Loading
Skip to content
Miracle Minds AI
  • Try “HIPAA”, “model risk”, “FinOps” or “agentic”.

Cloud & Infrastructure Strategy

A cloud strategy pays off when it is made workload by workload, each with a unit cost and a tested exit route, not when it starts by picking a provider. Most go wrong because AI inference changes the cost curve after the decision is made, and nobody owns the bill that follows.

Part of Cloud, Infrastructure & Applications, in the Think mode. 10 sources, checked .

Estimated share of cloud spend wasted in 2026, the first rise in five years, as AI workloads surged (Flexera survey)
29%
Estimated share of cloud spend wasted in 2026, the first rise in five years, as AI workloads surged (Flexera survey)
Flexera, 2026(opens in a new tab)
Forecast worldwide sovereign cloud IaaS spending in 2026, up 35.6% on 2025 (Gartner)
$80bn
Forecast worldwide sovereign cloud IaaS spending in 2026, up 35.6% on 2025 (Gartner)
Gartner, 2026(opens in a new tab)
Publicly reported outages over nine years attributed to third-party IT and data center providers, including cloud giants (Uptime Institute)
2 in 3
Publicly reported outages over nine years attributed to third-party IT and data center providers, including cloud giants (Uptime Institute)
Uptime Institute, 2026(opens in a new tab)

The problem

Cloud spend is now large enough, and wasteful enough, to be a board topic. Flexera's 2026 survey of more than 750 cloud decision-makers found managing cloud spend is still a top challenge for 85% of them, and that estimated wasted cloud spend rose to 29%, the first increase in five years, as cloud-based AI workloads surged.1 The FinOps Foundation's 2026 survey found 98% of practitioners now manage AI spend, up from 31% two years earlier.2 The cost problem has moved from virtual machines and storage to GPUs and tokens.

AI changes where the money goes. Gartner forecasts worldwide AI-optimized infrastructure-as-a-service spending of $42 billion in 2026, with inference ($23.3 billion) overtaking training ($19 billion) this year.3 Inference runs every time a user or agent asks something, so choices about latency, data location and GPU supply now drive the bill. Sovereignty adds a second axis: Gartner forecasts $80 billion of sovereign cloud IaaS spending in 2026 and expects it to shift 20% of current workloads from global to local providers.4

Regulators now ask how you would leave. The EU Digital Operational Resilience Act (DORA) has applied since 17 January 2025.5 Its technical standards require a documented exit plan for each contract supporting a critical or important function, reviewed and tested against unforeseen and persistent service interruptions.6 US bank guidance expects contingency plans to move an activity to another provider or back in-house,7 and in September 2026 the agencies proposed a shorter, risk-based replacement.8 Over nine years, third-party IT and data center providers, including cloud giants, accounted for about two-thirds of publicly reported outages.9

A good strategy is a placement map, not a provider choice. Each workload has a target platform, a unit cost you can track (per transaction, per claim, per inference), a regulatory classification, and an exit route with a rough cost and timeframe. The sequence follows dependencies: the data platforms, identity services and inference capacity AI needs come first, and the hardest legacy estates move when there is a business reason. The bill has a named owner from the first day.

What you get

  1. 01

    Workload placement map

    Every material workload classified by latency, data sensitivity, regulatory status and cost profile, with a target platform: public cloud, sovereign or regional cloud, private cloud, colocation or on-premises.

  2. 02

    Unit-cost model

    A cost model for each workload in business units (per transaction, per customer, per inference), with today's run rate, the target, and the commitments and GPU capacity it implies.

  3. 03

    AI-first sequencing roadmap

    The order of moves that gets data, identity and inference capacity ready for AI soonest, with dependencies, decision points and the modernization work each move needs.

  4. 04

    Exit and concentration plan

    For services that support critical or important functions, an exit route for stressed and non-stressed scenarios, concentration exposure by provider, and the contract terms that make the exit workable.

  5. 05

    Ownership and governance model

    Who owns each part of the bill, who approves placement exceptions, and how FinOps, security and architecture review new workloads before they are built.

Where it goes wrong

The failure modes we design against from the first week, and the control for each.

  • Choosing a provider before mapping workloads

    A platform chosen for a discount or a relationship forces every workload to fit it, including the ones that should stay put or need sovereign hosting. Control: make placement a per-workload decision with written criteria, and treat provider commitments as an output of the map, not an input.

  • Sizing the AI estate on training, not inference

    Inference is forecast to be the larger share of AI-optimized IaaS spend from 2026, and it grows with every user and agent step.3 Control: model inference demand per use case, at peak and steady state, before signing GPU reservations or commitment deals.

  • An exit plan that exists only on paper

    DORA's standards require exit plans that are realistic, feasible and periodically tested.6 UK firms are expected to cover both stressed and non-stressed exits.10 Control: estimate the cost and timeframe of each exit, keep data in portable formats, and test the plan for critical services.

