Contact

Cloud Repatriation: When Leaving Cloud Pays, and What It Actually Costs

Cloud repatriation is the move of workloads from public cloud to infrastructure you control, whether that is owned hardware in a colocation facility, rented bare metal, or a hybrid of dedicated and cloud. It is not a rejection of cloud. It is a decision about which workloads belong where, made on total cost rather than on the bill. Consider this your independent cloud repatriation review, written by an advisory practice that places workloads with operators and is paid by them rather than by you.

Summary: For companies spending seven figures a year on cloud with steady, predictable workloads, the movable share of that spend is usually a third to a half, and the saving on the movable share is usually 30 to 50 percent once every cost is counted. At that scale the destination is not a few cabinets. It is a wholesale colocation lease at several hundred kilowatts to several megawatts, or a bare metal fleet, or both. The two things cloud prices worst are constant utilization and high egress. If your workloads have both, the gap is wide. If they are bursty and light on data transfer, it is narrow or negative, and staying on cloud is the right answer. Which side you are on is a question of arithmetic, not philosophy, and at seven figures the arithmetic is worth doing properly.

If you are already there and want to know where your company lands, tell us what you are dealing with. Spend, the main workloads, and what is driving the question. It comes to the principal directly and you will have a first read within 24 hours, including whether the engagement is worth running for you.

What is cloud repatriation, and why is it back?

Cloud repatriation is the migration of applications, databases, compute and storage from AWS, Microsoft Azure, Google Cloud or a multi-cloud setup to dedicated infrastructure. In practice that means one of three things: owned hardware in a colocation facility or on-premise, rented bare metal servers from a provider, or a hybrid where steady workloads move and elastic ones stay. The cloud-to-on-premise version is rarer than it was, because most companies would rather rent space in a professional facility than run a server room, but the economics are the same.

It came back for a simple reason. Cloud is priced for elasticity. You pay a premium for the ability to scale down that a 24/7 workload never uses, and you pay again for every gigabyte that leaves the provider’s network. Companies that were small enough for that premium to be an IT line item are now large enough for it to be a board question. Dropbox saved $74 million over two years by moving off AWS. 37signals left in 2023 and published the numbers.

AWS itself launched a product in October 2025 that discounts egress by up to 80 percent for real-time bidding, because adtech companies were leaving over it.

And the largest repatriation motion in 2026 is AI. Companies that rent GPU capacity from cloud providers and neoclouds are moving to owned rack-scale hardware in colocation, because a GB300 NVL72 rack running training or steady inference at 140 to 160 kW is the most expensive thing to rent and the most predictable thing to own. That is a different scale of decision from moving a web tier: it is multi-megawatt, liquid-cooled, and it needs a facility that can actually serve the density. Our AI and GPU colocation page covers the facility side. This page covers whether to make the move at all.

None of that means cloud is wrong. It means the calculation has changed for a specific kind of workload, and the companies that run it are doing the math.

Cloud vs colocation vs bare metal: how the four paths compare

Every repatriation decision resolves into one of four scenarios. The table below is the comparison most companies build in their heads without writing down, and it is worth writing down, because the differences in speed and capital are larger than the differences in cost.

Scenario What it is Time to production Capital required P&L treatment Best for
Stay on cloud, optimized Reserved instances, savings plans, rightsizing, egress renegotiation Weeks None Opex, fully variable Bursty workloads, small footprints, teams with no infrastructure capacity
Bare metal Rented dedicated servers from a provider, no hypervisor, no capital Weeks to three months None Opex at a lower rate Steady compute that needs to move fast, or a first step before colocation
Colocation with owned hardware You buy servers, lease space and power in a data center. At scale this is a wholesale lease, 500 kW and up, with committed power and a ramp schedule Nine to fifteen months Hardware amortized three to five years Capex plus depreciation; opex for the facility Large steady footprints and AI clusters where three-year economics matter more than speed
Hybrid Steady workloads on dedicated, variable ones on cloud, private interconnect between them Follows the dedicated path chosen Depends on the dedicated path Mixed Most companies above $50,000 a month. This is where the analysis usually lands.

The scenario nobody should skip is the first one. Before moving anything, price what cloud costs once you have actually optimized it. Reserved capacity, committed use discounts, cutting the cloud waste that accumulates in any account over $50,000 a month, and a serious conversation with your account team can reduce cloud costs by 15 to 25 percent on a bill that has never been negotiated. The repatriation case has to beat optimized cloud, not list-price cloud.

How does each scenario show up in the P&L?

This is the question finance teams ask and infrastructure teams rarely answer, and it matters more than the headline saving.

Cloud as it is today is all operating expense, fully variable, and you own nothing at the end. Bare metal stays operating expense but at a lower rate, so EBITDA improves without capital outlay and nothing lands on the balance sheet. Colocation with owned hardware moves a large share of the cost into capital expenditure and depreciation. EBITDA improves more, because the recurring cost drops further, but there is a cash outlay up front and a depreciating asset to carry. Hybrid puts the steady workloads on whichever dedicated path fits and leaves the variable ones on cloud, so the P&L shows a mix.

