> ## Content Index > Fetch the complete content index at: https://www.sotkn.com/llms.txt > Use this file to discover other available public pages before exploring further. # Reserved vs On-Demand GPUs: Who Pays for the Idle Hour? - URL: https://www.sotkn.com/explained/reserved-vs-on-demand-gpu/ - Published: 2026-10-24T09:49:51.000Z - Updated: 2026-10-24T09:49:51.000Z - Description: A reserved GPU bills for its whole term; an on-demand GPU bills only while it runs. Neoclouds live on the first, the second sets the price signal, and in 2026 the discount for committing has been shrinking. - Author: Sean - Tags: Versus, #desk-explained EXPLAINED · Versus · Last reviewed: 24 Oct 2026 · Next review: after CoreWeave and Nebius report third-quarter 2026 results, or when AWS updates Capacity Block prices (scheduled January 2027) A reserved GPU is booked for a set term and billed whether or not it runs; an on-demand GPU is billed only for the time it actually runs. **In short.** Reserved and on-demand are two ways to buy the same GPU hour. Reserved means booking a block for weeks, months or years and paying for every hour of it; on-demand means paying only for the hours you run. The real difference is who carries the idle hour: on-demand leaves it with the operator, reserved hands it to the buyer. That is why neoclouds live on reservations: CoreWeave said committed take-or-pay contracts were 98% of its revenue in the second quarter of 2026\. In 2026 the discount for committing has been shrinking. AWS’s reserved H100 block now costs about 13% less than on-demand, down from about 25% in September, and short reservations sell at a premium per megawatt. The stuck step is powered capacity: a reservation only starts billing once the hall is live and the GPUs are delivered. Why an investor cares: the split between reserved and on-demand decides how predictable a GPU cloud’s revenue is, who absorbs a demand slowdown, and what lenders will fund. A neocloud with mostly reserved capacity has contracted cash; one selling mostly by the hour has a price-taker’s business. Our [GPU as a service explainer](https://www.sotkn.com/explained/gpu-as-a-service-explained/) covers the hour itself, and [take-or-pay GPU contracts](https://www.sotkn.com/explained/take-or-pay-gpu-contracts-explained/) covers the clause that makes a reservation bill when it sits idle. This page puts the two ways of buying side by side. ## What is the difference between reserved and on-demand GPUs? A reserved GPU is paid for by the calendar; an on-demand GPU is paid for by the clock. With a reservation the buyer books a block of GPUs for a term and owes every hour of it, whether the cluster runs flat out or sits waiting for the next training job. With on-demand the buyer starts an instance, pays while it runs and stops paying when it stops. Nebius, for example, bills by the second. There is a third shape, spot or preemptible capacity: spare GPUs sold cheaply on the condition that the provider can take them back. All three can be the same chip in the same hall. What differs is the price, the certainty of getting the GPUs, and who is left holding an empty hour. | | Reserved (committed) | On-demand | Spot (preemptible) | | ---------------------------------- | --------------------------------------------------- | ------------------------------------- | ----------------------------------------- | | **What you buy** | A block of GPUs for a set term | GPU time, started and stopped at will | Spare GPU time the provider can take back | | **Term** | Days to years; neocloud deals often multi-year | None | None; can end at short notice | | **When the bill runs** | Every hour of the term, from start date or delivery | Only while the instance runs | Only while it runs, at a floating price | | **Who carries the idle hour** | The buyer | The operator | The operator, selling it cheap | | **H100 posted price, 24 Oct 2026** | AWS Capacity Block $5.97 | AWS $6.88 · Nebius $4.50 | Nebius $0.79 to $4.49 | | **Supply certainty** | Guaranteed once delivered | Only if spare capacity exists | None | | **What it tells investors** | Contracted cash and backlog | The walk-up price of scarcity | How full the clusters are | *Prices per GPU-hour from AWS and Nebius public price pages, checked 24 Oct 2026; AWS on-demand is p5.48xlarge in US East (N. Virginia) divided by eight GPUs. List rates, before negotiated discounts.