Cloud Repatriation Math: How to Know If Moving Workloads Back On-Prem Actually Saves Money
Cloud repatriation is no longer a contrarian position. What was once the territory of a few outspoken CTOs is now a mainstream infrastructure strategy.
A Barclays CIO Survey from Q4 2024 found that 86% of CIOs planned to move at least some workloads from public cloud to private or on-premises environments — the highest rate ever recorded. IDC research found that 80% of enterprises expected to repatriate compute or storage workloads within twelve months.
The case studies are compelling. 37signals, the company behind Basecamp and HEY, famously saved $7 million over five years by leaving the public cloud for their own hardware. Dropbox built its own infrastructure under the codename "Magic Pocket" and reported tens of millions in annual savings.
But for every 37signals, there are organizations that ran the numbers poorly, repatriated the wrong workloads, underestimated operational costs, and ended up paying more than they would have in the cloud. Repatriation isn't inherently right or wrong. It's a math problem — and the math is more complicated than most analysis gives it credit for.
The Cloud Cost Side of the Ledger
Before you can evaluate repatriation, you need an accurate picture of what you're actually paying for a workload in the cloud. This is harder than it sounds.
Your cloud provider shows you compute costs. But the true cost of a cloud-hosted workload includes compute, storage, data transfer (particularly egress fees, which are often underestimated), support contract allocation, and the management overhead of the teams maintaining it.
Public cloud spending exceeded budgets by an average of 15% in 2024 according to Flexera — and that gap is largely driven by costs that weren't fully modeled upfront: egress fees, overprovisioned resources, reserved instances that don't match actual usage patterns.
Start here: take a workload you're considering for repatriation and calculate its fully loaded monthly cloud cost over the past six months. Include compute, storage, data transfer, and a proportional share of any support or tooling costs. Average it out. That's your baseline.
The On-Premises Cost Side of the Ledger
This is where repatriation analysis tends to go wrong — in both directions.
Organizations that are too optimistic about repatriation ignore the real costs of running infrastructure. Organizations that are too pessimistic overweight capital expenditures and undervalue the long-term economics of owning predictable workloads.
A rigorous on-premises cost model needs to account for:
Hardware acquisition and depreciation. Servers aren't free, but they're also not an ongoing expense the way cloud consumption is. A server purchased for $15,000 and depreciated over three years has a monthly hardware cost of roughly $417 — before power, before space, before maintenance. Amortize the capital expenditure over the expected useful life and treat it as a monthly cost equivalent.
Power consumption. A modern server in production draws 200–400 watts under typical load. At an average commercial electricity rate, factor in the Power Usage Effectiveness (PUE) of your facility — typically 1.5 to 1.8 for most enterprise datacenters, meaning you're spending 50–80% more on facility power overhead per watt of IT load. This is a real cost that's easy to undercount.
Rack space and facilities. Whether you own your datacenter or colocate, there's a cost per rack unit per month. For colocation, this is explicit in your contract. For owned facilities, it's a share of your facility cost that needs to be modeled.
Personnel. Who manages the hardware? Who handles OS patching, firmware updates, hardware failures? Cloud providers handle this labor for you. On-premises, it's either your team or a managed service provider. This is often the biggest underestimated cost in repatriation analysis, especially for organizations with lean IT teams.
Maintenance contracts and software licensing. Hardware maintenance agreements, operating system licenses, and any software you were getting "for free" bundled with your cloud service all need to be accounted for.
The Three-Year TCO Model
The right frame for repatriation analysis isn't monthly cost — it's three-year total cost of ownership. Cloud and on-premises have fundamentally different cost structures (variable versus mostly fixed), and a month-to-month comparison doesn't capture the full picture.
Build a three-year model with three scenarios:
Stay in cloud: Project your current cloud spend forward, accounting for expected growth in workload size and cloud provider price changes. Include reserved instances or savings plans if you'd purchase them. The U.S. Bureau of Labor Statistics Producer Price Index for cloud computing services rose 6.4% between September 2023 and May 2024 — factoring in some price inflation is not unreasonable.
Repatriate to owned hardware: Model Year 1 with capital expenditure as a lump sum and ongoing monthly operating costs. Years 2 and 3 carry only operating costs. Calculate the payback period — the point at which cumulative on-premises costs drop below cumulative cloud costs.
Repatriate to colocation: Similar to owned hardware, but capital expenditures are lower (you may be able to lease or buy less hardware), and facility costs are explicit monthly fees rather than amortized. Often a middle path between full cloud and owned infrastructure.
The Workloads That Repatriate Well
Not every workload is a repatriation candidate. The economics favor on-premises for workloads that are:
Predictable and steady-state. Cloud's value proposition includes elasticity — the ability to scale up and down on demand. If a workload runs at roughly consistent utilization month over month, you're not getting value from that elasticity, and you're paying for the option you're not using. Analysis from Broadcom suggests modern private cloud infrastructure can deliver 40–50% lower TCO compared to public cloud for steady-state workloads.
High in storage or egress volume. Egress fees are where cloud providers quietly earn significant margin on workloads involving large data volumes. If your workload moves a lot of data out of the cloud — to end users, to on-premises systems, to analytics tools — those fees compound quickly and are often the single biggest driver of repatriation economics.
Subject to compliance or data sovereignty requirements. For some workloads in regulated industries, the question isn't purely financial — it's operational. But even where compliance is the primary driver, the financial model still matters.
What Undermines Repatriation ROI
The analysis looks good on paper, then repatriation disappoints. This usually happens for a few predictable reasons:
Personnel costs were underestimated. The cloud abstracts away enormous operational labor. When you repatriate, that labor comes back in-house, and it's expensive.
The workload scaled faster than projected. Fixed on-premises capacity is great when utilization is predictable. When demand grows faster than expected, you're either buying more hardware (capital expenditure mid-cycle) or running saturated infrastructure.
Hidden cloud costs weren't captured upfront. If your cloud cost baseline was incomplete — missing egress, missing management tooling, missing support allocation — the comparison was flawed from the start.
Building the Case
A credible repatriation analysis requires accurate data on both sides of the ledger. That means complete cloud cost attribution for the workload in question — every service, every data transfer, every associated fee — and a rigorous on-premises model that accounts for all the costs above.
The organizations that make repatriation work are the ones that do the math honestly, model all three scenarios, and let the numbers tell them where each specific workload belongs. That's not a philosophical position about cloud versus on-premises. It's infrastructure cost management done right.
Reduce helps organizations model on-premises and cloud costs in a unified view — so repatriation decisions are based on complete, accurate data rather than incomplete cloud bills and optimistic on-prem estimates.