Logiciel Solutions Contact Us
Success Stories Tech News Investors Contact Us
whitepaper
8 min read

Cloud Migration Playbook for AI Workloads.

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.

In depth

Why AI Workloads Change The Way A Cloud Migration Has To Be Planned.

01

AI changes what the destination cloud is for.

The basic cloud migration process is familiar: assess, choose a strategy per workload, build a landing zone, move, then optimize. What has changed is the job the target platform does once the move is over. In Flexera's 2026 survey, 81% of respondents said they use generative AI, up from 47% in 2024, so a database that once served monthly reports may now feed agents and real-time retrieval. GPU capacity is reserved and rationed by region, inference runs on every prompt and agent step, and training sets and vector indexes are too large to move twice. Model endpoints, prompt logs and retrieved context also create new places where regulated data can be exposed. Retrofitting GPU placement, egress controls and model-access policies after cutover costs more than designing for them up front.

In shortRetrofitting GPU placement, egress controls and mode…
02

Replatforming the databases that feed models is usually the highest-value move.

AWS Prescriptive Guidance defines seven migration strategies (retire, retain, rehost, relocate, repurchase, replatform and refactor) and recommends against refactoring during large migrations, because modernizing while migrating adds complexity. Microsoft's Cloud Adoption Framework and Google Cloud describe the same journey under different names, so one crosswalk keeps multi-cloud teams aligned. For an AI program, the useful question is which strategy leaves each workload's data and interfaces ready for models with the least rework. Moving systems of record to managed engines such as Amazon RDS, Azure SQL, Cloud SQL or AlloyDB gives models stable endpoints, native backups and IAM-based access without a full rewrite, while rehosted databases rarely expose clean APIs for agents. Replatform what feeds models, rehost the rest, and refactor AI-facing applications in a later wave.

In shortrefactor AI-facing applications in a later wave
03

AI is now driving cloud waste as well as cloud growth.

Gartner expects AI-optimized IaaS spending to reach $42.3 billion in 2026, up about 96% from 2025, with inference at $23.3 billion overtaking training at $19 billion for the first time. Spend control has not kept pace. Managing cloud spend is the top challenge for 85% of Flexera respondents, and wasted spend rose to 29% in 2026, the first increase in five years. The FinOps Foundation reports that 98% of practitioners now manage AI spend, up from 31% two years earlier, and recommends getting visibility and allocation in place before optimizing. The costliest AI pattern is a model in one cloud reading a database in another on every request. Cost controls built into the landing zone are cheaper than any optimization added after cutover.

In shortCost controls built into the landing zone are cheape…
04

Compliance has to be mapped before regulated data moves.

Security and compliance is the top challenge for cloud AI initiatives, cited by 53% of cloud leaders in Flexera's 2026 survey, ahead of data quality for model training at 40%. Under HHS guidance, a cloud provider that creates, receives, maintains or transmits ePHI is a HIPAA business associate even when it holds only encrypted data and no key, and a signed BAA must come before use. Federal workloads need FedRAMP-authorized services, and GSA reported a record 114 authorizations in fiscal year 2025. Every hosted model endpoint in scope needs the same check as the database behind it. A late compliance review lands regulated data in services that were never cleared to hold it.

In shortA late compliance review lands regulated data in ser…
The detail

Three Migration Decisions That Set AI Performance And Cost.

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.

Zone · 01

Database cutover and rollback

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.

Zone · 02

GPU capacity by workload

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.

Zone · 03

Data placement and egress

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.

By the numbers

The figures that make it a board-level conversation.

29%
of cloud spend wasted in 2026, the first increase in five years, which Flexera ties to surging cloud AI workloads
$42.3 billion
forecast AI-optimized IaaS spending in 2026, with inference overtaking training for the first time
98%
of FinOps practitioners now manage AI spend, up from 31% two years earlier
Inside the report

What you'll take away.

01

Step 1 - Assign a 7 Rs strategy and an owner to every workload

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.

02

Step 2 - Build the landing zone with tagging, GPU quotas and compliance mapped

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.

03

Step 3 - Migrate in waves, moving data and its AI consumers together

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.

04

Step 4 - Track cost per outcome for each AI use case

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.

Questions

Frequently asked.

What is cloud migration, and what changes when the target is AI?

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.

What are the 7 Rs of cloud migration?

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.

How do you migrate a database to AWS, Azure or Google Cloud?

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.

How do you plan GPU capacity when migrating AI workloads to the cloud?

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.

How do egress fees and data gravity affect cloud migration costs?

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.

What compliance requirements apply to cloud migration in the US?

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.

What does the playbook include, and what evidence is it based on?

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.

Get the whitepaper

Have it emailed to you.

Drop your details and we'll send Cloud Migration Playbook for AI Workloads straight to your inbox - no spam, unsubscribe anytime.

Download whitepaper
Next step

Plan the migration around the AI workloads it has to carry.

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