The AWS Generative AI Stack
Amazon Bedrock, SageMaker AI, Amazon Q, Kiro, and the agentic toolset: what each service does, why build on AWS, and the cost tradeoffs.
The previous topics explained how generative AI works and how to judge whether it is worth using. This one is about the machinery: which AWS service you actually call, what it removes from your work, and what it charges you for. Task Statement 2.3 names the services directly, so this topic follows that list and draws the boundaries the exam probes between them.
The organizing idea is a single axis running from fully managed to fully controlled. Amazon Bedrock sits at one end, serving foundation models through one API. Amazon SageMaker AI sits at the other, letting you train and serve anything on infrastructure you size. SageMaker JumpStart sits between them. Above all three are the finished assistants and agent tooling: Amazon Q, Amazon Quick, Kiro, Strands Agents, and Amazon Bedrock AgentCore.
What This Topic Covers
- Amazon Bedrock as the managed foundation model API, plus Knowledge Bases, Guardrails, evaluations, and model customization
- how Bedrock handles customer prompts and completions with respect to model providers
- Amazon SageMaker AI and SageMaker JumpStart, and the requirements that push a workload from Bedrock to them
- Amazon Q and Amazon Quick as finished assistants, and who each one serves
- Kiro and spec-driven development, with specs, steering, and hooks
- Strands Agents for writing an agent and Amazon Bedrock AgentCore for running it
- the advantages of AWS generative AI services and the security, compliance, responsibility, and safety benefits of AWS infrastructure
- inference purchasing modes, model units and commitment terms, the true cost of a custom model, and how responsiveness, redundancy, and Regional coverage trade against price
Why It Matters
Service-selection questions are the most mechanical points on this part of the exam, and also the easiest to lose. The names are similar, several products were renamed recently, and a scenario usually hides one constraint that decides the answer. Reading for that constraint is a skill, and it is the same skill you use at work when someone asks whether a project belongs on Bedrock or SageMaker AI.
The cost material pays off even faster. Knowing that batch inference costs half of on-demand, or that a fine-tuned model cannot be served without Provisioned Throughput, changes architecture decisions before they become invoices.