* ## Is a reserved GPU cheaper than on-demand? Per hour, usually; overall, only if the buyer uses most of the hours booked. AWS publishes both prices for the same eight-GPU H100 server. On 24 October 2026 its on-demand rate in Northern Virginia was $55.04 an hour, or $6.88 per GPU. A Capacity Block, AWS’s dated reservation for a GPU cluster, cost $47.757 an hour, or $5.97 per GPU. That is about 13% less. The reservation bills every booked hour, so the discount only pays off if the cluster is busy. Our arithmetic: divide the reserved rate by the on-demand rate and you get the share of booked hours a buyer must actually use to come out ahead. At today’s AWS prices that is about 87%. A team that uses its block 60% of the time is effectively paying almost $10 for each H100 hour it runs. | AWS H100 (p5.48xlarge) | Reserved block, per GPU-hour | On-demand, per GPU-hour | Reserved discount | Use needed to beat on-demand | | -------------------------- | ---------------------------- | ----------------------- | ----------------- | ---------------------------- | | **Reading of 24 Sep 2026** | $5.19 | $6.88 | about 25% | about 75% of booked hours | | **Since 7 Oct 2026** | $5.97 | $6.88 | about 13% | about 87% of booked hours | *Our arithmetic from AWS’s posted Capacity Block and On-Demand rates. Break-even use = reserved rate ÷ on-demand rate. It assumes on-demand capacity is available when needed, which for large clusters it often is not.* The threshold moved in October. AWS raised Capacity Block prices on 7 October 2026, and the H100 block went from the $5.19 we read on 24 September to $5.97, a rise of about 15%. AWS says block prices follow supply and demand and are next due to change in January 2027\. On-demand prices were left alone, and the on-demand H100 rate is still the one set after AWS cut it 44% in June 2025\. Our read: the provider is charging more for the certainty, not for the chip. [GPU utilisation](https://www.sotkn.com/explained/gpu-utilization-explained/) is the number that decides which side of that line a buyer sits on. ![Reserved vs on-demand GPU prices: what one H100 hour costs on AWS on-demand, AWS Capacity Blocks and Nebius on-demand and spot, 24 Oct 2026](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/reserved-vs-on-demand-gpu-h100-hour.png) Posted list prices per H100 GPU-hour, checked 24 Oct 2026\. AWS 24 Sep reading from Second Order. ## Why do short reservations cost more than long ones? Because in 2026 buyers pay extra to get capacity now. The usual rule is that a longer commitment earns a bigger discount. The neoclouds’ own numbers show the opposite at the short end, measured as annual revenue per megawatt of power the GPUs need. Nebius said in its second-quarter letter of 12 August 2026 that its 2026 base ran at about $12 million of annual contract value per megawatt. Its four large deals signed in the second quarter averaged more than $1 billion each at above $20 million per megawatt, and short-term capacity deals in early third quarter were above $40 million. CoreWeave told investors on 17 September 2026 that short-dated contracts of three to six months signed in the third quarter priced at about $40 million per megawatt. Our arithmetic: a short reservation is earning roughly twice the revenue per megawatt of the long deals signed one quarter earlier, and more than three times Nebius’s 2026 base. The buyer pays for the hall being ready today. Our [CoreWeave short-contract Wire](https://www.sotkn.com/wire/coreweave-40m-mw-short-contracts-4-2gw-wire/) has the detail. ## Why do neoclouds sell mostly reserved capacity? Because reserved contracts hand the risk of an idle GPU to the customer, and that contracted cash is what pays for the chips. CoreWeave’s 10-Q for the quarter to 30 June 2026 says it sells either through committed contracts, which are take-or-pay, or