Skip to content

Task Statement 2.3: Describe AWS Infrastructure and Technologies for Building GenAI Applications

AWS AIF-C01 Exam Focus: Task 2.3 evaluates your knowledge of the AWS service ecosystem and hardware infrastructure used to deploy Generative AI applications. You must identify key AWS GenAI services (Amazon Bedrock, SageMaker AI, SageMaker JumpStart, AgentCore, Strands Agents, Kiro), understand the strategic advantages of AWS managed services, recognize AWS security and data privacy guarantees, and evaluate trade-offs regarding cost, latency, availability, and hardware silicon (Trainium vs. Inferentia).

1. AWS Services & Features for Building GenAI Applications

AWS provides a layered stack for building Generative AI solutions, ranging from serverless managed APIs to custom hardware infrastructure.

Managed Generative AI & Agentic Services

  • Amazon Bedrock: A fully managed, serverless service that offers access to high-performing Foundation Models (FMs) from leading AI startups and Amazon via a single unified API.
    • Bedrock Knowledge Bases: Fully managed Retrieval-Augmented Generation (RAG) that automates data parsing, chunking, embedding generation, and vector database storage.
    • Agents for Amazon Bedrock: Managed agentic service that orchestrates multi-step tasks, maintains short-term memory, and executes API calls via AWS Lambda.
    • Guardrails for Amazon Bedrock: Configurable safety controls that filter harmful content, block PII, reject off-topic subjects, and prevent prompt injection attacks.
  • Amazon Bedrock AgentCore & Strands Agents: Managed runtime frameworks for building, executing, securing, and orchestrating complex multi-agent workflows with state management and enterprise identity integration.
  • SageMaker JumpStart: An ML hub within Amazon SageMaker AI that provides access to pre-built algorithms and open-weight foundation models (e.g., Meta Llama, Mistral) for one-click deployment and fine-tuning.

Machine Learning Platform & Developer Tools

  • Amazon SageMaker AI: A comprehensive platform for developers and data scientists to build, train, fine-tune, host, and monitor custom machine learning models and foundation models at scale.
  • Kiro, Amazon Q & Developer Assistants: Developer-focused AI tools that accelerate software delivery, automate code generation, generate test cases, and assist with infrastructure architecture.

Purpose-Built AWS Hardware Silicon

  • AWS Trainium: Purpose-built AWS silicon accelerator designed specifically for deep learning model training and pre-training at scale. Delivers up to 50% cost-to-train savings compared to equivalent GPU instances.
  • AWS Inferentia: Purpose-built AWS silicon accelerator designed specifically for deep learning model inference. Delivers high throughput and lowest latency at the lowest cost for deploying trained models in production.

2. Strategic Advantages of AWS GenAI Managed Services

Building GenAI applications on managed AWS services (e.g., Amazon Bedrock) provides key business and operational benefits:

  • Lower Barrier to Entry: Developers can integrate generative capabilities into existing software via simple REST API calls without needing machine learning PhDs or deep expertise in GPU cluster management.
  • Faster Speed to Market: Eliminates months of infrastructure setup. Teams can build complete functional prototypes and production RAG systems in days using managed features like Bedrock Knowledge Bases and Agents.
  • Operational Efficiency: Serverless architectures eliminate the operational overhead of server patching, hardware capacity planning, and scaling configuration.
  • Cost Optimization: Token-based pay-as-you-go pricing allows organizations to scale costs directly with user adoption, avoiding large upfront hardware investments.
  • Seamless Ecosystem Integration: Out-of-the-box integration with core AWS enterprise infrastructure (Amazon S3, AWS Lambda, Amazon OpenSearch, AWS KMS, AWS IAM, Amazon CloudWatch).

3. Security, Compliance, and Data Protection Standards

AWS enforces strict data privacy, security, and governance standards across all GenAI services.

Core Data Privacy Guarantees in Amazon Bedrock

  • No Model Training on Customer Data: Customer prompts, generated responses, and uploaded documents are NEVER used to train or improve the base foundation models provided by AWS or third-party model vendors (e.g., Anthropic, Meta).
  • Data Isolation: Customer data remains entirely within the customer's tenant and selected AWS region. Data is never shared with third-party model providers.

