3.6.1. Generative AI Integration and Multi-Account Amazon Bedrock Patterns
💡 First Principle: Model access is a shared platform capability. Decide where calls originate, which account's quotas and bill they consume, and where the governance controls sit, before any team writes a prompt.
Scenario: A company with 40 workload accounts in AWS Organizations wants every team to use generative AI. Only three approved models may be used, all traffic must stay off the internet, and each team must stay within its own throughput quota.
How applications call Bedrock. "Amazon Bedrock" is a serverless API to foundation models from Amazon and third-party providers. Applications call the bedrock-runtime endpoint with InvokeModel or the model-agnostic Converse API, which uses one request format across models so a model can be swapped without code changes. There are no servers or GPUs to manage, and on-demand use is billed per input and output token. Bedrock does not use prompts or completions to train models and does not share them with model providers.
Model access is on by default, so governance is a deny. In commercial Regions most serverless models are enabled on first invocation: Bedrock starts the AWS Marketplace subscription automatically (the caller needs aws-marketplace:Subscribe once), and Anthropic models need a one-time use-case form that the management account can submit for the whole organization. There is no per-account "request access" step to withhold. To restrict models, attach an SCP that denies bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream (and bedrock:* where appropriate) on every foundation-model ARN outside the approved list. Denying aws-marketplace:Subscribe alone does not stop the first invocation.
Private connectivity. Interface VPC endpoints exist for bedrock-runtime, bedrock, bedrock-agent and bedrock-agent-runtime. Add an endpoint policy, and a Deny unless aws:SourceVpce matches, so a call cannot silently take a public path.
| Pattern | How it works | Choose it when | Watch for |
|---|---|---|---|
| Distributed access | Each workload account calls Bedrock directly. SCPs allow-list models, an Organizations Bedrock policy enforces a guardrail, and logging is deployed to every account with StackSets | Teams own their applications and you want per-account quota and cost isolation with the least infrastructure | Invocation logging and guardrail settings are per account and per Region, so they must be deployed everywhere |
| Centralized GenAI gateway | A platform account exposes one internal API ("API Gateway" and "Lambda", or "AgentCore Gateway" model routing) that calls Bedrock; workload accounts reach it privately | You need one place for model routing, per-tenant rate limits, prompt logging or provider API keys, or most accounts must not call models at all | All traffic consumes that one account's Bedrock quotas, and the gateway becomes a critical service you operate |
Quotas follow the account. On-demand quotas (requests and tokens per minute) apply per account per Region, so the distributed pattern spreads load naturally while a central gateway needs quota increases or cross-Region inference profiles for extra throughput. Shared building blocks still work across accounts: a guardrail can be shared with the organization through a resource-based policy, and application inference profiles give each team a taggable handle for cost attribution (see 3.6.7).
⚠️ Exam Trap: "Turn on model access only in the accounts that need it" is not a control in commercial Regions, because access starts on first use. The answer that actually blocks unapproved models is an explicit deny in an SCP (or IAM policy) on the model ARNs.