on demand, which is pay-as-you-go. Committed contracts were 98% of its revenue in that quarter, the same as a year earlier. The same filing names the risk: a move by customers toward pay-as-you-go would hurt its ability to forecast. Lenders fund GPU clusters against those contracts, which is the subject of our [GPU financing explainer](https://www.sotkn.com/explained/gpu-financing-explained/). Reservations can also bring cash in before the hall is finished. Nebius said 70% of its second-quarter deals included prepayments covering 50% to 60% of the related capex, helping cut the expected payback on that spending to one year and ten months. Our read: for [CoreWeave](https://www.sotkn.com/stack/neoclouds/what-does-coreweave-do/) and [Nebius](https://www.sotkn.com/stack/neoclouds/what-does-nebius-do/), on-demand is the shop window, not the shop. It shows what a walk-up hour is worth, while almost all the money comes in through reservations. ## Where does spot GPU pricing fit? Spot is how a GPU cloud sells hours that would otherwise sit empty. The buyer gets a low price; the operator can take the GPU back when a full-price customer arrives. Nebius gives a 60-second warning before stopping a preemptible machine. Since 8 October 2026 Nebius’s spot price floats with how full its clusters are, and it can change as often as every 15 minutes. For an H100 the floor is $0.79 an hour and the ceiling is one cent below on-demand, $4.49\. Our arithmetic: eight H100s for a 730-hour month cost about $4,614 at the floor and $26,280 at Nebius’s on-demand rate. On AWS the same month on demand is about $40,179, and on a Capacity Block about $34,863\. Our [Nebius spot pricing Wire](https://www.sotkn.com/wire/nebius-spare-gpu-hours-spot-pricing-wire/) covers the switch. ![Reserved vs on-demand GPU: who carries the idle hour under on-demand, spot, reserved and short reservations, with powered capacity as the stuck step](https://storage.ghost.io/c/cd/10/cd10e222-131f-4ba4-b4a9-ca4214e8c12d/content/images/2026/10/reserved-vs-on-demand-gpu-idle-hour.png) Stuck step: powered capacity. Second Order. ## Who gets paid, and where is the stuck step? Under a reservation the operator is paid for every hour from the start of the term, and the buyer carries any idle time. Lenders are paid out of that contracted cash. Under on-demand the operator is paid only for hours that someone actually runs, so an empty cluster is the operator’s loss. On-demand prices still matter, because they set the walk-up price that buyers compare a reservation against. Nebius raised its on-demand H100 rate from $3.85 to $4.50 on 1 October 2026, which the [Nebius on-demand Wire](https://www.sotkn.com/wire/nebius-on-demand-gpu-hours-up-21-percent-wire/) covered. A higher walk-up price makes a reservation look better to a buyer, even as its own price rises. The stuck step is powered capacity. CoreWeave’s filing says committed contracts usually start on a fixed date or when it delivers the capacity in the contract. A reservation on paper is [remaining performance obligation](https://www.sotkn.com/explained/remaining-performance-obligation-explained/), not revenue, until the hall is energised and the GPUs are handed over. CoreWeave had about 4.2 GW of contracted power and about 1.5 GW active. Our read: the gap between those two numbers is the queue of reservations waiting to start billing. [Our CoreWeave power-on analysis](https://www.sotkn.com/analysis/coreweave-power-on-stuck-step-what-price-assumes/) looks at that clock. ## What would change the view? - **The reserved share falls.** If CoreWeave’s committed share drops well below 98%, or Nebius reports more pay-as-you-go revenue, the idle-hour risk is moving back to the operators. - **The short-term premium disappears.** If three-to-six-month deals stop pricing above long deals per megawatt, the scarcity of ready capacity is easing. - **AWS cuts block prices in January.