The standard provider comparison — price per vCPU, region list, support tier — answers the wrong questions for AI work. Four different factors dominate.
1. Accelerator availability, not accelerator price
A cheaper hourly rate on capacity you cannot obtain is not cheaper. Before comparing anything else, establish:
- Which accelerator families the provider offers, and in which regions.
- What the default quota is for a new account, and how long an increase takes.
- Whether capacity is generally obtainable on demand or effectively requires a commitment.
This is the question that most often eliminates a provider outright.
2. Managed model services versus plain machines
Two quite different products get called “AI on cloud”:
- Managed model platforms, where you call hosted models through an API and never see an accelerator. Fast to adopt, priced per use, and the practical choice for most application work.
- Raw GPU instances, where you install the stack and run what you like. Necessary for custom training, custom serving and any requirement the managed catalogue does not cover.
The hyperscalers offer both and are usually chosen for the first. The VPS-oriented providers offer capable general compute and are frequently the better home for the surrounding application — see multi-provider cloud strategy.
3. Data gravity
Training data and model artefacts are large, and moving them repeatedly is expensive and slow. Whichever provider holds the data has a strong claim to hosting the compute, because the alternative is paying egress on every run.
If the dataset already lives somewhere, start the comparison from there rather than from a blank page.
4. What surrounds the model
A production AI feature is mostly not the model. It is an API, a queue, a database, storage, authentication, monitoring and a deployment pipeline. Those need to be good, cheap and boring, and that is an ordinary cloud comparison — see how to choose a cloud provider and the provider comparisons.
Three common shapes
- Application plus hosted model API. Application on whichever provider is cheapest and most predictable; model calls to a managed service. Lowest operational burden, and correct for most products.
- Application plus self-hosted inference. Application on a general provider, one GPU instance running the serving stack. Chosen for data residency or steady high volume — see running LLM inference on your own account.
- Everything on one hyperscaler. Simplest networking and no cross-provider egress, at the price of the highest bill and the strongest lock-in.
A short checklist
- Can I get the accelerator I need, in the region I need, on this account, this month?
- Where does the data already live?
- Do I need a model that a managed catalogue does not offer?
- What does the surrounding application cost on this provider?
- What does it cost to move away in a year?
Where to go next
Questions
Frequently asked
Which cloud provider is best for AI?
It depends on whether you need managed model services or raw accelerators, and on where your data already lives. Accelerator availability in your region eliminates more candidates than price does.
Do I need a hyperscaler for AI work?
For managed model catalogues and large-scale training, usually. For an application that calls a model API, the surrounding infrastructure can sit anywhere and is often cheaper elsewhere.
Does it matter where my training data is stored?
Considerably. Moving large datasets between providers costs egress on every run and adds latency, so the compute usually belongs next to the data.