GPU Cloud Radar

CoreWeave pricing: what does an H100 hour cost?

Published 2026-11-11 · Researched and price-checked on official sources

CoreWeave pricing runs $49.24 per hour for an eight-GPU HGX H100, which works out to $6.16 per GPU-hour, with spot capacity at $19.71. Those are the list rates on CoreWeave's own pricing page on November 11, 2026. The harder question for buyers now is credit risk, not sticker price.

Key facts

What matters

An eight-GPU HGX H100 lists at $49.24 per hour on demand, or $6.16 per GPU-hour. The same instance on spot is $19.71 per hour, exactly 60% less.
The 'inference single GPU price' column is not a discount: on all nine rows where both numbers appear, it equals the on-demand instance price divided by GPU count.
Spot instances can be reclaimed with a maximum seven-minute window, and spot capacity is not guaranteed in any region.
CoreWeave's 2032 bond yielded above 16% on November 6, 2026, and Goldman Sachs estimates data-center moratoriums could touch 45% of planned US capacity. That changes the case for prepaying.

CoreWeave pricing at a glance

CoreWeave publishes two GPU rate cards, one for North America and one for Europe. Both bill by the hour for a whole instance, not per card, which is why the headline number looks big until you divide it. The North America card, captured on November 11, 2026, lists these rates.

The arithmetic is the useful part. An eight-GPU HGX H100 at $49.24 per hour is $6.16 per GPU-hour, and a full day at that rate costs $1,181.76. The same instance on spot is $19.71 per hour, or $473.04 for the day. An eight-GPU A100 at $21.60 per hour is $2.70 per GPU-hour, $518.40 for a day.

Three rows carry no public on-demand number. GB300 NVL72, HGX B300 and the standard-memory RTX PRO 6000 Blackwell all read 'Contact sales' in that column, so any per-GPU math for those three would be a guess. Everything else on the card is public.

The per-GPU math behind the rate card

CoreWeave prints a third price column headed 'Inference Single GPU Price (Per Hour)'. It reads like a volume rate for inference work. It is not one. On every row where the on-demand and inference columns both carry a number, the inference figure is the on-demand instance price divided by the GPU count, rounded to the nearest cent.

HGX H100: $49.24 divided by 8 is $6.155, printed as $6.16.
HGX B200: $68.80 divided by 8 is $8.60, printed as $8.60.
NVIDIA L40: $10.00 divided by 8 is $1.25, printed as $1.25.
GH200: a single-GPU instance at $6.50, so the per-GPU price is the instance price.

Footnote 2 on the page explains the scope: per-GPU pricing is available only to customers on the CoreWeave inference platform, and the note says to contact an account executive. Treat that column as a billing unit for inference customers, not a self-serve discount.

Footnote 1 changes how the Blackwell rows should be read. Each GB200 or GB300 instance contains two Superchips, and each Superchip carries one Grace CPU and two Blackwell GPUs. So the GB200 NVL72 line is a four-GPU instance at $42.00 per hour, or $10.50 per GPU-hour. That is about 70% more per GPU-hour than an H100, which is the number that matters if you are budgeting a training run rather than admiring the rack.

Spot is 60% cheaper and comes with a seven-minute timer

Spot is where the discount lives. On the North America card, H100 spot is $19.71 against $49.24 on demand, exactly 60% lower. B200 spot is $34.11 against $68.80, about 50% lower. A100 spot is $9.65 against $21.60, 55% lower. The Europe card tracks closely on demand and differs slightly on spot: H100 there is $19.51, H200 is $20.64, A100 is $9.51.

CoreWeave's documentation is explicit about the trade. Spot instances are preemptible, meaning the company can reclaim them at any time if the capacity is needed elsewhere. When that happens, the documented timeline is short.

T+0:00: the node is cordoned so new jobs stop scheduling, and a Kubernetes warning event fires.
T+0:02: drain begins, a second warning event marks the start, and processes receive SIGTERM.
30 seconds after that: SIGKILL to anything still running. Processes cannot catch or ignore this one.
T+0:07 at the latest: the instance leaves the pool.

Capacity is not guaranteed and varies by availability zone; the docs point customers at Capacity Finder before building a pool. Preemption logs are retained for one week. Availability also differs by product and region: L40 and L40S show no spot price in Europe, and GB200 and GB300 show none in either region.

What the credit market changes about buying GPU capacity

Sticker price is half of a cluster decision. The other half is whether the operator will still be selling you the same capacity on the terms you signed. On November 6, 2026, CNBC reported that the yield on CoreWeave's 2032 bond with a 9.625% coupon had climbed above 16%, and that the company's five-year credit default swaps spiked above 1,289 basis points on the Thursday before easing about 8% to roughly 1,155.