** A lower Capacity Block rate at the next scheduled update would mean demand for guaranteed clusters has cooled. ## What to check next Watch CoreWeave’s and Nebius’s third-quarter 2026 results for the committed share of revenue, per-megawatt pricing on new deals and active versus contracted power. Then watch AWS’s Capacity Block update, scheduled for January 2027\. The [who are the neoclouds](https://www.sotkn.com/explained/who-are-the-neoclouds/) page lists the sellers, and [neocloud vs hyperscaler](https://www.sotkn.com/explained/neocloud-vs-hyperscaler/) explains why the same GPU hour earns differently in each. ## Common misconceptions - **People assume reserved is always cheaper.** Actually it is cheaper per hour but bills idle hours too; at AWS’s current H100 prices it only wins above about 87% use. - **People assume on-demand prices are what neoclouds earn.** Actually almost all CoreWeave revenue comes from committed contracts; the posted hourly price is the shop window. - **People assume longer commitments always get bigger discounts.** Actually in 2026 short three-to-six-month reservations earn the most per megawatt, because buyers pay for capacity that is ready now. - **People assume a signed reservation is revenue.** Actually it starts billing only when the capacity is delivered; until then it is backlog. ## FAQ **What is the difference between reserved and on-demand GPUs?** A reserved GPU is booked for a fixed term and billed for every hour of that term, used or not. An on-demand GPU is billed only while it runs and can be stopped at any time. Reserved buys certainty of supply; on-demand buys flexibility. **Is a reserved GPU cheaper than on-demand?** Usually per hour, but not always overall. On AWS on 24 Oct 2026 an H100 Capacity Block cost $5.97 per GPU-hour against $6.88 on-demand, about 13% less. Because a reservation bills idle hours too, it only works out cheaper if the buyer uses roughly 87% or more of the booked time. **What is a spot or preemptible GPU?** A spot or preemptible GPU is a spare GPU sold cheaply that the provider can take back when a full-price customer needs it. Nebius gives 60 seconds of warning before stopping one, and since 8 Oct 2026 its H100 spot price has floated between $0.79 and $4.49 an hour. **Why do neoclouds prefer reserved contracts?** Because reserved contracts move the risk of an idle GPU from the operator to the customer, and lenders will fund GPUs against that contracted cash. CoreWeave said committed take-or-pay contracts were 98% of its revenue in the second quarter of 2026. **What is a GPU capacity block?** A capacity block is a reservation for a cluster that starts and ends on set dates. On AWS the fee is paid up front at the price in force on the day of purchase, even if the block starts after a price change. **Why are short GPU reservations priced higher than long ones?** Because buyers who need capacity now pay for scarcity. Nebius said its short-term deals in early Q3 2026 earned above $40 million of annual contract value per megawatt, against above $20 million for its large Q2 deals and a 2026 base of about $12 million. **Is this a list of stocks to buy or sell?** No. This page explains how reserved and on-demand GPU pricing works and where the money goes. It does not rank companies or give buy or sell calls. Not investment advice. Do your own research. Related: [GPU as a service](https://www.sotkn.com/explained/gpu-as-a-service-explained/); [take-or-pay GPU contracts](https://www.sotkn.com/explained/take-or-pay-gpu-contracts-explained/); [GPU utilisation](https://www.sotkn.com/explained/gpu-utilization-explained/); [what is a neocloud](https://www.sotkn.com/explained/neocloud-explained/); [AI tokens vs GPU hours](https://www.sotkn.com/explained/ai-tokens-vs-gpu-hours/); [GPU financing](https://www.sotkn.com/explained/gpu-financing-explained/). ## Stack doors Layer: [Neoclouds](https://www.sotkn.com/stack/neoclouds/). Also: [Hyperscalers](https://www.sotkn.com/stack/hyperscalers/) and [AI Data Centers](https://www.sotkn.com/stack/ai-data-centers/). Markets: [US](https://www.sotkn.com/markets/us/), where the listed GPU clouds report. [Get the Wire](https://www.sotkn.com/#/portal/signup) — short daily notes when GPU prices, contracts and capacity move. Not investment advice. Do your own research. **How