Which big GPU cloud is cheapest for H100 and B200? CoreWeave vs Nebius vs Lambda vs Crusoe

Which big GPU cloud is cheapest for H100 and B200? On each provider's own pricing page on November 3, 2026: for the H100, Crusoe at $3.90 per GPU-hour on demand, with Lambda a cent behind at $3.99 on an 8-GPU instance. For the B200, Lambda at $6.69 per GPU-hour (8x), well under Nebius ($8.50) and CoreWeave ($8.60). CoreWeave is the most expensive on demand, but if your job can be interrupted its published spot prices ($2.46 an H100, $4.26 a B200, per GPU) undercut everyone's on-demand rate. Two dates to know: Crusoe's own page says its H100 goes to $4.50 on December 1, 2026, which hands the cheapest on-demand H100 to Lambda.

Which cloud has the cheapest H100 and B200, side by side?

Here is every published per-GPU price from the four providers, read on their own pricing pages on November 3, 2026. CoreWeave lists prices per 8-GPU node, so we divided by eight.

Crusoe doesn't appear in the B200 chart because its page says "Contact sales" for the B200. If you're comparing H200s instead, Crusoe was cheapest at $4.29 until December 1; see our H200 rental cost guide.

Is Crusoe's H100 still the cheapest after December 1?

No, if the prices on its page hold. Crusoe's table has two columns, "Current on-demand" and "New on-demand starting Dec 1, 2026", and the H100 moves from $3.90 to $4.50 per GPU-hour. That puts it level with Nebius ($4.50) and above Lambda ($3.99 on 8x). The H200 moves from $4.29 to $5.40, the same as Nebius's H200. So if you book on-demand H100 time today for a run that crosses December 1, price the December hours at $4.50.

For a single H100 rather than a node, the cheapest self-serve option among the four is Lambda's 1x H100 PCIe at $3.29 per hour. Lambda's 1x H100 SXM is $4.29.

Why is CoreWeave the most expensive, and when is it the cheapest?

CoreWeave sells whole 8-GPU HGX nodes, not single cards: on November 3, 2026 an H100 node was $49.24 an hour and a B200 node $68.80 on demand. Divided by eight that's $6.16 and $8.60 per GPU, the highest of the four. Those per-GPU figures also appear in a separate column on CoreWeave's page, but its footnote says GPU-based pricing "is exclusively applicable to the CoreWeave inference platform customers."

The premium buys scale and contracted capacity, not a faster chip. CoreWeave's Q2 2026 results release (CoreWeave Investor Relations) says its revenue backlog "was approximately $104 billion as of June 30, 2026", and its tables show $2,575 million of quarterly revenue against $1,212 million a year earlier, about 112% more by our arithmetic.

The twist is spot. CoreWeave's table has a "Spot Price (Per Hour)" column: $19.71 for the H100 node and $34.11 for the B200 node in North America, or $2.46 and $4.26 per GPU. That's below every on-demand price on this page, Crusoe's and Lambda's included. Spot capacity can be taken back, and it runs under CoreWeave's Spot Terms of Service, so it suits checkpointed training and batch jobs, not a demo that has to stay up.

What do Nebius and Lambda clusters cost for longer runs?

More per GPU than their self-serve instances, which surprises people. Lambda's 1-Click Clusters (2 weeks to 1 year, 16 or more GPUs) were listed on Nov 3, 2026 at $9.86 down to $8.87 per B200 GPU-hour and $6.16 down to $5.54 per H100, depending on cluster size, against $6.69 and $3.99 for an 8x instance. You pay for a reserved, interconnected cluster, not just the GPUs. Lambda says to contact it for reserved capacity "at our lowest prices", and Nebius says commitment discounts let you "pay less than on-demand rates by reserving large-scale clusters for multiple months". Neither publishes those reserved rates.

Is there anything cheaper than the big four?

For the H100, yes, by a lot. For the B200, only some.

• Hyperstack listed on-demand H100 at $2.50 (PCIe), $2.60 (NVLink) and $3.20 (SXM), H200 SXM at $3.99, B200 at $6.00 and B300 at $7.40 per hour on Nov 3, 2026. That B200 price is the lowest data-center on-demand B200 we found on a public price list that day.
• GMI Cloud listed H100 from $2.60 and H200 from $3.20 per GPU-hour; its B200 "from $5.00" is marked contact sales.
• Runpod (Secure Cloud pods) listed H100 PCIe at $2.89, H100 SXM at $3.99, H200 at $5.29 and B200 at $7.99 per hour. The same page's product data also lists Community Cloud rates (B200 $5.98, H100 PCIe $1.99), but its docs say "Runpod is no longer accepting new hosts for Community Cloud. Existing Community Cloud resources remain available," so that capacity is limited to the hosts already there.
• Vast.ai, a marketplace, had H100 SXM from $2.40 and H100 PCIe from $2.54 per hour on demand in our 15:20 UTC read on Nov 3, 2026. Its cheapest on-demand B200 was $10.25, more than any big cloud, so the marketplace isn't the cheap route for Blackwell right now.

Marketplace and budget clouds trade support, guaranteed capacity and multi-node networking for the price. For one or two GPUs on a personal or small-team job, that trade usually pays; for a 64-GPU training run, it usually doesn't. Today's live prices for every card we track are on the cloud-GPU compare page, and the cheapest H100 rental guide goes deeper on single-H100 options.

H100 or B200: which should you rent?

Rent by what fits and what finishes fastest, not by the hourly rate alone. A B200 has 180 GB of memory per GPU on these price tables, against 80 GB for an H100, so a model that needs two H100s may fit on one B200. Check the fit for your model and context length in our VRAM calculator, then compare the cost of the whole job in H100 vs B200 rental cost. If you're weighing renting against owning a card at all, see renting vs buying an RTX 5090.

Crusoe is the one of the four with AMD GPUs on its GPU price table: on Nov 3, 2026 the MI300X (192 GB) was $3.45 per GPU-hour, with no change listed for December 1, and the MI355X was contact-sales. That matters only if your stack runs on ROCm rather than CUDA.

The rest of the cloud-GPU guides

This page covers the big neocloud tier. For one or two GPUs, most people rent at the budget and marketplace end: start with the live cloud-GPU compare page, then Runpod vs Vast vs Lambda, the Vast.ai review, the Runpod review, the best cloud GPU for fine-tuning, cloud vs local GPU for AI, Runpod Serverless pricing explained and the cheapest cloud GPU for ComfyUI and Flux.