured page names no currency code and states no tax treatment.1
Per-GPU-hour figures are Glamdring’s arithmetic: node price divided by the page’s GPU Count, rounded half up to the cent. For the GB200 NVL72 and GB300 NVL72, a footnote sets the count at four. Each instance holds two Superchips, and each Superchip has one Grace CPU and two Blackwell GPUs.1 Ten quotients run past the cent. Their exact values are $6.155 (H100 on-demand) and $6.305 (H200 on-demand). On spot they are $4.26375 (B200), $2.61625 (H200), $2.46375 (H100), $1.38625 (RTX PRO 6000 High Memory), $1.20625 (A100), $1.195 (RTX PRO 6000 Standard Memory), $0.985 (L40S) and $0.78375 (L40). Spot discounts are one minus spot divided by on-demand. CPU-only nodes, storage and networking sit on the same page and are left out of the ranking. How Glamdring treats company-published prices is set out in our research methodology.
What does each CoreWeave GPU node cost per hour?
The HGX B200 heads the ranking at $68.80 per node-hour, and the one-GPU GH200 sits last at $6.50. All figures below come from CoreWeave’s North America table, checked 28 September 2026.1
| Rank | GPU node | GPUs per node | On-demand ($ per node-hour) | On-demand ($ per GPU-hour) | Spot, North America ($ per node-hour) | Spot ($ per GPU-hour) |
|---|---|---|---|---|---|---|
| 1 | NVIDIA HGX B200 | 8 | $68.80 | $8.60 | $34.11 | $4.26 |
| 2 | NVIDIA HGX H200 | 8 | $50.44 | $6.31 | $20.93 | $2.62 |
| 3 | NVIDIA HGX H100 | 8 | $49.24 | $6.16 | $19.71 | $2.46 |
| 4 | NVIDIA GB200 NVL72 | 4 | $42.00 | $10.50 | N/A | N/A |
| 5 | NVIDIA A100 | 8 | $21.60 | $2.70 | $9.65 | $1.21 |
| 6 | NVIDIA RTX PRO 6000 Blackwell Server Edition (High Memory) | 8 | $20.00 | $2.50 | $11.09 | $1.39 |
| 7 | NVIDIA L40S | 8 | $18.00 | $2.25 | $7.88 | $0.99 |
| 8 | NVIDIA L40 | 8 | $10.00 | $1.25 | $6.27 | $0.78 |
| 9 | NVIDIA GH200 | 1 | $6.50 | $6.50 | N/A | N/A |
| Unranked | NVIDIA HGX B300 | 8 | Contact sales | n/a | $35.84 | $4.48 |
| Unranked | NVIDIA RTX PRO 6000 Blackwell Server Edition (Standard Memory) | 8 | Contact sales | n/a | $9.56 | $1.20 |
| Unranked | NVIDIA GB300 NVL72 | 4 | Contact sales | n/a | N/A | n/a |
The node ranking and the per-GPU order disagree. The GB200 NVL72 ranks fourth by node price because its node holds four GPUs. Per GPU it is the most expensive listing on the page.1 The GH200 is the cheapest node to switch on and the seventh-cheapest of the nine per GPU.
The GB300 NVL72 is the only GPU node with no self-serve price of either kind. CoreWeave lists no spot price for the GB200 NVL72 or the GH200.1
What does a CoreWeave GPU cost per GPU-hour?
Glamdring Research reads CoreWeave’s per-GPU-hour rate as a comparison unit. On demand it runs from $1.25 (L40) to $10.50 (GB200 NVL72). Outside CoreWeave’s inference platform, the smallest purchase is a whole node.1
In ascending order, the on-demand per-GPU rates are L40 $1.25, L40S $2.25 and RTX PRO 6000 High Memory $2.50. A100 follows at $2.70, then H100 $6.16, H200 $6.31, GH200 $6.50, B200 $8.60 and GB200 NVL72 $10.50.1
CoreWeave prints these figures itself. The North America table carries a column called Inference Single GPU Price (Per Hour). Each of its nine prices matches node price divided by GPU count at the rounding used here. It lists $6.16 for the H100, $6.31 for the H200 and $10.50 for the GB200 NVL72.1 A footnote limits that column: “Pricing based on GPUs is exclusively applicable to the CoreWeave inference platform customers.”1 The Europe table has no such column.
The decision changes with utilisation. A team that fills all eight GPUs pays the per-GPU rate. A team that needs two H100s still pays $49.24 an hour for the node, which is $24.62 per working GPU. The lowest self-serve on-demand entry points are the one-GPU GH200 at $6.50 an hour and the eight-GPU L40 node at $10.00.1 Test the choice against one real workload: the GPU count it keeps busy, its hours per month and the memory it needs per GPU.
How much cheaper is CoreWeave spot pricing?
