Compare offer types
What differs between the ways compute is offered: who runs the hardware, how you pay, how long you commit and where your data sits. A framework, not a ranking.
Framework, not a statistic
The short answer
The main differences between offer types are who runs the hardware, how you pay, how long you commit for and where your data sits. This table sets those side by side. It is a framework for asking questions, not a ranking: the right type depends on your work, your rules and your timing. Payment is described by model only; we link to providers for prices.
What differs by offer type
| Offer type | Who runs the hardware | Payment model | Commitment shape | Where data sits |
|---|---|---|---|---|
| Hyperscale cloud instances | The cloud provider runs the buildings, servers and virtualisation. | Pay as you go by time, or a commitment for a rate, through one cloud account. | None for on-demand; a term for reservations and savings plans. | In the provider regions you choose, under the provider terms for that region. |
| GPU clouds and bare metal | The GPU cloud runs the servers, often in a colocation facility owned by someone else. | By time for on-demand, by term for reservations and clusters, sometimes by quote. | None for on-demand; a fixed term for reserved or dedicated clusters. | In the facilities the provider uses, which the contract should name. |
| Marketplaces and aggregators | Many separate hosts or clouds run the hardware; the platform runs the listing, billing and routing. | By time or by call, through the platform account; some use bids or host-set rates. | Usually none; reservations where the platform arranges them. | With whichever host runs the job, which may change from job to job. |
| Decentralised and community networks | Many independent operators run their own machines; a protocol or company coordinates them. | Per job or per time, often through a network token, credits or bids. | Usually none. | On operator machines that may be anywhere, unless the network lets you choose. |
| Regional and sovereign clouds | A provider based in, or operating under the law of, the chosen country or bloc. | By time or by contract, sometimes through public sector frameworks. | Ranges from none to multi-year sovereign contracts. | Kept in the named country or bloc, as far as the contract says. |
| Managed inference and serverless GPU | The platform runs the hardware and the model serving software. | Per token, per call or per second of execution; dedicated deployments by time. | Usually none; reserved throughput on a term. | Where the platform processes requests, and wherever it stores logs. |
| Chip-maker and accelerator clouds | The chip maker or its cloud runs its own accelerators. | By time for instances, per token or call for APIs, by contract for dedicated systems. | None to a term, depending on the shape. | In the regions where the chip is installed. |
| National and institutional HPC, and free or funded programmes | A public body, university or research centre, or a company giving credits on its own cloud. | Allocation, grant or credit by application; a free tier for some. | A project period set by the scheme. | On the named system or cloud, under the scheme’s rules. |
| Colocation, wholesale, powered shell and build-to-suit | The operator runs the building; you or your partner run the servers. | Rent for space and power, on a lease or service contract. | Usually a multi-year term; longest for wholesale and build-to-suit. | On your own machines, in a site you choose. |
| On-premises hardware, and leasing or financing it | You run the hardware, or a vendor manages it on your site. | Purchase, or lease and finance payments over a term. | The life of the hardware or the lease term. | On your own premises or your chosen colocation site. |
| Capacity brokers, reserved contracts, secondary and spot markets | The underlying provider runs the hardware; a broker may sit between you. | Committed term or block for reservations; time-based for spot; fees or margins for brokers. | Fixed term or window for reservations and blocks; none for spot. | With the underlying provider, which a brokered contract should name. |
| Edge and telecom sites, and converted sites | An edge or network operator runs many small sites; a site owner runs a converted site. | By time or call at the edge; lease or service contract for converted sites. | None for edge functions; a term for leases. | Spread across many locations at the edge; at one site for a converted building. |
Questions that differ most by type
Every type needs the questions on the buyer's checklist. These are the ones whose answer changes most from type to type.
| Offer type | Data residency | Exit | Also ask |
|---|---|---|---|
| Hyperscale cloud instances | Which regions offer the accelerator you need, and are backups, logs and support access kept in the same place? | How do you export machine images, data and configuration, and what happens to committed spend if you leave early? | Is quota for the accelerator granted on request, and how long does an increase take? |
| GPU clouds and bare metal | In which country and which facility will the machines sit, and is that fixed in the contract? | How are disks wiped at the end, how do you get your data out, and can a reservation be cut short or passed to someone else? | What network joins the machines, for example InfiniBand, and is it shared with other customers? |
| Marketplaces and aggregators | Can you choose and pin the country of the host, and does the platform tell you who the host is? | Is your data removed from a host when the rental ends, and can you leave the platform without losing reservations? | Who is your contract with: the platform, the host, or both? |
| Decentralised and community networks | Can you choose the country of the operator, and can operators see your data or model? | If payment is in a token, how do you convert and account for it, and what are the tax and accounting rules where you are? | How is the result checked, so you know the operator really ran the work? |
| Regional and sovereign clouds | Which law governs the provider and its parent company, and can a foreign authority compel access? | Can data and keys be returned and deleted in-country, under a written procedure? | Who holds the encryption keys, and where are the staff who can access the systems? |
| Managed inference and serverless GPU | Where are requests processed, are prompts and outputs stored or used for training, and can you switch that off? | Can you run the same open model elsewhere with the same results, or are you tied to a proprietary format or API? | What happens to a cold start, the delay when an idle model or function wakes up? |
| Chip-maker and accelerator clouds | In which regions is the chip available, which may be fewer than for GPUs? | Will your models and code run elsewhere without the custom chip, and at what effort? | Which frameworks and precisions are supported, and how much porting work is expected? |
| National and institutional HPC, and free or funded programmes | Where is the system, and are there rules on which data you may bring to it? | What happens when the allocation or credit runs out, and do you have a paid route to continue? | What must you report or publish in return, and who owns the results? |
| Colocation, wholesale, powered shell and build-to-suit | Where exactly is the site, who owns the land and building, and what local rules apply? | How much notice to leave, who removes the equipment, and who owns any fit-out you paid for? | When will the power be ready, and is the grid connection signed or still in a queue? |
| On-premises hardware, and leasing or financing it | The data stays where you put it; check any remote support access the vendor needs. | What is the resale or return route at the end of the lease, and how are disks wiped? | What lead time does the vendor give for delivery, and is there power and cooling ready for it? |
| Capacity brokers, reserved contracts, secondary and spot markets | With a broker, which provider and site will you actually get, and can you refuse a substitute? | Can the reservation be cancelled, resized or resold, and what does that cost? | Is "reserved" a billing discount or a physical set-aside of machines? |
| Edge and telecom sites, and converted sites | Which countries do the edge locations sit in, and can you limit where requests run? | For converted sites, what are the planning, environmental and power terms if the use changes again? | For converted sites: is the power connection rated and permitted for data centre use, not only the old use? |
Related
- Glossary: On-demand, Spot, Reserved, Capacity block, Colocation, Powered shell, Sovereign cloud, Egress, SLA (service level agreement), Data residency.
- Guide: What is a capacity block, and the other contract shapes
- Guide: How to get GPUs this week
- Provider directory and free and funded compute
Definitions and questions written in our own words, checked on 3 October 2026 and due for review by 3 January 2027. Sourced facts stay on the provider, programme and policy pages this page links to. This is general information to help you ask better questions. It is not legal, tax or procurement advice. For export controls, sanctions and contracts, check the primary rules and talk to a qualified adviser. Something wrong? Tell us.