Every cloud provider now offers a layered stack of AI services: foundation-model APIs at the top, pre-trained task-specific services in the middle, and full ML platforms at the bottom. This lesson surveys the landscape and helps you map services across clouds.
Foundation Model Gateways
| Cloud | Service | Models available |
|---|---|---|
| AWS | Amazon Bedrock | Anthropic Claude, Meta Llama, Mistral, Cohere, Stability AI, Amazon Nova |
| Azure | Azure AI Foundry (incl. Azure OpenAI Service) | OpenAI GPT family, Meta Llama, Mistral, Phi (Microsoft), and 1,800+ models in the catalogue |
| Google Cloud | Vertex AI Model Garden | Google Gemini, Anthropic Claude, Meta Llama, Mistral, and 200+ open-source models |
All three follow the same pattern: a single API endpoint per cloud, pay-per-token pricing, regional data residency, enterprise security (private endpoints, customer-managed keys, no training on your data), and built-in safety filters. Choose based on which frontier models you need and which cloud your data lives in.
Pre-Built Task-Specific Services
Before reaching for a foundation model, check whether a pre-built service already solves your problem at lower cost and latency.
| Task | AWS | Azure | Google Cloud |
|---|---|---|---|
| Speech-to-text | Amazon Transcribe | Azure AI Speech | Speech-to-Text |
| Text-to-speech | Amazon Polly | Azure AI Speech | Text-to-Speech |
| Translation | Amazon Translate | Azure AI Translator | Cloud Translation |
| Text analysis (NER, sentiment) | Amazon Comprehend | Azure AI Language | Cloud Natural Language |
| Image / video analysis | Amazon Rekognition | Azure AI Vision | Cloud Vision / Video Intelligence |
| Document understanding (OCR, forms) | Amazon Textract | Azure AI Document Intelligence | Document AI |
| Conversational AI / virtual agents | Amazon Lex | Azure AI Bot Service / Copilot Studio | Dialogflow / Conversational Agents |
| Personalisation / recommendations | Amazon Personalize | (via Azure ML) | Vertex AI Search and Conversation |
| Forecasting | Amazon Forecast (now in SageMaker) | Azure ML AutoML | Vertex AI Forecast |
| Fraud detection | Amazon Fraud Detector | (via Azure ML) | (via Vertex AI) |
Full ML Platforms
For custom models — your own data, your own architecture, your own training — use a managed ML platform:
- Amazon SageMaker — notebooks, training jobs, hyperparameter tuning, model registry, real-time and batch inference, feature store, pipelines, and the SageMaker AI Studio IDE.
- Azure Machine Learning — compute instances, AutoML, designer (drag-and-drop), MLflow integration, managed endpoints, and tight integration with Azure DevOps and GitHub Actions.
- Vertex AI — Workbench notebooks, AutoML, custom training, Pipelines (Kubeflow-based), Feature Store, model monitoring, and deep integration with BigQuery.
All three offer the same conceptual building blocks — they differ mainly in pricing, ecosystem integration, and the maturity of specific features.
Build vs Fine-Tune vs API: A Decision Framework
| You should... | When... |
|---|---|
| Use an API as-is | The task is general (summarisation, classification, Q&A), latency > 1s is acceptable, and prompting + RAG meets quality requirements. |
| Use a pre-built service | The task matches a packaged service (speech, translation, OCR) and you need low latency, simple billing, and high accuracy out of the box. |
| Fine-tune a foundation model | You have hundreds to thousands of high-quality labelled examples and need a specific style, format, or domain knowledge baked in. |
| Train a custom model | You have a large proprietary dataset, the task is highly domain-specific, and inference cost or latency matters at scale. |
In practice, most teams start with an API + RAG, only fine-tune when prompting plateaus, and almost never train from scratch.
Generative AI–Specific Services
- AWS Bedrock Knowledge Bases — managed RAG over S3 documents.
- AWS Bedrock Agents — orchestrate multi-step LLM workflows with tool calls.
- Azure AI Foundry Agent Service — managed agents over your data and APIs.
- Azure AI Search — vector + keyword hybrid search, often paired with Azure OpenAI for RAG.
- Vertex AI Search — Google-grade search and RAG over your enterprise data.
- Vertex AI Agent Builder — build conversational agents and grounded chatbots.
Mapping to Certifications
| Certification | Focus areas covered above |
|---|---|
| AWS AI Practitioner (AIF-C01) | Bedrock, SageMaker, Comprehend, Rekognition, Polly, Transcribe, Translate, responsible AI on AWS |
| Azure AI Fundamentals (AI-900) | Azure AI Foundry, Azure OpenAI, AI Vision, Speech, Language, Document Intelligence, responsible AI principles |
| Google Cloud Digital Leader | Vertex AI, Gemini, Document AI, Vision AI, Speech AI, AI/ML use cases on Google Cloud |
You have now completed AI and ML Fundamentals. From here, pick the cloud closest to your work, dive into the certification-specific course, and start building. The fundamentals you have learned — supervised vs unsupervised, the ML lifecycle, transformers, prompting, RAG, responsible AI — apply on every cloud and to every model that comes next.