"You rarely build the whole stack yourself. Cloud platforms offer ready-made building blocks — from raw GPUs to pre-trained AI — so you focus on your product."
Level: Intermediate · Time: ~11 min · Prerequisites: Module 39
Learning Objectives
By the end of this module, you will be able to:
- Explain what cloud AI services offer
- Tell "build your own" from "ready-made" AI
- Recognize the major providers
- Choose the right level of service
1. The Cloud AI Spectrum
Cloud providers offer AI at every level, from do-it-yourself to done-for-you:
| Level | You get | Example use |
|---|---|---|
| Compute | Rented GPUs/machines | Train your own model |
| ML platforms | Managed training + serving | End-to-end projects |
| AutoML | Automated model building | Fast baselines |
| Pre-trained AI APIs | Ready models (vision, speech, language) | Add AI in one call |
Explain like I'm new: It's like building a kitchen. You can buy raw land and build from scratch (compute), rent a fitted kitchen (ML platform), or just order takeout (pre-trained API). Pick the level that matches how custom your needs really are.
Concept: Moving up the spectrum trades control for speed. Raw compute gives you total control but you build everything; a pre-trained API gives you almost none but ships in an afternoon. Most real products mix levels — a pre-trained speech API here, a small custom model there — choosing per feature rather than committing the whole system to one rung.
2. Build vs. Use
- Build your own when your problem is unique and you have good data.
- Use a pre-trained service when your need is common (transcribe speech, detect objects, translate) and speed matters.
A quick decision test: ask "is this problem solved elsewhere, or is it my problem?" Transcribing English speech, detecting faces, translating between major languages, extracting text from receipts — these are solved, and a provider has trained on far more data than you'll ever gather. But predicting your customers' churn from your product's usage patterns is uniquely yours; no vendor's off-the-shelf model has seen your data. That's when building earns its cost. A good rule of thumb: if you can describe the need in one generic sentence a stranger would understand — "turn speech into text" — a service probably exists; if explaining it requires details only your business knows, you're likely building.
Key idea: Don't train a model for a solved problem. If a pre-trained API does it well, use it — you'll ship in hours, not months. Reserve custom training for genuinely unique needs where off-the-shelf falls short.
3. The Major Providers
| Provider | ML platform | AI APIs |
|---|---|---|
| Microsoft Azure | Azure ML | Azure AI Services |
| AWS | SageMaker | Bedrock, Rekognition… |
| Google Cloud | Vertex AI | Vision, Speech, Gemini… |
They also provide managed serving, deployment, and monitoring — much of Modules 42–46, as services.
Real-world use case: A startup needs to transcribe support calls. Instead of collecting audio data and training a speech model (months), they call a cloud speech-to-text API (minutes) and spend their real effort on the product. Later, if they need something custom, they can move to a platform.
4. Trade-offs
Managed services save enormous effort but bring cost (pay per use — watch it), some vendor lock-in, and less control. Weigh convenience against these, and design so you could switch if needed. The practical defense against lock-in is a thin abstraction layer: wrap the provider's call behind your own small interface, so swapping "Provider A's speech API" for "Provider B's" is a one-file change, not a rewrite scattered across your codebase.
Real-world use case: A team prototypes on a cloud vision API and grows to millions of calls a month. At that volume the per-call bill exceeds the cost of hosting an open model themselves, so they switch — and because every call went through one wrapper function, the migration takes an afternoon, not a sprint.
Common mistake: Ignoring cost until the bill arrives. Managed and per-call services are convenient but can get expensive at scale. Estimate costs early and monitor them (Module 49).
✅ Checkpoint
- Name the levels of the cloud AI spectrum.
- When should you use a pre-trained API vs. build your own?
- Name one trade-off of managed services.
Answers: 1) Compute, ML platforms, AutoML, pre-trained AI APIs. 2) Pre-trained for common/solved needs and speed; build your own for unique problems with good data. 3) Cost, vendor lock-in, or less control.
Key Takeaway: Cloud AI spans a spectrum — raw compute, managed ML platforms, AutoML, and pre-trained AI APIs — from Azure, AWS, and Google. Use pre-trained services for common/solved problems (ship fast); build your own only for unique needs. Weigh the trade-offs: cost, lock-in, and control — and watch spending early.
Further Learning
Part of "Zero to AI Engineer." Simplified from the AI Engineer curriculum.