Most cloud migration programs were designed to cut data center costs, and AI workloads break that brief. A model that runs inference on every prompt, reads a vector index or needs reserved GPU capacity has different placement, cost and security needs than a lift-and-shift web app.
Flexera's 2026 survey found wasted cloud spend rising to 29%, the first increase in five years, and tied the rise to cloud AI workloads. This playbook maps the AWS 7 Rs to the Microsoft and Google frameworks, compares the three managed database migration services, and covers GPU capacity, egress, hybrid integration, FinOps and US compliance.
Its evidence comes from provider documentation, analyst forecasts, industry surveys and US government guidance current as of November 2026. It does not cover pipeline design or vector store architecture, and it reports no results from any single migration program.
Most migration failures trace to skipped assessment, unclear ownership or cost surprises. For AI programs, three decisions carry most of that risk, and each one is best settled before the first wave moves rather than after it. Each produces one named artifact that a program lead can check.
AWS DMS, Azure Database Migration Service and Google Cloud Database Migration Service all run an initial full load followed by continuous replication, so the source stays live until cutover. Cross-engine moves, such as Oracle or SQL Server to PostgreSQL, need schema conversion first. Run change data capture for a full business cycle, validate counts and checksums, and move applications before AI consumers. The artifact is a tested rollback plan for every move group.
Training and fine-tuning are bursty and suit time-boxed reservations, such as AWS EC2 Capacity Blocks for ML, which run for up to six months and terminate running instances when the block ends. Inference is steady and grows with adoption. Estimate requests per second for each use case, convert them to GPU-hours from benchmark throughput, and check regional quotas before choosing a region. The artifact is a capacity plan with approved quotas per region.
Large datasets pull compute toward them, and providers charge for data leaving their network. Google Cloud's free transfer waiver applies only to a full exit completed within 60 days, not to routine multi-cloud traffic. Co-locate inference, retrieval indexes and data replicas in one region and provider, and use private interconnects for steady hybrid traffic. The artifact is a data flow map with egress budgeted for every path.
Inventory applications, databases and dependencies, classify data and flag PHI, PCI and CUI, and record which models, retrieval pipelines or agents will read each source. Add a manual pass for batch jobs and file shares.
Enforce tags for application, owner, environment and AI use case, set budgets and anomaly alerts on GPU and model API spend, and sign BAAs or confirm FedRAMP and SOC 2 scope before data moves.
Benchmark storage latency against captured p99 numbers, run replication alongside production, and co-locate inference endpoints with their data. Each wave passes validation and rollback tests before the next one starts.
After cutover, right-size, commit only once usage is stable, refactor AI-facing applications in a separate wave, and report unit metrics such as cost per 1,000 inferences or cost per resolved ticket.
Cloud migration is moving applications, databases and infrastructure from on-premises data centers, or from another cloud, onto a platform such as AWS, Microsoft Azure or Google Cloud. When the goal is AI, the migration must also place data near GPU capacity, control egress and inference costs, and meet security rules for model access to sensitive data.
Those needs are cheaper to design in before cutover than to retrofit after it.
The 7 Rs are AWS's seven migration strategies: retire, retain, rehost, relocate, repurchase, replatform and refactor. Each workload is assigned one. AWS recommends against refactoring during large migrations because it adds complexity.
For AI programs, replatforming databases to managed engines is often the best balance, since it gives models stable endpoints, native backups and IAM-based access without a full rewrite.
Use the provider's managed database migration service. AWS DMS, Azure Database Migration Service and Google Cloud Database Migration Service all run a full load followed by continuous replication, so the source stays live until cutover.
Cross-engine moves need schema conversion first. Run replication for a full business cycle, validate counts and checksums, test the rollback plan, then cut over applications before AI consumers.
Separate training from inference. Reserve training capacity in time-boxed blocks, such as AWS EC2 Capacity Blocks for ML, which run for up to six months. Size inference from expected requests per second and benchmark throughput on candidate instances.
Check quotas and regional availability before choosing a target region, and commit to inference capacity only after usage stabilizes.
They make the location of data a cost decision. Large datasets pull compute toward them, and providers charge for data leaving their network, so a model in one cloud reading a database in another pays on every request.
Co-locate models, retrieval indexes and data replicas in one region, replicate read-only copies instead of querying across clouds, and budget egress explicitly in the migration business case.
HIPAA requires a signed business associate agreement before a cloud provider handles ePHI, even if the provider holds only encrypted data and no key. Federal workloads need FedRAMP-authorized services at the right impact level.
Migration also changes the control environment behind a SOC 2 report. Many organizations use the NIST AI Risk Management Framework to structure risk reviews for models deployed after migration.
It includes a 7 Rs guide for AI programs, a crosswalk of the AWS, Microsoft and Google frameworks, a comparison of the three managed database migration services, a four-phase roadmap, a 14-step checklist and an eight-dimension readiness scorecard.
Its facts come from provider documentation, Gartner forecasts, Flexera, FinOps Foundation and McKinsey surveys, and HHS, GSA and NIST guidance. It reports no results from a specific migration.
Drop your details and we'll send Cloud Migration Playbook for AI Workloads straight to your inbox - no spam, unsubscribe anytime.
Bring your application inventory and the AI use cases your migration has to support. We will work through 7 Rs decisions, database cutover, GPU placement and cost allocation with your cloud and data teams. SECTION 7 - FAQ - 5 to 8 questions
Book a migration review