Our last piece looked at one GPU cloud, Nebius, nudging its on-demand H100 up to $4.50 an hour. Widen the lens and a single provider's increase is almost a rounding error. In October 2026 the same H100 — the same silicon, the same 80GB — rents for anywhere from under $3 an hour to more than $11, depending only on where you put it. For a team pricing an AI server, where you rent now matters more than which card you rent.
The same chip, a near-fourfold price range
Checked against providers' own pricing pages in the second week of October 2026, on-demand H100 rates spread remarkably. At the low end sit the specialist GPU clouds: Runpod lists the H100 PCIe at $2.89 an hour, the SXM variant at $3.99 and the NVL at $3.19; Lambda lists a single H100 SXM at $3.99 per GPU-hour. Nebius, after its October 1 rise, sits at $4.50. AWS's prepaid Capacity Blocks for a p5 node work out to $5.97 per GPU-hour. Then the hyperscaler on-demand rates pull away: Google Cloud's a3-highgpu-8g is $88.49 an hour for the eight-GPU node, about $11.06 per card.
A survey of the same pricing pages on October 10 put Microsoft Azure's eight-GPU H100 virtual machine higher still, near $12.29 a card, Oracle's eight-H100 shape at $10.00 per GPU-hour, and marketplace capacity on Vast.ai as low as $1.68. From the cheapest verified list price to the dearest, that is a near-fourfold range on providers' own pages — wider still once the marketplace floor is counted — for a card that is, physically, identical.
What the gap is on the monthly bill
GPU servers run flat out and in multiples, so the hourly gap compounds into real money. A single H100 left running around the clock for a 730-hour month costs about $2,110 at Runpod's PCIe rate and about $8,074 on Google Cloud on demand — a difference of nearly $6,000 a month for the same unit of compute. NVLink cards are usually taken eight to a node: that is roughly $16,900 a month on the cheap tier against about $64,600 on Google's on-demand node, close to $47,700 a month apart. Nebius's $4.50 lands at about $3,285 a month per card; AWS's prepaid block at about $4,360. The point is not that one number is right and the others wrong — it is that the provider decision moves the bill by more than any chip-generation upgrade or any single provider's price change does.
Why identical hardware prices so differently
Part of it is what is bundled. The hyperscaler rate buys an on-demand machine inside a full platform — managed networking, identity, support, instant capacity and no commitment — and that convenience is priced in. The specialist clouds sell closer to raw metal: you get the card and little else, often from a smaller region footprint, sometimes on a community or marketplace tier where another customer's idle hardware is being resold.
Commitment is the other axis. On-demand is the most expensive way to run a steady load everywhere; AWS's Capacity Block rate already sits below its own on-demand rate, and committed or reserved terms on any of these clouds run well under the hourly figure. The quoted hour is the price of flexibility, and the hyperscalers charge the most for it. A nominally faster card per dollar does not help if the rate you can actually book it at is triple what a specialist charges for last generation's silicon.
What a buyer should do
Treat the GPU market as you would any commodity with a wide spread: shop it, and match the commitment model to the workload. For bursty, experimental or inference-with-idle work, on-demand on a specialist cloud captures most of the saving with none of the lock-in. For anything that runs most of the month — production inference, long training jobs — a reserved term or prepaid block, even on a pricier-looking provider, usually beats cheap on-demand once utilisation is high.
And read past the headline hour. Egress, storage, the minimum node size and regional availability all land in the bill, and a $3 card with expensive egress can lose to a $5 one without. Our AI workload page covers sizing the server to the model rather than to the marketing, and the catalog lets you compare providers on price per resource instead of per press release.
Figures here are from the providers' own published compute pricing pages — Runpod, Lambda, Nebius, AWS, Google Cloud, Oracle and Microsoft Azure — as listed in the second week of October 2026, cross-checked against an October 10 survey of the same pages by IntuitionLabs.