B200 GPU price per hour: 6 clouds ranked, Sep 2026
Ranked by the lowest listed on-demand B200 price per GPU-hour, checked 28 September 2026, Modal leads at $6.25 (GPU only; CPU and memory billed separately), ahead of Lambda at $6.69 and RunPod at $6.79.
Modal’s own pricing page gives the lowest on-demand B200 price of the six GPU clouds we ranked: $6.25 per GPU-hour, converted from $0.001736 per second, checked 28 September 2026.1 CoreWeave is the dearest at $8.60, the per-GPU share of its $68.80 8-GPU node.2 That’s a spread of $2.35 an hour for the same chip, and the host, billing unit and minimum size behind each price explain most of it.
TL;DR
- Ranked by the lowest listed on-demand B200 price per GPU-hour, checked 28 September 2026, Modal ($6.25), Lambda ($6.69) and RunPod ($6.79) sit within 54 cents of each other.1,3,4
- Modal’s rate buys the GPU alone. CPU and memory are billed on top, so its lead narrows or disappears once a host is attached.1
- Crusoe and DigitalOcean list no on-demand B200 price. Crusoe sends B200 instance buyers to sales; DigitalOcean’s GPU Droplet list has no B200 at all.5,6
- A B200 hour costs 1.40 to 2.05 times an H100 hour at the same provider. It carries 180 GB of memory against 80 GB, so it’s cheaper per gigabyte of GPU memory everywhere we could compare.3,4,2
- Interruptible capacity roughly halves the price: Together AI lists preemptible B200s at $4.09 and CoreWeave’s spot node works out to $4.26 per GPU.7,2
- Nebius lists $8.50 from 1 October 2026, up from $7.15, which drops it from fourth to fifth.8
- The hourly rate is only one input to what a model costs to run. Our explainer on what controls AI inference cost covers the rest of the bill.
How this ranking was built
Glamdring Research checked eight GPU clouds that rent capacity by the hour: Modal, Lambda, RunPod, Nebius, Together AI, CoreWeave, Crusoe and DigitalOcean. We captured each company’s own pricing page on 28 September 2026, plus the Vast.ai pricing page as a marketplace reference.
- Source: each provider’s public pricing page, captured 28 September 2026.
- Field: the lowest listed on-demand B200 price per GPU-hour at each provider. HGX B200 and B200 SXM6 listings both count.
- Unit: US dollars per GPU per hour, before tax. Lambda lists its prices “plus applicable sales tax/VAT/GST”.3 Nebius shows all prices “without any applicable taxes, including VAT”.8
- Conversions: CoreWeave prices a whole 8-GPU node, so its node price is divided by eight. Modal prices per second, so its rate is multiplied by 3,600. No other figure is changed.
- Row count: nine pages captured, eight providers checked, six listing an on-demand B200 price, six ranked.
- Not ranked: Crusoe and DigitalOcean, which list no on-demand B200 price; Vast.ai, whose prices are set by hosts on a marketplace; spot, preemptible and reserved rates, which have their own section below.
These are company statements about company prices, and our research methodology treats them that way. The ranking sits in our model infrastructure coverage.
Which cloud rents the B200 cheapest on demand?
Modal does, at $6.25 per GPU-hour, checked 28 September 2026.1 Lambda is second at $6.69 on its 8-GPU B200 instance, and RunPod third at $6.79.3,4 CoreWeave is last at $8.60.2
| Rank | Provider | B200 listing | On-demand (US$ per GPU-hour) | How the page lists it |
|---|---|---|---|---|
| 1 | Modal | Nvidia B200 | 6.25 | $0.001736 per second, × 3,600 |
| 2 | Lambda | B200 SXM6 180 GB, 8-GPU instance | 6.69 | $6.69 per GPU-hour, plus tax |
| 3 | RunPod | B200 180 GB pod | 6.79 | $6.79/hr |
| 4 | Nebius | HGX B200 | 7.15 | $7.15 per GPU-hour; $8.50 from 1 October 2026 |
| 5 | Together AI | HGX B200, GPU Clusters | 8.19 | $8.19 per GPU-hour |
| 6 | CoreWeave | HGX B200, 8-GPU node | 8.60 | $68.80 per node-hour, ÷ 8 |
Each row comes from that provider’s own pricing page, checked 28 September 2026.1,3,4,8,7,2
The top three are close. Modal, Lambda and RunPod sit within 54 cents, and the median of all six is $6.97. The break comes after Nebius: Together AI and CoreWeave are both more than a dollar above the median.