  • A bill nobody owns

    Wasted spend rose to 29% in 2026 as AI workloads grew.1 Control: give every account, subscription and model endpoint an accountable owner, tag spend when resources are created, and report unit cost to the business owner each month, not just total spend to finance.

Standards and rules

  • EU DORA technical standards on ICT services supporting critical or important functions

    Primary text on ec.europa.eu(opens in a new tab)

    EU financial entities need a documented exit plan for each such ICT contract, reviewed and tested periodically, so every placement decision for a critical service must come with a workable way out.

  • UK PRA SS2/21 Outsourcing and third party risk management

    Primary text on bankofengland.co.uk(opens in a new tab)

    UK banks and insurers must keep tested exit strategies for stressed exits (such as provider failure) and non-stressed exits, and track concentration risk across material outsourcing; a March 2026 update, effective 18 March 2027, follows new PRA rules on incident and third-party reporting.

  • US Interagency Guidance on Third-Party Relationships: Risk Management (2023)

    Primary text on federalregister.gov(opens in a new tab)

    US banks are expected to plan how they would transition an outsourced activity to another provider or bring it in-house, and to negotiate contracts that let them change providers without undue cost.

How to measure it

  • Unit cost per business transaction or per inference for each major workload, and its trend
  • Share of cloud spend with a named owner and complete tagging
  • Workloads placed according to the map versus approved exceptions
  • Critical services with a tested exit plan, and the estimated time to exit each one
  • Commitment and GPU reservation utilization against plan

Questions buyers ask

Should we be multi-cloud?

Only where a workload has a reason for it. Running the same application across two providers doubles the skills, tooling and security work. A common pattern is one primary provider for most workloads, a second for specific services or to limit concentration, and private or sovereign capacity where regulation or data location requires it. Regulators ask for credible exit plans and concentration awareness, which you can meet without running everything everywhere.

When should we run AI inference on our own GPUs instead of a cloud API?

When volume is high and steady, when data cannot leave your environment, or when latency demands local serving. Managed APIs suit early and spiky demand because you pay only for use. Self-hosting suits sustained load, where reserved capacity and efficient serving software can cut unit cost, but you take on capacity planning, patching and model operations. Compare both on cost per successful outcome, not on price per token or per GPU hour.

Does DORA affect our cloud strategy if we are not an EU bank?

DORA applies to around 20 types of EU financial entity, including insurers and investment firms, and brings critical ICT providers under EU oversight. If you are a US or UK group with EU entities, those entities need exit plans for critical services and a register of their ICT contracts. UK PRA rules and US bank guidance set similar expectations, so designing for exit once usually covers all three.

Sources

Every figure on this page comes from the sources below. We checked each link on .

  1. 1.Flexera Finds Cloud Value is Rising While AI Waste Grows(opens in a new tab)Flexera, March 18, 2026 · flexera.com
  2. 2.State of FinOps Survey: AI Value and Skills Top Priorities as FinOps Matures Across Technology Value (98% Manage AI, 90% SaaS, 64% Licensing, 48% Data Center)(opens in a new tab)FinOps Foundation, The Linux Foundation, February 19, 2026 · linuxfoundation.org
  3. 3.Gartner Forecasts Worldwide AI-Optimized IaaS Spending to Grow 96% Through 2026(opens in a new tab)Gartner, August 10, 2026 · gartner.com
  4. 4.Gartner Says Worldwide Sovereign Cloud IaaS Spending Will Total $80 Billion in 2026(opens in a new tab)Gartner, February 9, 2026 · gartner.com
  5. 5.Digital Operational Resilience Act (DORA)(opens in a new tab)European Insurance and Occupational Pensions Authority · eiopa.europa.eu
  6. 6.Commission Delegated Regulation supplementing Regulation (EU) 2022/2554 with regard to regulatory technical standards specifying the detailed content of the policy regarding contractual arrangements on the use of ICT services supporting critical or important functions provided by ICT third-party service providers, C(2024) 1531 final(opens in a new tab)European Commission, March 13, 2024 · ec.europa.eu
  7. 7.Interagency Guidance on Third-Party Relationships: Risk Management(opens in a new tab)Board of Governors of the Federal Reserve System, FDIC and OCC (Federal Register), June 9, 2023 · federalregister.gov
  8. 8.OCC Bulletin 2026-46: Proposed Third-Party Risk Management Guidance(opens in a new tab)Office of the Comptroller of the Currency, September 11, 2026 · occ.gov
  9. 9.Uptime Announces Annual Outage Analysis Report 2026(opens in a new tab)Uptime Institute, May 13, 2026 · uptimeinstitute.com
  10. 10.SS2/21 Outsourcing and third party risk management(opens in a new tab)Prudential Regulation Authority, Bank of England, March 29, 2021 · bankofengland.co.uk