This comparison of machine learning as a service providers covers the three that matter to most UK businesses: Amazon Web Services (AWS), Microsoft Azure and Google Cloud. The short answer is that their offerings are close enough that the best provider is usually the one your systems already run on. The harder part is working out what each one sells, because the product names have changed several times. Every name below was checked on the provider’s own site on 2 October 2026.
Machine learning as a service means renting machine learning from a cloud provider instead of building and hosting it yourself. You send data to a service and get an answer back, or you use the provider’s computers and tools to build a model of your own.
What the providers sell: three layers
All three providers offer the same three layers, and knowing which one you need narrows the choice quickly.
- Ready-made services. You send a document, an image or a recording, and get back the text, the fields or a transcript. No machine learning knowledge is needed, only a developer to connect it to your system.
- Hosted language models. Large general-purpose models that read and write text, used for summarising, answering questions from your documents, classifying messages and drafting replies. You pay for what you use and the provider runs the model.
- A platform for training your own model. Tools for data scientists to build, train and run a model on your own historical data, for example to forecast demand.
Most businesses need the first or second layer. The third is for cases where you hold a lot of your own data and want a prediction that no general service provides.
The three providers side by side
| Layer | AWS | Microsoft Azure | Google Cloud |
|---|---|---|---|
| Ready-made services | Amazon Textract, Rekognition, Transcribe, Translate | Foundry Tools: Document Intelligence, Vision, Speech, Translator | Document AI, Vision AI, Speech-to-Text, Translation AI |
| Hosted language models | Amazon Bedrock | Microsoft Foundry, including Azure OpenAI | Gemini Enterprise Agent Platform, with Gemini and Model Garden |
| Train your own model | Amazon SageMaker AI | Azure Machine Learning | Gemini Enterprise Agent Platform; BigQuery ML |
| Without writing code | Amazon SageMaker Canvas | Automated ML and the designer in Azure Machine Learning | AutoML |
What each product used to be called
Older articles, including the earlier version of this one, use names that no longer match what you will see when you log in.
- AWS. Amazon SageMaker was renamed Amazon SageMaker AI on 3 December 2024. The plain name Amazon SageMaker now refers to a wider platform for data, analytics and AI, of which SageMaker AI is one part.
- Azure. Azure Machine Learning has kept its name. The portal for language models has been Azure AI Studio, then Azure AI Foundry, and is now Microsoft Foundry. The ready-made services, once Cognitive Services and later Azure AI services, are now called Foundry Tools.
- Google. The earlier version of this article described Google’s AI Platform. That was replaced by Vertex AI, which Google now presents as Gemini Enterprise Agent Platform, with “formerly Vertex AI” beside the name. The documentation addresses still contain the words vertex-ai.
The renaming is itself worth noting. In our experience, machine learning products are renamed, merged and withdrawn more often than other cloud services. Microsoft’s own list of ready-made services marks six as retired, among them Anomaly Detector, Personalizer and QnA Maker, and it has announced that prompt flow, a tool for building language model applications, will be retired on 20 April 2027.
How they differ in practice
AWS keeps the layers as separate products that you assemble. Bedrock gives access to language models from a range of companies through one service. SageMaker AI is a full workbench for teams with machine learning engineers, and SageMaker Canvas offers a visual route for people who do not write code.
Azure is the natural fit for a business on the Microsoft stack. Access is controlled through Microsoft Entra, the same identity system as Microsoft 365. Microsoft Foundry offers OpenAI’s models alongside models from Microsoft, Anthropic, Meta and others. Azure Machine Learning can export a model trained by its automated tool to ONNX, an open model format, which a .NET application can then run directly. That matters if the system the prediction feeds into is written in C#.
Google builds its offering around its own Gemini models, with third-party and open models available through Model Garden. BigQuery ML is distinctive: it lets analysts create and run models using SQL, the query language they already know, on data held in Google’s BigQuery data warehouse. If your reporting data already lives there, that is a short path to a first model.
All three charge by usage, not by licence: by the page, by the minute of audio, by the amount of text processed or by the hour of computing. Each offers a free tier or trial credit for experimenting. Prices change often, so check the provider’s own pricing page and estimate from your real volumes.
How to choose
Start with where your systems are. Moving data between clouds adds cost, delay and another set of security controls. If your applications and databases are on Azure, use Azure’s services unless there is a specific reason not to. If you have not yet chosen a cloud, our comparison of AWS and Azure covers the wider decision.
Decide which layer you need. If the task is reading documents or working with text, you need the first or second layer, and all three providers are capable. Test each candidate on a sample of your own documents and compare the results. If you need to train a model, the choice of machine learning framework matters as much as the provider.
Check where the data goes. All three run data centres in the UK, but not every service or model is available from them. Before sending anything that includes personal or confidential information, find out which region will process it and what the provider’s terms say about retaining it or using it to improve their models. Take advice on the data protection position if personal data is involved.
Plan for the provider to change. Given how often these products are renamed and retired, keep the connection to the provider in one place in your code, so that swapping one service for another is a contained change. Keep your own test examples, so a replacement can be measured against the original. This is how we approach adding AI to existing systems.
Be clear about the benefit first. The provider is a secondary question. Our article on the business benefits of machine learning covers how to tell whether a task is worth doing this way at all.
Questions to ask before you commit
- Which layer are we buying: a ready-made service, a hosted model or a training platform?
- In which region will our data be processed and stored?
- Is our data kept, and can it be used to train the provider’s models?
- What will it cost per month at our real volumes, and what happens to the cost if volumes double?
- If this service is withdrawn or its price changes, how much of our system has to change?
A supplier who has built with these services should be able to answer the first four from the provider’s documentation and the fifth from their own design.