Over a 30-day month of 720 hours, one B200 costs about $4,500 at Modal’s rate and $6,192 at CoreWeave’s. That’s the GPU line alone, before storage, network and tax.
Lambda’s price depends on instance size. The same B200 SXM6 lists at $6.69 per GPU on the 8-GPU instance, $6.79 on four GPUs, $6.89 on two and $6.99 on one.3 A buyer who needs one GPU pays $6.99 at Lambda, which moves it behind RunPod’s single-pod $6.79.4
What does the hourly B200 price include?
It includes a different host at each provider, and at Modal it includes no host at all. Modal lists the B200 GPU at $0.001736 per second and bills CPU at $0.0000131 per physical core per second and memory at $0.00000222 per GiB per second, separately.1
The other five price a machine. RunPod’s B200 pod comes with 28 vCPUs and 283 GB of RAM.4 Lambda’s 1-GPU B200 instance carries 26 vCPUs, 360 GiB of RAM and a 2.75 TiB SSD; the 8-GPU instance carries 208 vCPUs, 2,900 GiB and 22 TiB.3 CoreWeave’s node price covers eight B200s, 128 vCPUs, 2,048 GB of RAM and 61.44 TB of local storage.2
As an illustration, give a Modal B200 the same host as Lambda’s 1-GPU instance: 13 physical cores, the 2-vCPU equivalent of Lambda’s 26 vCPUs, and 360 GiB of memory. The hour then comes to about $9.74, against Lambda’s $6.99.1,3 Modal’s pricing assumes you won’t provision that much. It says customers “never pay for idle resources” and pay for “actual compute time”.1 For a short or bursty job with a lean host, its $6.25 holds. For a long job on a heavy host, a bundled machine is cheaper.
Minimum size matters as much as the rate. CoreWeave’s on-demand B200 comes as an 8-GPU node at $68.80 an hour.2 Its table also lists a single-GPU B200 price of $8.60, but a footnote limits GPU-based pricing to customers of the CoreWeave inference platform.2 Together AI’s $8.19 is its GPU Clusters rate, and it separately lists Dedicated Inference on B200 at $8.99 per GPU-hour.7
Which providers only quote B200 prices through sales?
Crusoe is the clearest case among the eight. Its GPU instance table lists the 180 GB HGX B200 with “Contact sales” in both the on-demand and spot columns, while it prints $3.90 per GPU-hour for the H100.5 The only B200 figure Crusoe publishes is $9.65 an hour for a self-serve managed inference deployment, which is a dedicated model endpoint, not an instance rental.5
DigitalOcean doesn’t list the B200 at all. Its GPU Droplet page, repriced from 1 August 2026, lists B300, H200, H100, L40S, RTX Ada and AMD Instinct GPUs, with no B200 in the on-demand, spot or reserved plans.6
Several ranked providers also move B200 capacity to sales beyond a certain size or term:
- RunPod lists B200 multi-node Clusters as “Contact sales”.4
- Lambda’s 1-Click Clusters of 16 or more B200s for a year or longer carry no printed price and point to its team.3
- Together AI prints reserved B200 rates up to 180 days, then asks buyers to contact it for 181 days or more.7
- Modal offers volume-based discounts on its custom-priced Enterprise plan.1
The pattern is consistent. Short, small B200 rentals have public prices. Long terms and big clusters are negotiated.
How much more does a B200 cost than an H100?
At the same provider, a B200 hour costs between 1.40 and 2.05 times an H100 hour, checked 28 September 2026. The gap in dollars runs from $2.30 at Modal to $4.20 at Together AI. Each pair below compares the provider’s B200 with its own SXM-class or HGX H100 on the same terms.
| Provider | B200 (US$ per GPU-hour) | H100 SXM or HGX (US$ per GPU-hour) | Gap (US$) | B200 ÷ H100 |
|---|---|---|---|---|
| Modal | 6.25 | 3.95 | 2.30 | 1.58 |
| Lambda, 8-GPU instance | 6.69 | 3.99 | 2.70 | 1.68 |
| RunPod | 6.79 | 3.49 | 3.30 | 1.95 |
| Nebius | 7.15 | 3.85 | 3.30 | 1.86 |
| Together AI, GPU Clusters | 8.19 | 3.99 | 4.20 | 2.05 |
| CoreWeave, 8-GPU node | 8.60 | 6.16 | 2.45 | 1.40 |
Sources: each provider’s pricing page, checked 28 September 2026.1,3,4,8,7,2 Modal’s H100 SXM5 converts from $0.001097 per second, and CoreWeave’s HGX H100 from a $49.24 8-GPU node.1,2 The median ratio across the six is 1.77.
