▲ Claude Academy
Sign in

Vault note

Track 7 — Deployment & Scale

You can build with Claude; now run it in production, on a cloud, under governance. This track covers accessing Claude through Amazon Bedrock and Google Cloud's Vertex AI — the enterprise surfaces where models live inside your existing cloud account, IAM, networking, logging, and billing — plus a review of the cost-and-throughput levers (prompt caching, the Batch API) that keep scaled workloads affordable and reliable.

Audience: developers, platform, and ops. · Est. time: ~6–8 hours. · Prereq: track-6-agents-skills-subagents and cloud access (setup-checklist, section d — AWS/Bedrock or GCP/Vertex). · Outcome: deploy Claude on a cloud with governance.

Sequence

Do the platform you (or your org) actually use. If you're cloud-agnostic, do Bedrock first; the concepts transfer.

1. Claude on Amazon Bedrock · ⏱ ~2.5–3 h

  • Course: claude-with-amazon-bedrock — enabling Claude models in Bedrock, region/model availability, calling Claude via the Bedrock SDK (and how it maps to the Messages API you already know), IAM roles vs. static keys, logging/observability, and cost controls.

2. Claude on Google Cloud Vertex AI · ⏱ ~2.5–3 h

  • Course: claude-with-vertex-ai — enabling Claude in Model Garden, the Vertex SDK, gcloud / application-default auth, service accounts, regions, and Vertex logging and quotas.

3. Batch & caching review — scale economics · ⏱ ~1–2 h

  • Review: revisit the prompt-caching and Batch API portions of building-with-the-claude-api and your p08-prompt-caching-and-rag project, now through a production lens — when batching wins, how caching cuts cost/latency at scale, and how this interacts with cloud billing and rate limits.

Governance checklist (carry this into any deployment)

  • Auth: prefer IAM roles / service accounts over long-lived keys; least privilege.
  • Secrets: keys in a secrets manager, never in code or git (see setup-checklist security note).
  • Observability: request/response logging, error tracking, latency and token dashboards.
  • Cost: spend limits, model-tier discipline (Haiku/Sonnet where they suffice), caching + batching for volume.
  • Reliability: retries/backoff, timeouts, graceful degradation, region/quotas planning.
  • Data: know your platform's data-handling and residency guarantees before sending regulated data.

You're ready to move on when…

  • You can call Claude through either Bedrock or Vertex and map it back to the Messages API.
  • You can explain auth on your chosen platform (IAM role / service account) and why it beats static keys.
  • You can name the cost/throughput levers (prompt caching, Batch API, model tiering) and when each applies.
  • You can list the governance concerns above for a real deployment.

See also