Deploying Agents:
AWS vs Azure vs GCP

• Most mature ecosystem
• Best documentation
• Lambda + Bedrock + DynamoDB
• Most Stack Overflow answers
• Widest model selection (Bedrock)

• Pricing complexity
• Service naming inconsistent
• IAM permissions labyrinth

• Best enterprise integration (AD)
• Exclusive OpenAI partnership
• Good compliance checkboxes
• Strong Windows support
• Enterprise agreement discounts

• Service naming confusing
• Pricing unpredictable
• Documentation scattered

• Cleanest APIs
• Fastest cold starts (100-500ms)
• Best ML infrastructure (Vertex AI)
• Native Gemini access
• Great for batch processing

• Fewer enterprise features
• Smaller community
• Less mature than AWS

• 3–5 engineers, limited budget, need to ship fast
• Using Claude as primary LLM
• Go-To: AWS (Lambda + DynamoDB + S3 + Bedrock)
• Typical cost at 10K requests/month: ≈ $120 (mostly LLM)

• Existing Azure infrastructure + Active Directory
• Strong compliance needs (SOC 2, HIPAA, etc.)
• Go-To: Azure (Functions + Cosmos DB + Azure OpenAI)
• At 1M requests/month: ≈ $8,500 (often 20% off with EA)

• Custom model fine-tuning + large-scale embeddings
• Heavy batch processing, Gemini-first strategy
• Go-To: GCP (Cloud Run + Vertex AI + Vector Search + BigQuery)
• At 500K requests/month: ≈ $5,200

• Saving >30% on $50K+/month spend
• LLM requirements force it
• Acquisition/merger (no choice)

• Marginal 10-15% savings
• "Better" features (theoretical)
• Resume building

• Phase 1 (Months 1–2): Ship new features on the new cloud only
• Phase 2 (Month 3): Move background jobs and batch workloads
• Phase 3 (Months 4–5): Migrate read-heavy endpoints (search, dashboards)
• Phase 4 (Month 6): Carefully move write endpoints with state sync
• Phase 5 (Months 7–8): Shift core agent execution, keep old cloud on standby
• Phase 6 (Month 9): Monitor, then decommission old infra

• Different products live on different clouds, with dedicated teams
• Compliance: e.g., EU data on Azure EU, US data on AWS US
• Very high LLM spend (>$100K/month) for LLM arbitrage across providers

• Need expertise in 3 platforms, 3 security models, 3 billings
• Separate monitoring, cross-cloud networking, more failure modes
• Adds roughly 20–30% operational overhead

• All three have serverless compute, managed DBs, object storage, LLMs, monitoring
• Differences are mostly in APIs, pricing models, and enterprise features
• Your constraints (infra, team, LLM, compliance, geography) should drive the choice

• Do: Pick one cloud, optimize inside it, and focus on shipping
• Don't: Migrate for marginal gains or chase clouds for resume points
• Treat cloud choice as a constraint problem, not a religion