The premium buys memory first. Lambda, RunPod and CoreWeave list the B200 at 180 GB per GPU and the H100 at 80 GB, which is 2.25 times the memory.3,4,2 Every ratio in the table is below 2.25, so per gigabyte of GPU memory the B200 is the cheaper chip:
| Provider | B200, US cents per GB-hour | H100, US cents per GB-hour |
|---|---|---|
| Lambda | 3.72 | 4.99 |
| RunPod | 3.77 | 4.36 |
| CoreWeave | 4.78 | 7.69 |
The decision changes with the model. A job that fits in 80 GB pays the full premium for memory it doesn’t use, so the H100 is cheaper per hour and likely per task. A model that needs two H100s just to hold its weights, 160 GB between them, can fit on one 180 GB B200, and then the per-gigabyte figures above are the relevant ones. Memory isn’t throughput, though. Confirm the per-task cost on one representative job before committing.
What do spot and reserved B200 prices cost?
Interruptible B200 capacity costs about half the on-demand rate where it’s listed. Commitment discounts are smaller, and one provider charges more for clusters than for instances. Checked 28 September 2026:
| Provider | Spot or preemptible (US$ per GPU-hour) | Reserved or committed (US$ per GPU-hour) | On-demand (US$ per GPU-hour) |
|---|---|---|---|
| Nebius | from 0.99 | Up to 35% below on-demand for multi-month clusters | 7.15 |
| Together AI | 4.09 | 7.99 (7–30 days), 7.79 (31–90), 6.79 (91–180); 181+ days on request | 8.19 |
| CoreWeave | 4.26 in North America ($34.11 node ÷ 8); 4.36 in Europe ($34.87 ÷ 8) | Not listed | 8.60 |
| Lambda | Not listed | 1-Click Clusters, 2 weeks to 1 year: 9.86 (16 GPUs), 9.36 (64), 8.87 (256+) | 6.69 |
| RunPod | Not listed for B200 | Contact sales | 6.79 |
| Modal | Not listed | Volume-based discounts on Enterprise | 6.25 |
Sources: each provider’s pricing page, checked 28 September 2026.8,7,2,3,4,1
Nebius’s “from $0.99” is a floor on a moving price. It says spot prices are dynamic, can change as often as every 15 minutes, and top out one cent below the current on-demand rate.8 Together AI’s $4.09 preemptible B200 is almost exactly half its $8.19 on-demand rate, and CoreWeave’s spot node is about half its on-demand node.7,2
Commitment saves less. Together AI’s 91–180 day rate of $6.79 is 17% below its on-demand price.7 Nebius’s “up to 35%” would put its B200 near $4.65 at today’s rate, but the page gives no B200-specific figure.8 Lambda runs the other way: its cheapest 1-Click Cluster rate, $8.87 for 256 or more B200s, sits $2.18 above its $6.69 8-GPU instance, and the page doesn’t say what the premium covers.3
How does the Vast.ai marketplace price compare?
Vast.ai’s median B200 listing was $10.63 an hour on 28 September 2026, above every fixed on-demand price in the ranking.9 Its lowest B200 offer at that capture was $9.38.9
The marketplace moves fast. Vast.ai says its prices are “set by supply and demand across 40+ data centers”.9 Its B200 detail page, captured about 90 minutes later the same day, headlined B200 rental “for $6.25/hr”, said prices “update hourly” and rated availability “Low”.10 Both pages label the card at 192 GB, where the ranked clouds list 180 GB.9,10 A median of host-set offers is a different figure from a list price, so Vast.ai sits beside the ranking rather than in it.
Why do other published B200 prices disagree?