we check this** Figures come from company filings and price pages, each dated in the list below: CoreWeave’s Form 10-Q for the second quarter of 2026 and its 17 September 2026 investor presentation, Nebius’s second-quarter 2026 shareholder letter, and the public price pages of Nebius and AWS, each re-checked on 24 Oct 2026\. The 24 September AWS Capacity Block price is our own earlier reading of the same AWS page. Per-GPU prices, monthly costs, discounts and break-even use are our arithmetic from those posted rates, and the working is shown on the page. Posted prices are list rates, not what large customers pay after negotiation. Where we interpret, it is labelled “our read”. This page explains how reserved and on-demand GPU pricing works. It does not rank companies or give buy or sell calls. Not investment advice. Do your own research. **Last reviewed:** 24 Oct 2026 · **Next review:** after CoreWeave and Nebius report third-quarter 2026 results, or when AWS updates Capacity Block prices (scheduled January 2027) Spot an error? [Tell us](mailto:sean@thekopinotes.com) and we will correct it and note the change here. [Our editorial standards →](https://www.sotkn.com/editorial-standards/) ## References 1. [CoreWeave — Form 10-Q for the quarter ended 30 Jun 2026 (filed 12 Aug 2026)](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000366/crwv-20260630.htm?ref=sotkn.com) — sells through committed contracts, which are take-or-pay, or on-demand, which are pay-as-you-go; committed contracts 98% of revenue in Q2 2026 and Q2 2025; a shift to pay-as-you-go would hurt forecasting 2. [CoreWeave — Investor presentation, Form 8-K exhibit 99.2 (17 Sep 2026)](https://www.sec.gov/Archives/edgar/data/1769628/000176962826000429/crwv2026convertpresentat.htm?ref=sotkn.com) — short-dated contracts of 3–6 months signed in Q3 at about $40M per MW of annualised revenue; contracted power about 4.2 GW, active about 1.5 GW 3. [Nebius Group — Letter to shareholders Q2 2026, Form 6-K exhibit 99.2 (12 Aug 2026)](https://www.sec.gov/Archives/edgar/data/1513845/000110465926094568/tm2622968d1%5Fex99-2.htm?ref=sotkn.com) — annual contract value per MW: 2026 base about $12M, Q2 deals above $20M, Q3 short-term deals above $40M; 70% of deals with prepayments covering 50–60% of capex; payback 1 year 10 months 4. [Nebius — AI Cloud pricing (checked 24 Oct 2026)](https://nebius.com/prices?ref=sotkn.com) — HGX H100 on-demand $4.50 per GPU-hour, preemptible from $0.79; B300 $9.50 on-demand, from $0.99 preemptible; highest preemptible rate one cent below on-demand; commitment discounts for multi-month clusters 5. [Nebius — Compute pricing documentation (checked 24 Oct 2026)](https://docs.nebius.com/compute/resources/pricing?ref=sotkn.com) — GPU prices updated from 1 Oct 2026; preemptible H100, H200, B200, B300 and RTX PRO 6000 moved to dynamic spot pricing from 8 Oct 2026; billing per second 6. [AWS — Amazon EC2 Capacity Blocks for ML pricing (checked 24 Oct 2026)](https://aws.amazon.com/ec2/capacityblocks/pricing/?ref=sotkn.com) — p5.48xlarge (8× H100) $47.757 per instance-hour, $5.970 per accelerator-hour; fee paid up front at the rate in force when bought; next scheduled price update January 2027 7. [AWS — Amazon EC2 On-Demand pricing (checked 24 Oct 2026)](https://aws.amazon.com/ec2/pricing/on-demand/?ref=sotkn.com) — p5.48xlarge Linux, US East (N. Virginia) $55.04 per hour; p6-b300.48xlarge $142.416 per hour 8. [AWS News Blog — Up to 45% price reduction for EC2 NVIDIA GPU instances (5 Jun 2025)](https://aws.amazon.com/blogs/aws/announcing-up-to-45-price-reduction-for-amazon-ec2-nvidia-gpu-accelerated-instances/?ref=sotkn.com) — P5 (H100) on-demand price cut 44% from 1 Jun 2025 S **[Sean](https://www.sotkn.com/author/sean/)** Writes Second Order by TKN — a plain-English map of AI infrastructure and semiconductors for non-specialist investors. Focus: who gets paid, where money sticks, and which step is stuck. Covers AI infrastructure, semiconductors, and second-order stack reads. [Author page →](https://www.sotkn.com/author/sean/) · [Editorial standards →](https://www.sotkn.com/editorial-standards/)