Spot cuts CoreWeave’s North American node prices by 37.3% to 60.0% on the seven nodes that list both an on-demand and a spot price.1
| GPU node | On-demand ($ per node-hour) | Spot ($ per node-hour) | Spot discount |
|---|---|---|---|
| NVIDIA HGX H100 | $49.24 | $19.71 | 60.0% |
| NVIDIA HGX H200 | $50.44 | $20.93 | 58.5% |
| NVIDIA L40S | $18.00 | $7.88 | 56.2% |
| NVIDIA A100 | $21.60 | $9.65 | 55.3% |
| NVIDIA HGX B200 | $68.80 | $34.11 | 50.4% |
| NVIDIA RTX PRO 6000 Blackwell Server Edition (High Memory) | $20.00 | $11.09 | 44.6% |
| NVIDIA L40 | $10.00 | $6.27 | 37.3% |
Hopper nodes carry the deepest spot discounts. The Blackwell B200 node sits near half price, and the L40 has the shallowest discount of any GPU node.1 In cash terms, an HGX H100 node running for a 720-hour month costs $35,452.80 on demand and $14,191.20 on spot. The gap is $21,261.60. The estimate assumes continuous use and no lost spot hours.
The cheapest GPU node on the page at any price is the North American L40 on spot, at $6.27 an hour.1 The page links to CoreWeave’s Spot Terms of Service, which set the conditions on spot capacity. This ranking covers the published prices only.1
What hardware comes with each node?
Every CoreWeave GPU node is a fixed bundle. The page lists GPU count, VRAM, vCPUs, system RAM and local storage beside a single hourly price.1
| GPU node | GPUs | VRAM (GB, as listed) | vCPUs | System RAM (GB) | Local storage (TB) |
|---|---|---|---|---|---|
| NVIDIA GB300 NVL72 | 4 | 279 | 144 | 960 | 61.44 |
| NVIDIA GB200 NVL72 | 4 | 186 | 144 | 960 | 30.72 |
| NVIDIA HGX B300 | 8 | 270 | 192 | 4,096 | 61.44 |
| NVIDIA HGX B200 | 8 | 180 | 128 | 2,048 | 61.44 |
| NVIDIA RTX PRO 6000 Blackwell Server Edition (High Memory) | 8 | 96 | 128 | 1,024 | 7.68 |
| NVIDIA RTX PRO 6000 Blackwell Server Edition (Standard Memory) | 8 | 96 | 128 | 512 | 7.68 |
| NVIDIA GH200 | 1 | 96 | 72 | 480 | 7.68 |
| NVIDIA HGX H100 | 8 | 80 | 128 | 2,048 | 61.44 |
| NVIDIA L40 | 8 | 48 | 128 | 1,024 | 7.68 |
| NVIDIA HGX H200 | 8 | 141 | 128 | 2,048 | 61.44 |
| NVIDIA L40S | 8 | 48 | 128 | 1,024 | 7.68 |
| NVIDIA A100 | 8 | 80 | 128 | 2,048 | 7.68 |
The two RTX PRO 6000 rows differ only in system RAM: 1,024 GB against 512 GB.1 The A100 node carries 7.68 TB of local storage. The H100 and H200 nodes each carry 61.44 TB. The HGX B300 has the most system RAM on the page at 4,096 GB.1 CoreWeave counts a terabyte as 2⁴⁰ bytes.1
Do CoreWeave prices differ between North America and Europe?
On-demand prices are identical in CoreWeave’s North America and Europe tables. Spot prices differ on most nodes, and Europe lists fewer of them.1
| GPU node | On-demand, both regions ($ per node-hour) | Spot, North America | Spot, Europe | Europe spot ($ per GPU-hour) |
|---|---|---|---|---|
| NVIDIA GB300 NVL72 | Contact sales | N/A | N/A | n/a |
| NVIDIA GB200 NVL72 | $42.00 | N/A | N/A | n/a |
| NVIDIA HGX B300 | Contact sales | $35.84 | $36.70 | $4.59 |
| NVIDIA HGX B200 | $68.80 | $34.11 | $34.87 | $4.36 |
| NVIDIA RTX PRO 6000 Blackwell Server Edition (High Memory) | $20.00 | $11.09 | $11.01 | $1.38 |
| NVIDIA RTX PRO 6000 Blackwell Server Edition (Standard Memory) | Contact sales | $9.56 | $9.56 | $1.20 |
| NVIDIA GH200 | $6.50 | N/A | N/A | n/a |
| NVIDIA HGX H100 | $49.24 | $19.71 | $19.51 | $2.44 |
| NVIDIA L40 | $10.00 | $6.27 | N/A | n/a |
| NVIDIA HGX H200 | $50.44 | $20.93 | $20.64 | $2.58 |
| NVIDIA L40S | $18.00 | $7.88 | N/A | n/a |
| NVIDIA A100 | $21.60 | $9.65 | $9.51 | $1.19 |
Blackwell spot costs more in Europe. The B300 is $0.86 higher per node-hour and the B200 is $0.76 higher. Hopper and Ampere spot costs less: the H200 by $0.29, the H100 by $0.20 and the A100 by $0.14.1 Europe’s H100 spot price of $19.51 is 60.4% below on-demand, the deepest spot discount on any GPU node on the page.1 Europe lists no spot price for the L40 or the L40S.
Why do other published CoreWeave prices disagree?
