Insights/Article

What is GPUaaS? How GPU as a service is priced, who it suits and how to buy it

GPUaaS, or GPU as a service, means renting GPU compute from a provider instead of buying and running the hardware yourself. The provider owns the GPUs, houses them in a data center and operates them. You pay for access, usually per GPU per hour, and you do not carry the capital cost, the power contract or the delivery risk of the cluster.

There are two prices for GPU compute. One is what providers publish on their websites. The other is what they quote when a buyer commits to a cluster. We place GPUaaS capacity with customers every week, so this article sets what we see in quotes against what the market presents outwardly. Our figures are asking terms from providers in November 2026. They are not contract prices.

The short answer

For B300, the published on-demand rate at Nebius is $9.50 per GPU per hour. The quotes we have received recently are $4.35, $4.50 and $4.55, which is less than half. GB300 and the newer parts are not published at all. We see GB300 quoted anywhere from $5.70 to $6.60, and Vera Rubin around $7 in most cases. The gap between the two prices is explained by the contract, not the chip.

What the market presents outwardly

Published pricing is on-demand pricing: hourly, flexible and the most expensive way to buy. Nebius lists HGX B300 at $9.50 per GPU per hour, B200 at $8.50, H200 at $5.40 and H100 at $4.50. CoreWeave lists HGX B200 at $68.80 per eight-GPU node, which is $8.60 per GPU, and H200 at about $6.31. Its HGX B300 and GB300 NVL72 are marked contact sales. Lambda publishes B200 cluster rates of $8.87 to $9.86 per GPU per hour for terms of two weeks to a year.

The newest hardware is where the published price disappears. GB300 has no public price at the providers above, and nothing is published for Vera Rubin. Anyone planning a cluster on current generation hardware cannot work from a rate card. They need quotes.

What we see in the market

Quotes for committed capacity sit far below the on-demand list. Recent B300 quotes have come in at $4.35, $4.50 and $4.55 per GPU per hour on three to five year terms. GB300 quotes run from $5.70 to $6.60. Vera Rubin is quoted at around $7 in most cases. Each step up the roadmap carries a premium, set by a market where supply of the newest parts is tight.

The spread inside each chip is wide, and it is explained by three things. The first is contract length. A provider financing a cluster wants enough contracted revenue to cover the hardware, so the longer term buys the lower rate, and a one year commitment can price well above a three or five year one. The second is the ready for service (RFS) date. The sooner the capacity is available, the more expensive it is, because racked and powered capacity carries a premium over capacity that is built to order. The third is prepayment. Providers ask for anything from 20% to 50% or more up front, and some price the biggest discounts only for large prepayments on the longest terms.

Small reserved blocks sold ready to go are priced differently again. We hold all-inclusive quotes for B300 servers well above $9 per GPU per hour, in line with the published on-demand rate, because the buyer is paying for speed and flexibility. The $4.50 quote and the $9.50 listing are both real. They are different products.

Hopper is scarce

Most published rate cards still show H100 and H200, and that gives a misleading picture. We see very few clusters of Hopper available anywhere. One provider told us directly that it has no Hopper available and is adding none. Where H200 capacity exists it is in small blocks. The few H200 asks we hold are between $6.25 and $8.86 per GPU per hour on one to five year terms, above the published on-demand rate of about $5.40 to $6.31, and we have nothing on H100. A buyer who plans around a Hopper price list is likely to find the capacity is not there.

Who GPUaaS suits

GPUaaS suits teams that want current generation GPUs without the capital outlay, the power and cooling contract or the delivery risk of building a cluster. It fits best where demand is proved and can be committed for several years, because that is where the lower rates sit. It fits less well where demand is uncertain, because reserved capacity is paid for whether or not the GPUs are busy.

What is in the quote and what is not

Ask every provider the same questions. Is the rate all-in, or are storage, networking and power separate lines? Is the term a minimum or a take-or-pay? How much is prepaid, and is any of it refundable or credited back if delivery slips? Which hardware generation is it, and what happens if you want to move to the next one during the term? What is the RFS date in writing, and what is the remedy if it is missed?

Then check the details that never appear on a rate card: cooling method, the network between the cluster and your data, support hours, and where the cluster physically sits.

White label GPUaaS and working with a partner

Resellers, software companies and cloud brands often want to sell GPU compute under their own name without owning the hardware. That is white label GPUaaS, and the commercial questions are who holds the capacity commitment, who carries the utilisation risk and who owns the customer relationship. A GPUaaS partner places the capacity, brings the provider and the customer together and structures those terms so each side knows what it has agreed to.

How to compare offers properly

Put every offer on one basis. Convert each quote to an all-in cost per GPU per hour, including storage, networking and any fees. Account for the prepayment, which ties up cash for the length of the term. Compare the RFS dates, because capacity that is cheaper but arrives months later has its own cost. Our cost of compute calculator and GPU cluster planner help set the options out like for like.

Armason: neutral introductions between sites, demand and capital

Armason introduces AI cloud providers and offtakers who need capacity to operators who can deliver it, in the US, Europe, the UK and beyond, with no bias towards either side of a transaction. If you are weighing GPUaaS against buying, or hold GPU capacity you want placed, we would like to hear from you.

This article is general information. Armason figures are asking terms from providers as at November 2026 and are not contract prices. Published figures are attributed to their sources. It is not investment or legal advice.

Related reading

· Cost of compute calculator

· GPU cluster planner

· Data centre offtake UK: how demand gets matched to capacity in a constrained market

Sources referenced in this article

· Nebius: AI cloud pricing (on-demand GPU-hour rates)

· CoreWeave: cloud pricing (per-instance on-demand rates)

· Lambda: AI cloud pricing (1-Click Clusters)

Related Insights
Work With Armason

If you’re sourcing data centre capacity or delivering it, we can help structure the engagement.

MI Roundup

Weekly market intelligence, every Monday.