Most of the disagreement comes from which providers are counted, the date and the unit. Thunder Compute’s guide, last reviewed 18 September 2026, puts the B200 at $5.99 to $16.11 per GPU-hour.11 Its low end is Hyperbolic at $5.99 and Hyperstack at $6.00, neither of which we checked.11 Its high end is the hyperscalers: Oracle Cloud at $14.00, AWS at $14.24 and Google Cloud at $16.11, all normalised from 8-GPU instances.11
Where the guides and this ranking cover the same provider, the figures mostly agree. Thunder Compute’s prices for Modal ($6.25), Lambda ($6.69), RunPod ($6.79), Nebius ($7.15) and CoreWeave ($8.60) match our 28 September capture.11 Its Vast.ai figure doesn’t: $6.82, described as the median of two verified US or Canadian hosts, against the platform-wide median of $10.63 we captured.11,9
Older figures drift for another reason. Northflank’s guide, published 6 August 2025, lists RunPod’s B200 at $8.64 an hour.12 That matches RunPod’s current Serverless B200 rate, not the $6.79 pod rate.4 The same guide lists Northflank’s own bundled B200 at $5.87 and Google Cloud at $18.53.12 A B200 price is only comparable once the product line, the date and the per-GPU unit are pinned down.
What changes in October 2026?
Nebius has published its next B200 rate: $8.50 per GPU-hour from 1 October 2026, up $1.35, or 19%, from $7.15.8 At $8.50 it moves from fourth to fifth, behind Together AI’s $8.19 and just ahead of CoreWeave’s $8.60. Its H100 rises at the same time, from $3.85 to $4.50, so its B200-to-H100 ratio barely moves, from 1.86 to 1.89.8
Other pages were repriced recently. RunPod dates its pricing page 27 September 2026, and DigitalOcean’s new GPU pricing took effect on 1 August 2026.4,6 Together AI’s page carries a banner saying on-demand B200s are now available on its GPU Clusters.7 This is Glamdring’s first B200 price ranking, so there’s no earlier release to compare.
What the ranking cannot tell you
The ranking covers eight selected clouds, not the whole market. Hyperbolic’s $5.99 and Hyperstack’s $6.00, as reported by Thunder Compute on 18 September, would both rank ahead of Modal if they still held.11 We didn’t check their pricing pages.
It doesn’t measure availability. A listed price says what a provider charges, not whether a B200 can be launched in your region today, and Vast.ai rated its own B200 availability low on the day we checked.10
It doesn’t measure cost per task. The host bundled with each GPU differs, Modal bills its host separately, and storage, network and idle time add to the GPU line. Run one representative job on the two or three cheapest options that fit it, and compare the cost of the finished job. Confirm the instance size, region and tax in the current quote.
One ranked figure carries a label question. RunPod’s page offers Community Cloud and Secure Cloud views, and the page text we captured doesn’t say which view carries the $6.79.4
More compute price rankings are in our published research, and new ones go out in our free email briefing.
Frequently asked questions
Why is the Nvidia B200 so expensive?
Thunder Compute, a GPU cloud that doesn’t rent B200s itself, gives three reasons: high demand for large-scale training, capacity reserved for large enterprise customers, and the cost of the liquid-cooled, high-power systems the chip requires.11 It lists the B200 at 1,000 W and says liquid cooling is required.11 In rental terms, the premium over an H100 at the same provider ran from 1.40 to 2.05 times on 28 September 2026. The main difference those pricing pages list is memory: 180 GB per B200 against 80 GB per H100.3,2
What does it cost to buy a B200 instead of renting one?
Nvidia doesn’t publish a list price, so purchase figures come from resellers and estimates. Thunder Compute estimates $45,000 to $55,000 for a single B200 at street prices and $450,000 to $550,000 for an 8-GPU HGX B200 system.11 A European reseller lists the 8-GPU DGX B200 at €455,000 to €628,000, for businesses in Europe only.13 At Modal’s $6.25 an hour, $45,000 to $55,000 buys roughly 7,200 to 8,800 GPU-hours of rental, or 10 to 12 months of continuous use, before counting the server, power and cooling an owned card needs.
Is the B200 better than the H100?
It has more than twice the memory, 180 GB against 80 GB on the pages we checked, and costs 1.40 to 2.05 times as much per hour at the same provider.3,4,2 For a model that needs two H100s just to hold its weights, one B200 can do it and costs less per gigabyte of memory. For a job that fits comfortably on an H100, the older chip is cheaper per hour. Which is cheaper per task depends on the workload, so test one representative job on both.
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.