Infrastructure & Network Security

  • Encryption at Rest & In Transit: Data is encrypted in transit via TLS 1.2+ and at rest using keys managed through AWS Key Management Service (AWS KMS).
  • Private Connectivity via AWS PrivateLink: Applications inside an Amazon VPC can communicate with Amazon Bedrock privately without exposing traffic to the public internet.
  • Granular Access Control: Access to specific models, Bedrock Knowledge Bases, and Agent APIs is governed by AWS Identity and Access Management (IAM) policies and roles.
  • Compliance Frameworks: AWS GenAI services are HIPAA eligible and compliant with SOC 1/2/3, ISO, and PCI-DSS standards.

4. Cost, Latency, and Infrastructure Trade-Offs

When architecting GenAI applications on AWS, you must evaluate several key operational trade-offs:

1. Consumption Model Trade-Offs (On-Demand vs. Provisioned Throughput)

  • On-Demand Pricing:
    • Mechanics: Pay purely per 1,000 input/output tokens processed.
    • Trade-off: Highly cost-effective for variable or low-volume workloads, but subject to default API rate limits during peak regional usage.
  • Provisioned Throughput:
    • Mechanics: Commit to fixed Model Units (MUs) for 1-month or 6-month terms.
    • Trade-off: Guarantees dedicated compute capacity, consistent latency, and higher throughput SLAs. Higher fixed cost; required for deploying custom fine-tuned models in Bedrock.

2. Regional Coverage vs. Latency

  • Service & Model Availability: Not every foundation model or Bedrock feature is available in every AWS Region.
  • Cross-Region Latency: Invoking a model hosted in a distant AWS Region to access a specific FM introduces higher network latency and cross-region data transfer fees compared to hosting in the local region.

3. Hardware Silicon Selection (Trainium vs. Inferentia vs. GPUs)

  • Training Phase: Use AWS Trainium (EC2 trn1 instances) to lower the cost of pre-training or large-scale fine-tuning compared to traditional NVIDIA GPUs.
  • Inference Phase: Use AWS Inferentia (EC2 inf2 instances) to achieve the lowest cost-per-inference and low-latency token generation for self-hosted models on SageMaker.

4. Managed API (Bedrock) vs. Self-Hosted (SageMaker Endpoints)

  • Choose Managed API (Bedrock) for lowest operational effort, zero server management, and token pricing.
  • Choose Self-Hosted (SageMaker) when you require deep access to open-source model weights, custom container runtimes, non-standard model architectures, or predictable instance-based pricing at continuous high volume.

5. Exam Decision Rules & Common Pitfalls

High-Yield Decision Rules

  1. Select AWS Trainium when a scenario asks for the most cost-effective AWS hardware to train or fine-tune a deep learning model.
  2. Select AWS Inferentia when a scenario asks for the most cost-effective, low-latency AWS hardware to host and serve inference for a deep learning model.
  3. Select Amazon Bedrock when the requirement is serverless API access to top-tier FMs with zero infrastructure management.
  4. Select Provisioned Throughput in Bedrock when an enterprise requires guaranteed capacity SLAs, no rate-limit throttling, or hosting a custom fine-tuned model.
  5. Select AWS PrivateLink when network compliance mandates that GenAI API traffic must remain strictly inside the private VPC without traversing the public internet.
  6. Select SageMaker JumpStart when you need a pre-built hub to test, fine-tune, and deploy open-weight foundation models directly inside your SageMaker environment.

Common Exam Traps

  • Trap 1: Confusing Trainium and Inferentia
    • Remember: Trainium is for Training models. Inferentia is for Inference (serving) models.
  • Trap 2: The Data Privacy Myth
    • Watch out for distractor options claiming that Amazon Bedrock exposes customer data to third parties or uses customer prompts to train base models. AWS explicitly guarantees that prompts are NEVER used to train base models.
  • Trap 3: On-Demand for Fine-Tuned Bedrock Models
    • You CANNOT run custom fine-tuned models in Amazon Bedrock using standard On-Demand token pricing. Custom fine-tuned models require Provisioned Throughput.