GPU cloud prices are easy to find. Hyperscalers publish price lists and pricing APIs, neoclouds publish rate cards, and index publishers turn them into daily series. In 2026, exchanges proposed futures that settle on such series [24][26].
Each of these uses assumes a posted price measures something stable. It does, but the thing it measures is narrow: what one configuration costs, under one set of terms, at one moment, if capacity is there. This explainer shows what that means in practice.
One machine, many prices
Take one instance with fixed hardware: AWS's p5.48xlarge, with 8 NVIDIA H100 GPUs [6]. On 9 October 2026, AWS posted these prices for it in US East (N. Virginia), among others [3][4][5].
| Pricing type | Term and payment | Per GPU-hour |
|---|---|---|
| On-Demand | None | $6.88 |
| Capacity Block | Reserved block for a future date, paid upfront | $5.970 |
| Compute Savings Plan | 1 year, no upfront | $5.42 |
| EC2 Instance Savings Plan | 1 year, no upfront | $4.33 |
| EC2 Instance Savings Plan | 3 years, all upfront | $2.59 |
| Spot | Interruptible | $3.61 to $4.64 by Availability Zone |
Same chips, same region, same day. The on-demand rate is about 2.7 times the three-year rate. Spot prices differed by up to 28% between Availability Zones in the same region [5].
Each row answers a different question. On-demand is capacity with no commitment at a static hourly rate [1]. A Capacity Block reserves instances for a future date, cannot be cancelled and is paid whether or not it is used [2][3]. A Savings Plan commits to a stated amount of compute use, measured per hour, for one or three years [4]. Spot is spare capacity the provider can reclaim with two minutes' notice at AWS [1] and 30 seconds at Azure [11]. Google says its Spot prices can change up to once a day [10].
Posted prices move by decision
Posted prices change when the provider decides to change them. AWS cut H100 on-demand prices by 44% in June 2025 [7]. It raised Capacity Block prices by about 15% in January 2026 and about 20% in July 2026 [8][9]. In January an AWS spokesperson said its commitment not to raise On-Demand and Savings Plans prices was unchanged [8].
So two series built from the same provider can diverge: one from on-demand rates, one from reservations. One index publisher found its hyperscaler H100 on-demand index unchanged on 39 of 50 daily transitions between 1 March and 20 April 2026, and described it as a price ceiling posted above reserved capacity [14].
Availability flags are thin
Providers report availability, but in limited forms. Lambda's API lists "the regions in which this instance type is available" next to the price [12]. AWS's Spot placement score estimates how likely a Spot request is to succeed and "does not provide any guarantees" [13]. Azure shows hourly eviction rates from the past seven days [11]. None reports how many GPUs can be launched, for how long, or at what cluster size. A price without availability may describe capacity no one can buy at that moment.
A listing is not a transaction
A posted price is a listing: what the seller asks anyone for that configuration, at that time. Large contracts are priced privately. Nebius reported contracts at $20 million to $25 million of annual value per megawatt [20], a unit that never appears on a rate card. Nscale describes take-or-pay agreements with fixed pricing over the term [21].
Regulators have noticed the gap. The CFTC's 2026 request for comment on compute derivatives asks what proportion of compute transactions occur at publicly disclosed prices, and whether a contract should settle to a price computed from data the Commission may not be able to observe or verify [28]. It also asks what would stop a provider from moving a settlement index by adjusting a posted rate [28]. Index designs answer differently: ICE and Ornn describe Ornn's index as built only from printed transactions [26], while CME's proposed contracts reference Silicon Data indexes of on-demand rental rates [24][27].
Seven ways a comparison goes wrong
Before two posted prices can be compared, each of these has to match.
- GPU variant. An H100 SXM has 80 GB of memory and 900 GB/s NVLink. An H100 NVL has 94 GB and 600 GB/s [19]. Both are "H100."
- Interconnect. AWS posts a 1-GPU p5.4xlarge and the 8-GPU p5.48xlarge at the same $6.88 per GPU-hour [5]. Per GPU, the larger node has four times the network bandwidth plus the NVSwitch fabric [6].
- Node size. NVIDIA lists H100 SXM systems with 4 or 8 GPUs [19]. A single-GPU offer serves different work.
- Region. The same p5.48xlarge costs $8.944 per GPU-hour on demand in London and $6.88 in Virginia [5].
- Term and payment. Commitment alone moved the price 2.7-fold in the table above.
- Bundled CPU, memory and storage. The p5.48xlarge includes 192 vCPUs, 2 TiB of memory and about 30 TB of local NVMe [6]. A bare GPU listing is a smaller product.
- Billing rules. Upfront reservation fees, cancellation terms, operating system fees and interruption notice change the effective cost of equal hourly prices [1][2][3].
Index construction adds an eighth trap. One publisher's daily hyperscaler on-demand H100 index sat near $7.46 in May 2026 [15], while its quarterly median series put hyperscaler H100 medians at $4.80 to $6.10 for the third quarter through July [16]. Both are honest. They are built differently.
How the Rillor Compute Index methodology treats posted prices
The Rillor Compute Index methodology v1.0, published on 9 October 2026, handles each of these points [25]. A rental price published through a pricing API or page is a listing, and every observation carries its evidence class, so a listing never counts as a sale. Each observation records its source and capture time. Availability is recorded as the provider reports it. On-demand, spot and reserved prices stay separate, with reserved terms recorded in months, and hyperscaler and neocloud segments are shown separately. Prices are normalized per GPU-hour and mapped to GPU class and variant, interconnect, node size, region, term and pricing type. Each series takes a median for each provider, then the median across providers, so a provider with many listings counts once. Rillor's own quotes and agreements are excluded, and a series is published only when it meets the methodology's data rules. The series are for research and planning use and are not offered for settlement of any contract.
Read the full report
The full report, with the complete price tables, the normalization checklist, the method, its limits and 28 numbered sources, is available as a PDF.