The third-party CoreWeave figures we checked differ from CoreWeave’s page in small, traceable ways. The causes are rounding, one figure that does not match the page and a separate classic pricing page.
Thunder Compute’s guide, dated 1 September 2026, lists “8x H100” at $49.28 an hour.2 CoreWeave’s page shows $49.24.1 The same guide describes total instance cost as a GPU charge plus vCPUs and RAM.2 The current page prints each GPU node’s vCPUs, RAM and storage in the same row as one hourly price. The page shows no separate CPU or RAM charge for GPU nodes.1
ComputePrices lists the CoreWeave H200 at $6.30 per GPU-hour and the L40 with 40 GB of memory.3 The H200 node price of $50.44 divides to $6.305, which CoreWeave’s own inference column shows as $6.31. CoreWeave lists the L40 at 48 GB of VRAM.1 ComputePrices’ “from $0.78” headline matches the North American L40 spot price of $6.27 divided by eight.3
Spheron read CoreWeave’s rate card on 12 August 2026.4 Every on-demand and North American spot figure in its table matches the 28 September capture. Spheron shows one RTX PRO 6000 row, carrying the High Memory prices.4 CoreWeave’s page lists High Memory and Standard Memory as separate nodes.1 Spheron also reports that CoreWeave’s older “Classic” pricing page lists lower H100 rates.4 The current page links to that classic page.1 Glamdring did not capture it, so its figures are left out here.
What does CoreWeave charge beyond the GPU node?
Storage and public IP addresses carry charges. Data movement does not: CoreWeave lists egress, IOPS and data transfer as free.1
| Item | Listed price |
|---|---|
| AI Object Storage, Hot | $0.06 per GB per month |
| AI Object Storage, Warm | $0.03 per GB per month |
| AI Object Storage, Cold | $0.015 per GB per month |
| AI Object Storage, Archive (limited availability) | $0.0125 per GB per month |
| Distributed File Storage | $0.070 per GB per month |
| Egress and IOPS | Free |
| Data transfer to the internet and within CoreWeave | Free |
| NAT Gateway and VPC | Free |
| Public IP address | $4.00 per IP per month |
| Bring Your Own IP | $4.00 one-time setup fee per IP |
| Direct Connect, dedicated 10G / 100G / 400G | $1,250 / $12,500 / $50,000 per month |
| Direct Connect, virtual 10G | $1,500 per month (100G and 400G not available) |
| SUNK (Slurm on Kubernetes) | Free, excluding the CPU usage needed to run it |
| CoreWeave Kubernetes Service control plane | Free |
CoreWeave bills storage in binary gigabytes of 2³⁰ bytes.1 Holding 10 TB (10,240 GB) in Hot object storage adds $614.40 a month to a node bill. Direct Connect prices exclude cross-connect and third-party port charges.1 CPU-only nodes start at $3.36 an hour on demand for an Intel Ice Lake (6342) node.1
Committed usage is the lever the page does not price. CoreWeave “offers up to 60% discounts over our On-Demand prices for committed usage” and routes buyers to its sales team.1 Applied to the H100 node, 60% off $49.24 is $19.696 an hour. That is $0.014 below the North American spot price, with a commitment attached. The node rate is one input to a model’s serving bill. Our explainer on what controls AI inference cost covers the others.
What the ranking cannot tell you
These are CoreWeave’s listed prices on 28 September 2026. They are a company statement. They do not show what committed customers pay, because the page publishes a ceiling on the discount and no rate card for it.1
The page shows price. Nor does it show available capacity or delivery times. The ranking does not measure performance per dollar. It covers North America and Europe only, the two regions the page lists. Currency and tax treatment are not stated on the captured page. Confirm them in the current quote.
The ranking will be rebuilt when CoreWeave changes its pricing page. New pricing research goes out in our free research briefing.
Frequently asked questions
How much does CoreWeave charge for GPU rental?
CoreWeave’s on-demand GPU nodes cost $6.50 to $68.80 an hour on its pricing page checked 28 September 2026.1 That is $1.25 to $10.50 per GPU-hour once the node price is divided by its GPU count. Spot nodes in North America start at $6.27 an hour for an eight-GPU L40 node. Three nodes need a sales call for on-demand pricing.1
Can you rent a single GPU on CoreWeave?
The NVIDIA GH200 is the only one-GPU node on CoreWeave’s pricing page, at $6.50 an hour on demand.1 CoreWeave also lists per-GPU prices. A footnote says those apply exclusively to CoreWeave inference platform customers. Every other listed node holds four or eight GPUs and is priced as a whole.1
How much does CoreWeave discount committed usage?
CoreWeave advertises discounts of up to 60% off on-demand prices for committed usage.1 It does not publish which commitment length or volume earns which discount. At the full 60%, an HGX H100 node would cost $19.696 an hour. CoreWeave’s North American spot price for the same node is $19.71. Get the committed rate in writing from CoreWeave’s sales team before comparing it with spot.
Sources checked
Company-owned pages establish what a company says. They do not prove a market conclusion. Each source is dated so readers can judge each claim.