Two things follow from that. First, for an adtech, SaaS or data company, infrastructure usually sits in cost of revenue rather than below the line, so a saving improves gross margin directly. That is a better story for investors than an operating expense reduction. Second, the choice between owning hardware and renting it is a financing decision, and it should be made separately from the infrastructure decision. Hardware leasing is standard, and it lets you take the colocation path without the capex if that is what the balance sheet prefers.

How Metro Colo Advisory works on this

Most companies at this scale have three options: run the analysis internally, hand it to a large consultancy, or go straight to quotes. Each has a problem. Internal teams have never priced dedicated infrastructure and tend to leave out the operations and platform costs, which overstates the saving by a third.

Large consultancies charge for the name and staff it with associates who have never negotiated a colocation lease.

Quotes price a configuration you have not decided on yet, and each vendor prices their own product.

What we do instead.

A fixed-fee engagement, four to six weeks, that produces the scenario-by-scenario P&L this page describes. Every cost counted, priced from operator contracts the practice has visibility into rather than from published rates, with an editable model your finance team keeps. The fee is the same whether the answer is move, stay, or somewhere in between. That is what keeps the recommendation clean.

Why the practice is positioned for this.

The placement side has been in front of more than fifteen operators in the last ninety days with live requirements. That is where the dedicated-side pricing in the model comes from. It reflects what a facility will actually sign, not what its website says, and that difference is usually larger than the margin the whole decision turns on.

What you get.

The four scenarios as a P&L over three to five years with cash timing. The movable list and the stay list, workload by workload. What egress actually costs you and what it would cost on dedicated infrastructure. A recommendation with the migration sequence and the risks per workload. And the model, so when your spend or cloud pricing changes, you rerun it rather than starting over.

What it costs.

Engagements run from $50,000 to $195,000 depending on scope. Half on signature, half on delivery. Against seven figures of annual cloud spend, the fee is what it costs to know the number before the decision rather than after.

When the answer is move

The placement side takes the requirement to operators across North America and internationally and brings back competitive proposals for wholesale colocation, bare metal, or both. That is paid by the operator, not by you, and the analysis fee does not depend on it.

How much do egress fees actually cost?

Egress is what a cloud provider charges to move data out of its network. Ingress is free. Egress is priced per gigabyte, and for a business that sends a lot of data outward, it is often the largest single line on the bill and the one that disappears most completely on dedicated infrastructure.

The published rates are similar across providers. AWS egress costs, GCP egress costs and Azure egress costs all run roughly 5 to 9 cents per gigabyte at volume, with the first tier higher and cross-region transfer priced separately. The egress cost line is the one that grows fastest as a business scales, because it tracks traffic rather than infrastructure. For a company moving 500 terabytes a month outward, that is $25,000 to $45,000 a month in egress before compute or storage.

On dedicated infrastructure you pay for a network port and a transit commit, and the marginal cost per gigabyte falls to a fraction of a cent.

The workloads where egress dominates are the ones that answer requests: real-time bidding, ad serving, content delivery, API platforms, anything that returns data to the outside world at high volume. AWS’s RTB Fabric product, launched in October 2025, is the clearest evidence of how much this matters. It cuts egress by up to 80 percent for real-time bidding traffic, but only when both sides of the exchange are on AWS, and it is not available in every region. Google and Microsoft have not announced an equivalent. If your egress bill is large and you are not on AWS, or you are on AWS but your counterparties are not, that line is a strong candidate to move.

How do you calculate cloud TCO against dedicated infrastructure?

Total cost of ownership is the comparison over three to five years, not year one. Whether the question is cloud vs colocation, cloud vs on-premise cost, or colo vs cloud for a single workload, the method is the same. Dedicated infrastructure loses year one to migration cost and, if you buy hardware, capital outlay. It wins years two through five if the workload is steady. The crossover point is usually somewhere in year two, and the cumulative gap by year five is what the decision rests on.

The cloud side of the model is the bill by workload, with an optimized-cloud line to make the comparison fair. The dedicated side has to carry everything a cloud bill hides:

  • Hardware, amortized over three to five years, with a refresh cycle built in. 
  • Facility: colocation rent per kilowatt or bare metal rental, power, cross-connects, remote hands. Primary-market colocation averaged about $196 per kilowatt per month in CBRE’s most recent survey. 
  • Interconnect: private links back to whatever stays on cloud, plus transit and peering for the traffic that used to ride on the cloud provider’s network. 
  • Operations: whoever runs the hardware. Usually a managed service provider rather than new hires, and usually the line companies forget. 
  • Platform layer: whatever replaces the managed services you use today. Managed databases, load balancers, Kubernetes, observability. Each is a product on cloud and a project on dedicated infrastructure. 
  • Migration: one-time cost of moving, including the months both environments run in parallel. 
  • Risk buffer: capacity headroom, redundancy, and the cost of a hardware failure when the replacement takes eight weeks. 


