Decide which of three things you need
AI workloads on cloud fall into three groups, and they point at different providers.
- Calling a hosted foundation model over an API, where you never touch a GPU.
- Training or fine-tuning a model, where you need a managed platform and accelerators.
- Running your own inference on raw GPU instances, where you want hardware and nothing else.
Most teams need the first. Paying for the third because it sounds more capable is a common and expensive mistake.
Hosted foundation models
If you are building on top of an existing model, the managed services are the sensible route. Amazon Bedrock, Azure OpenAI Service, Google Vertex AI and OCI Generative AI all expose models over an API with no infrastructure to manage.
The differences are which model families each carries, which regions they run in, and how access is gated. Azure OpenAI is the route to GPT-family models with enterprise controls; Bedrock carries a range of providers including Claude; Vertex AI is the home of the Gemini family alongside a large model garden.
Access is often quota-gated rather than simply available, so an account with existing approval is worth more than one without.
Training and fine-tuning
If you are training, you want a managed platform rather than bare instances: experiment tracking, distributed training, checkpointing and a deployment path are all work you would otherwise build.
SageMaker and Vertex AI are the most complete options here. Azure Machine Learning covers the same ground within the Microsoft ecosystem. IBM watsonx targets the same lifecycle with governance emphasis for regulated environments.
Raw GPU instances
For inference on your own model, or training where you already have the tooling, a GPU instance is often the cheaper answer. Providers outside the hyperscalers compete hard here.
DigitalOcean has GPU Droplets, Linode and OVHcloud both offer GPU plans, and Hetzner sells dedicated GPU hardware. The per-hour cost is typically well below hyperscaler pricing, with the trade being that you build the surrounding tooling yourself.
Quotas will be your first obstacle
GPU capacity is quota-controlled everywhere, and new accounts start at or near zero. You will usually need to request an increase, explain the workload, and wait.
This is worth planning around. An account that already has GPU quota or model access approved removes a genuine delay from the project.
Watch the cost shape
- GPU instances bill by the hour whether or not they are computing. Stop them when idle.
- Hosted model APIs bill per token, which scales with usage rather than time.
- Training data movement can generate significant egress if storage and compute are not in the same region.
- Checkpoints and datasets accumulate storage cost quietly.
Related accounts
Accounts on the platforms discussed here: Buy AWS Accounts, Buy Google Cloud Accounts, Buy Azure Accounts.
Related cloud accounts
Accounts on the platforms discussed here, with every configuration and price published up front:
- Buy AWS Accounts — from $25.00. Over 200 services, from EC2 to Bedrock.
- Buy Google Cloud Accounts — from $30.00. Vertex AI, Gemini models and BigQuery.
- Buy Azure Accounts — from $25.00. Azure OpenAI and deep Microsoft integration.
Further reading
- Serving a Static Site: Storage, CDN and What It Costs
- AWS vs UpCloud: Which Cloud Account Should You Buy?
- AWS vs Alibaba Cloud: Which Cloud Account Should You Buy?
- AWS vs Linode: Which Cloud Account Should You Buy?
Official documentation
Provider behaviour, quotas and pricing are defined by the platform itself. These are the first-party references for the services discussed above:
Frequently asked questions
No. Hosted foundation model services run the hardware for you and bill per token.
It depends on the model family you want. Azure for GPT-family access, Google Cloud for Gemini and Vertex tooling, AWS for the widest choice through Bedrock.
Almost always a quota of zero on that instance family in that region. It has to be requested and approved.
