Cloud Modernization

The cloud repatriation panic is a measurement error

The statistic your board keeps forwarding you counts a company that moved one database the same as a company that left entirely. Underneath the bad number is a real shift worth understanding.

Every few months a board member forwards me the same article. The headline says something like 83% of companies are moving workloads back out of the cloud, and the question that follows is always some version of: did we make a mistake?

It's a fair question. It's also built on a statistic that doesn't mean what the headline says it means — and untangling that is genuinely useful, because underneath the bad number there is a real and much more actionable shift happening in how modernization gets done.

The repatriation number is a measurement artifact

The figure originates in a Barclays CIO survey, and IDC has published something similar. What it actually measures is the share of organizations that have moved at least one workload out of public cloud. A company that repatriated a single chatty database scores identically to a company that exited the cloud entirely.

The number behind the number

IDC's own data puts organizations planning full-scale repatriation at 8–9%. So of the widely quoted 83–86%, roughly nine-tenths represents selective workload redistribution — which is not a retreat from cloud, it's the normal exercise of judgment about where individual workloads belong. Gartner expects 90% of organizations to be running hybrid infrastructure by 2027. Hybrid is the destination, not the failure state.

The other tell: hyperscaler revenue keeps compounding. IDC expects global public cloud spending to cross $1 trillion in 2026, up 21% year over year, with platform services growing fastest at 37%. Those two facts cannot both be true and also support a mass-exodus narrative.

There's a sharper version of this point that I've found lands well with boards, and I'll credit it where it's due — the analyst Michael Coté put it as: that which never moved can never move back. Most enterprise workload never left the data center in the first place. A large share of what gets counted as "repatriation" is a category error about systems that were never migrated.

What is actually true, and it's less comfortable

The repatriation story is wrong. The disappointment underneath it is real. It just has a different cause.

Cloud waste rose roughly 29% in 2026 — the first increase in five years, according to Flexera's long-running State of the Cloud survey. That reversal happened in a year when 63% of respondents had established FinOps teams and half were using unit economics. Discipline went up and waste went up anyway, because AI workloads arrived faster than the governance around them.

The FinOps Foundation's 2026 survey, representing more than $83 billion in annual cloud spend, contains the line that best describes where most organizations now sit: "We have hit the 'big rocks' of waste and now face a high volume of smaller opportunities." The easy savings are gone. What's left is architectural, and architecture is not a procurement exercise.

Three things I've learned doing this work

1. Lift-and-shift buys you a bill, not a capability

I spent nearly two decades inside a global energy major's downstream business, and the ~$10M in savings I'm associated with there did not come from migration. It came from decommissioning a mainframe, retiring overlapping applications, and rationalizing a portfolio against actual business capability rather than org-chart ownership. Moving those same systems to someone else's data center unmodified would have changed the invoice and nothing else.

McKinsey has been making a version of this argument for years: the cloud prize is captured by companies that change their operating model, not by companies that change their hosting. Adoption is not value. It is the precondition for value, and the two get conflated constantly in board reporting.

2. The mainframe question has quietly flipped

If you're carrying a mainframe or a comparable legacy core, the honest 2026 answer is probably not "migrate off it." BMC's twentieth annual mainframe survey found positive perception at an all-time high of 97%, with 72% of respondents reporting increased capacity — growth driven substantially by new applications. Two-thirds of respondents are now millennial or Gen Z, which takes a good deal of air out of the skills-cliff argument.

BMC sells mainframe tooling, so read the enthusiasm accordingly. But the directional point survives the discount: the platform is not dying, which means the business case for a big-bang migration off it is weaker than the pitch deck implies. Modernize the interfaces, expose the data, leave the transaction engine where it earns its keep.

3. The genuine repatriation cases are workload-economics cases

The best-documented exit is 37signals: roughly $2M saved in the first year, around $10M projected over five. Real numbers, real company. Also an unusually favorable case — a stable SaaS with flat, predictable load, high egress, and a strong in-house operations team.

If your workload is steady-state, egress-heavy, and you have the platform staff to run it, the arithmetic may well favor owning the metal. That is a decision about one workload's economics. It is not a cloud strategy, and it does not generalize to a mid-market operator with variable demand and two infrastructure people.

What "modernization" should mean in 2026

The framing I use with clients is that modernization is a portfolio decision, not a platform decision. You are not choosing cloud or not-cloud. You are deciding, application by application, which of five things is true: this earns a real investment, this gets migrated as-is because it's fine, this gets its data exposed but its guts left alone, this gets retired, or this gets tolerated for now with a documented reason and a review date.

That last category matters more than people expect. Half the value of a portfolio review is converting undecided systems into deliberately deferred ones, with a named owner and a date. An undecided system generates anxiety in every planning cycle forever. A deferred one doesn't.

What to do about it

  1. Stop asking "should we repatriate?" and start asking "which workloads have the wrong economics?" The first question has no answer. The second one has five or six, and you can compute them.
  2. Put a unit-economics number on your top ten workloads. Cost per transaction, per booking, per order — whatever your business actually counts. Half of FinOps teams now do this. Without it, every cost conversation is a debate about invoices instead of value.
  3. Separate the AI line item from everything else, now. The waste reversal is being driven by AI workloads landing inside cloud budgets that were never structured to hold them. If AI spend isn't visible on its own, your cloud trend line has stopped being meaningful.
  4. Treat your legacy core as an integration problem before a migration problem. Expose the data, put an API layer in front of it, and revisit the replatforming question when you have evidence rather than a vendor's timeline.
  5. Make deferral explicit. Every application gets an owner, a disposition, and a review date — including the ones you're consciously leaving alone. Ambiguity is what makes portfolios expensive.

None of this is exciting, which is roughly the point. The organizations getting value out of cloud in 2026 are not the ones with the boldest migration strategy. They're the ones who did the boring portfolio work first and then had somewhere sensible to put their AI workloads when those arrived.

Sources

  1. Channelnomics, Breaking down the 83% public cloud repatriation number
  2. IDC via BizTechReports, Global public cloud spending to surpass $1 trillion in 2026 (March 2026)
  3. Flexera, 2026 State of the Cloud Report (n=753 cloud decision-makers)
  4. FinOps Foundation, State of FinOps 2026 (n=1,192; $83B+ annual cloud spend represented)
  5. BMC, 20th Annual Mainframe Survey (September 2025)
  6. DataCenterDynamics, 37signals on its repatriation savings
  7. Michael Coté, That which never moved can never move back
  8. McKinsey, Cloud insights

Kurt Wysock

Fractional & interim CTO and board technology advisor. Currently interim CTO at a luxury destination-club operator, head of the Technology & Architecture Office at a global consulting firm, and founder of JSummit Consulting. 30+ years in enterprise architecture across energy, hospitality, manufacturing, financial services, retail, and healthcare. TOGAF® 9 Certified. More about Kurt · Get in touch

Your AI problem is a data problem →

Want this applied to your systems?

The Production AI Blueprint is a fixed-fee engagement that sorts out your data, governance, controls, and ownership — and hands you a sequenced roadmap you own outright.