Repatriation is the infrastructure story of 2026, and for once the hype is earned. Somewhere between 80 and 86 percent of CIOs now plan to pull at least some workloads back off public cloud, according to the Barclays Q4 2024 CIO Survey and IDC. Only about 8 percent plan a full exit. Flexera's 2025 data shows roughly 21 percent of workloads that once moved to cloud have already come home, even as net-new cloud migration keeps going.

So this isn't an exodus. It's a rebalancing. Steady, high-utilization, predictable workloads win on dedicated infrastructure. Variable and cloud-native workloads stay in cloud. Right workload, right environment.

The math behind the move is real too. Three things converged. Broadcom's VMware licensing shift raised the minimum license count to 72 cores in April 2025, up from 16, and killed perpetual licenses. Smaller shops have reported cost jumps from 350 percent to over 1,000 percent. License an 8-core edge server and you now pay for 64 cores that don't exist. Meanwhile the EU's Digital Operational Resilience Act (DORA) became enforceable in January 2025 with no grace period, and in November the regulators named the first 19 critical ICT providers under direct supervision, including the three largest public clouds. And every hosting vendor with a landing page is quoting 30 to 70 percent savings, some as high as 82 percent.

Here's the part the savings slide leaves out.

The number and the failure mode come from the same missing thing

Every one of those savings figures is a compute comparison. Hardware versus cloud bill. None of them price in the people who keep the hardware running.

That omission is not small. Ponemon puts average patch-management staffing alone at about $1.1 million per year for organizations at scale. Infrastructure engineers run $75,000 to $100,000 each. Automox estimates on-prem patch management can reach three to five times the software list price once you count staffing, licensing complexity, and maintenance. Manual patching by itself can burn 2,400 staff-hours a year. Then add the redundancy and disaster-recovery spend to hit even a 99.5 percent availability target, plus the on-call rotation that answers the pager at 3 a.m.

A fair total-cost comparison has four buckets: infrastructure, people, risk, and compliance. Vendor pitches show you the first bucket and stop.

And this is where it gets interesting. The savings and the risk trace back to the exact same variable. Repatriation banks the 30 to 70 percent only if you already own the operational capability to run the metal. Skip that, and you don't just miss the savings. You inherit the failure mode. The documented ways repatriation goes wrong are consistent: understaffing, undersized hardware because cloud autoscaling masked your true consumption, and hidden cloud-native dependencies like managed databases and queues that don't exist once you leave. As one analysis put it plainly, a team that repatriates without scaling its operations capability can end up worse off than the cloud setup it left.

The success everyone quotes proves the point

The repatriation case study everyone cites is 37signals. It's a good one, and it says the opposite of what most people think it says.

They cut their cloud bill from $3.2 million a year to under $1 million, with projected savings over $10 million across five years. In the final phase they spent $1.5 million on 18 petabytes of storage that costs under $200,000 a year to run, replacing a roughly $1.5 million annual bill for the same in cloud.

Now read the fine print. Their Director of Operations was emphatic: they did not hire a single new person. Same 10-person operations team that ran it in the cloud runs it on their own hardware today. That line gets quoted as proof that anyone can do this cheaply.

But two facts change the story. First, they started with a mature 10-person ops team. Second, they never touch a physical server. The hardware is racked, powered, and maintained by the on-site team at their colocation facility. 37signals manages it remotely and built serious tooling to make that work.

So the celebrated example is not "buy a rack and save." It's a seasoned operations team paired with a managed physical layer. They didn't dodge the operational bill. They already had the team, and they rented away the metal. The DIY reader who quotes them is missing that both of those were already in place.

DORA made "own the metal" harder, not simpler

If your workloads touch EU financial services, repatriation does not hand compliance back to you in a clean box. DORA keeps the financial entity accountable even when it outsources, and it's shifting from paperwork to evidence. You now owe demonstrable, data-backed proof of operational resilience. In the 2024 dry run, only 6.5 percent of roughly 1,000 firms passed all the data-quality checks.

That evidence is an operational output. It comes from monitoring, patch records, backup testing, and incident logs that someone has to produce and keep current. Pulling a workload in-house doesn't remove that obligation. It moves the whole thing onto your side of the line.

The real choice isn't cloud versus metal. It's who runs the metal.

Notice that even the vendors cheering repatriation hedge the same way. The pitch quietly becomes "dedicated bare-metal performance with managed support around the physical layer" for leaner teams. The market has already conceded the argument. The decision was never public cloud against a rack in your closet. It's who owns the patching cadence, the uptime target, the capacity planning, and the compliance evidence.

This is exactly what LTFI is built to do. Every LTFI client runs on isolated, dedicated infrastructure, not shared tenancy. The servers are hardened Debian with automated patching, post-quantum SSH, fail2ban, AppArmor enforcement, and default-drop firewall policies, deployed on dedicated infrastructure with unified secrets management, DNS and SSL automation, and automated backup and recovery. Fleet management, monitoring, and 30-plus automated verification checks per deployment come with it.

That's the 37signals model without the prerequisite. You get repatriation economics without re-absorbing the headcount you offloaded to the cloud in the first place. Someone else owns the pager, the patch schedule, the capacity curve, and the evidence trail. You get the workload placement win and skip the operational trap.

Repatriation is a smart decision for the right workloads. Just finish the math the savings slide starts. The 30 to 70 percent is real, but only if the team behind it is too.

Talk to us about your infrastructure. ltfi.ai/contact

Tagged: Managed Dedicated Infrastructure

Cloud Repatriation Is 2026's Best-Selling Cost Story — But the 30-70% Savings Slide Never Shows the Team You Now Have to Run the Metal