One sentence in that report is aimed straight at customers: for neoclouds, any sign of consolidation or retrenchment in capex could mean fewer or shorter contracts for data center compute, or delays in prepayments. Goldman Sachs also estimated that data-center moratoriums could affect 45% of planned US capacity, concentrated in Ohio, Texas and Georgia, the states holding much of the planned build-out.

For a buyer, the practical reading is narrow and unglamorous. Pay for compute you can checkpoint or restart, because spot and on-demand pricing is what the market is already clearing at. If you need a reserved contract, keep the term short enough that a renegotiation is survivable, and get the exit and prepayment language in writing before you wire anything. Keep a second provider credentialed, even if you never run production there, so a capacity shortfall is a migration rather than a shutdown. None of that is a forecast about CoreWeave's solvency. It is the standard counterparty check you would run on any supplier that wants a year of cash up front.

Which GPU to rent, by workload

The rate card sorts into three price bands once you convert to per-GPU-hour.

Cheapest per card: L40 at $1.25 per GPU-hour (48 GB), then L40S at $2.25 and A100 80 GB at $2.70. For serving a model that fits in 48 GB, L40S at $2.25 is the value pick on this card, with spot at $7.88 per instance, or $0.99 per GPU-hour.
Mainstream training: H100 at $6.16 per GPU-hour and H200 at $6.31, a 2.4% premium for the extra 61 GB of memory per card. B200 jumps to $8.60.
Blackwell racks: GB200 NVL72 at $10.50 per GPU-hour, with GB300 priced only through sales.

Note what is absent: there is no public reserved or committed-use rate on this page. CoreWeave sells long-term capacity through sales conversations, so the numbers above are the ceiling of what you should pay, not the floor of what you might negotiate.

Who should use CoreWeave, and who should not

CoreWeave fits teams that already run Kubernetes and can absorb a preemption. The spot timeline is workable if your job checkpoints every few minutes and your scheduler tolerates nodes disappearing. It is a bad fit for a single long training run with no checkpointing, or for anything with an uptime promise attached.

Budget shape matters too. An eight-GPU H100 at $49.24 per hour is $35,452.80 for 720 hours of on-demand use, which is the number to compare against a reserved quote. If your utilization sits under about a third of the month, on-demand wins on price even before you price in the commitment risk.

If you are still choosing a provider, our cheapest GPU cloud for LLM inference comparison covers the smaller-card options, and RunPod alternatives covers the serverless end of the market. For per-card rates at other hosts, see GPU dedicated server pricing.

At a glance

Instance	GPUs	On-demand per hour	Spot per hour	Per GPU-hour
NVIDIA HGX H100 (80 GB)	8	$49.24	$19.71	$6.16
NVIDIA HGX H200 (141 GB)	8	$50.44	$20.93	$6.31
NVIDIA HGX B200 (180 GB)	8	$68.80	$34.11	$8.60
NVIDIA GB200 NVL72	4	$42.00	N/A	$10.50
NVIDIA A100 (80 GB)	8	$21.60	$9.65	$2.70
NVIDIA L40S (48 GB)	8	$18.00	$7.88	$2.25
NVIDIA L40 (48 GB)	8	$10.00	$6.27	$1.25
NVIDIA GH200 (96 GB)	1	$6.50	N/A	$6.50

CoreWeave GPU list prices, North America, verified November 11, 2026

FAQ

How much does CoreWeave cost per GPU hour?

On the November 11, 2026 rate card, an H100 works out to $6.16 per GPU-hour on demand and $2.46 on spot, an H200 to $6.31, a B200 to $8.60, and an A100 to $2.70. These are list prices for full instances billed hourly, before any negotiated reserved rate.

Is CoreWeave spot capacity reliable enough for production?

It is preemptible by design. CoreWeave's docs say it can reclaim spot instances whenever capacity is needed elsewhere, with a maximum seven-minute window from the first warning to removal, and spot capacity is not guaranteed in any zone. Production use requires checkpointing and a scheduler that can lose nodes.

Should I prepay for a multi-year CoreWeave contract?

Verified prices alone do not answer that, but the counterparty picture is part of the decision: CoreWeave's 2032 bond yielded above 16% in early November 2026, and the company's own contract terms for neoclouds are exposed to delayed prepayments if capex slows. Run on-demand or spot first, and if you commit, keep the term short and the exit terms written.