What is a capacity block, and the other contract shapes
The ways compute is sold, from on-demand to capacity blocks to reserved contracts, with what each one promises and what providers on this site say they offer.
The short answer
A capacity block is a fixed amount of accelerator capacity booked for a fixed window, paid for up front or on booking, so it is there when the window opens. It sits between on-demand (no promise) and a long reservation (a long promise). Providers use their own names and terms for the same shapes, so compare what is promised, not what it is called.
The shapes
| Shape | What it means | What you give up |
|---|---|---|
| On-demand | Start and stop when you like and pay for the time used. | No promise the capacity is there when you ask. |
| Spot or interruptible | Spare capacity on terms that let the provider take it back. | Your job can be stopped. |
| Capacity block | A set amount of capacity booked for a set window in advance. | Flexibility: the window and size are fixed. |
| Reserved or committed | A longer commitment for supply certainty, a different rate, or both. | You pay whether or not you use it. |
| Dedicated or bare metal contract | Whole machines set aside for you, see bare metal. | Lead time and a contract. |
| Allocation by application | Public programmes give compute after a proposal, see 52 programmes. | Time to apply, and the scheme's rules. |
| Colocation and powered shell | You bring the machines, or the fit-out, and rent the space and power. | You run the hardware or build the inside. |
| Sovereign cloud | A service designed to keep data and operations under a chosen country's or bloc's legal control. | Depends on what the contract actually guarantees. |
What providers on this site say
These are the purchase options the providers' own pages mention, as recorded in our directory. A provider missing here may still offer the shape; we only list what we have checked. We have not yet recorded any provider's own page for a capacity block product, so this site does not say who offers one. Ask providers directly, and tell us if you find a product page.
| Provider | What the provider's page says | Source |
|---|---|---|
| Alibaba Cloud | ECS lists subscription, pay-as-you-go, reserved and preemptible purchasing models. | Source |
| Amazon Web Services (EC2) | EC2 can be bought on demand, as spot capacity, or under longer-term savings plans. | Source |
| Fireworks AI | Offers three deployment tiers: serverless, on-demand dedicated deployments, and reserved capacity. | Source |
| Genesis Cloud | Offers committed-use discounts on selected instances for terms of a few months, and free credits for new users. | Source |
| Google Cloud (Compute Engine) | Spot VMs and committed use discounts are listed as purchase options. | Source |
| Google Cloud (Compute Engine) | Checked 14 November 2026: Google documents A5X with NVIDIA Vera Rubin GPUs as bare-metal instances. It directs users to contact their account team to reserve capacity; Spot and Flex-start are not supported, and Google warns that the Compute Engine SLA does not apply to A5X. | Source |
| Hyperbolic | Offers three access tiers: on-demand GPUs, reserved clusters with commitments under one year, and private cloud for longer-term dedicated capacity. | Source |
| Hyperstack (NexGen Cloud) | Pricing page describes per-minute on-demand billing, reservations and spot virtual machines. | Source |
| io.net | The home page describes on-demand GPU clusters and says it supports containers, Ray clusters and bare metal. | Source |
| Jarvislabs | The pricing page describes per-minute billing, spot instances and reservations from one month. | Source |
| Lambda | Lambda lists on-demand instances, 1-Click Clusters and single-tenant Superclusters as separate products. | Source |
| Massed Compute | Offers on-demand instances, bare metal single-tenant servers and clusters with InfiniBand. | Source |
| Massed Compute | The pricing page lists on-demand GPU rental by the hour with no contracts. | Source |
| Microsoft Azure | Spot virtual machines, reserved instances and savings plans are listed as ways to buy. | Source |
| Prime Intellect | Describes on-demand GPU access across cloud providers from a single platform, and requests for large clusters drawn from more than 50 providers (a company claim). | Source |
| RunPod | Pods come as reserved or spot, the latter being interruptible. | Source |
| Salad | Also offers reserved dedicated bare metal and virtual machines on contract. | Source |
| Shadeform | On-demand instances, on-demand clusters and reserved commitments are listed as options. | Source |
| UpCloud | Offers GPU servers, and lists reservations for B200 and B300 clusters in a planned Nordic AI data centre. | Source |
| Vast.ai | Both on-demand and interruptible rental modes are offered. | Source |
| Verda | Pricing page describes on-demand, spot, reserved and serverless billing models. | Source |
| Voltage Park | Pricing page lists on-demand Ethernet, on-demand InfiniBand and long-term reserve options. | Source |
How to compare two offers
- Write down the start date, the length, the number of accelerators and the network between them for each offer.
- Ask what happens if you need less, or more, part-way through.
- Ask what happens if the capacity is late or fails: is there a credit, a replacement, or nothing?
- Check the term "reserved" with the provider: a billing discount on usage and a physical set-aside of machines are different promises.
Sources
- Provider facts above: each row links the provider's own page.
- Glossary: definitions, not sourced statistics.
- Free and funded compute: programme pages for allocation routes.
Built from facts already on this site, written on 3 October 2026 and due for review by 3 January 2027. Each fact keeps the check date shown on the page it comes from. This is general information, not advice. Something wrong? Tell us.