A model that leaves out operations and the platform layer will overstate the saving by a third or more. Most in-house comparisons do, which is why companies that move on a spreadsheet built from provider websites are often surprised in year two.

Bare metal or colocation: which dedicated path?

This is the decision inside the decision, and it is mostly about speed and capital.

Bare metal is a dedicated physical server you rent from a provider. No hypervisor, no other tenants on the machine, no capital outlay. Bare metal vs cloud is usually a 25 to 40 percent saving on steady compute, and it is the comparison most companies should run first because it needs no capital to test. Provisioning takes days to weeks because the hardware already exists. It costs more per month than owning over three years, but it lands the saving in a quarter rather than a year, and it lets you test the move before committing capital. For a company that needs a result before a fiscal year end, bare metal is usually the only path that gets there.

Colocation with owned hardware is the lower-cost path over three years and the slower one to reach. Procurement, delivery, racking and burn-in run nine to fifteen months from decision to production. The saving is larger because you own the asset, and the facility contract is negotiable in ways a bare metal rental is not. For a large steady footprint where the three-year number matters more than the first-quarter number, it is the right answer.

Many companies do both: bare metal for the first wave, colocation for the second, once the workload profile is proven on dedicated infrastructure. That sequence gets the saving early and the lower cost later, and it turns the migration from one bet into two smaller ones.

Which workloads should move, and which should stay?

The distinction is utilization and data movement, and it is worth applying workload by workload rather than to the bill as a whole.

Strong candidates to move

  • AI training and steady inference on GPU clusters. Rented GPU capacity is the highest-cost line in cloud, and a cluster that runs at consistent utilization is the clearest repatriation case there is. At rack scale this means owned GB300 or Rubin hardware in a liquid-cooled wholesale colocation facility 
  • Workloads that run 24/7 at consistent resource usage: databases, data warehouses, core application tiers, batch processing on a fixed schedule 
  • Anything with high outbound data transfer: bidding, serving, content delivery, high-volume APIs 
  • Storage-heavy applications where the cloud bill is mostly capacity you never release 
  • Applications with data residency or compliance requirements that make physical control an advantage

Better left on cloud

  • Workloads with extreme variability, spiking ten times or more against baseline 
  • Anything under roughly $20,000 a month in isolation, where the operational overhead of dedicated infrastructure exceeds the saving 
  • Applications deeply built on proprietary managed services that would need re-platforming to move 
  • Workloads requiring true global distribution across many regions simultaneously 
  • Anything expected to change shape dramatically in the next twelve months

The result is almost always hybrid. The steady core moves, the variable edge stays, and private interconnect keeps them connected without going over the public internet. Our hybrid cloud colocation guide covers the architecture; this page covers the decision.

A worked example at enterprise scale

The numbers below are illustrative, built for a data platform company at $1.2 million a month on cloud, and they are the shape of the analysis rather than any client’s figures.

Line Cloud today Hybrid, year one Hybrid, year three
Cloud compute and storage $720,000 $310,000 $290,000
Cloud egress $280,000 $60,000 $55,000
Cloud managed services $200,000 $120,000 $110,000
Bare metal rental 0 $210,000 $40,000
Colocation facility 0 $25,000 $95,000
Hardware amortization 0 $20,000 $130,000
Interconnect and transit 0 $35,000 $40,000
Operations and platform 0 $70,000 $75,000
Migration, amortized 0 $60,000 0
Monthly total $1,200,000 $910,000 $835,000
Annual saving $3.5 million $4.4 million

The colocation footprint in year three is roughly 2 to 3 MW, which is a wholesale lease rather than a cabinet agreement, with committed power, a ramp schedule and the contract terms that come with it. Three things to notice. Egress is the largest single reduction and it arrives immediately. Year one uses bare metal for speed and starts the colocation build in parallel, which is why bare metal rental is high in year one and low in year three. And the operations and platform line, $70,000 to $75,000 a month, is the one that in-house models usually leave out. Without it, this analysis would show $5.3 million a year instead of $4.4 million, and the company would find the difference the hard way.

 

If your numbers look anything like that, send us the shape of the problem. Annual spend, the main workloads, which clouds, and what is driving the question. That is enough for a first conversation, and if the numbers do not justify the engagement, we will say so.

What does a cloud exit strategy actually look like?

The sequence matters as much as the destination. A cloud exit strategy is the order in which workloads move, the parallel-running period for each, and the point at which the cloud commitment is reduced. Done in the wrong order, it costs more than staying.

  1. Workload assessment, weeks one to three. Billing export by workload, utilization data, and a conversation with the people who run the systems. This produces the movable list, the stay list, and the kilowatt requirement for the dedicated side. 
  2. Path selection and facility or provider search, weeks three to six. Bare metal, colocation or both. If colocation, two or three facilities that fit the connectivity, density and compliance profile, with competitive proposals. Our site selection guide covers the facility criteria. 
  3. Contracting, weeks six to ten. Facility or provider contract through legal review, with the escalation, minimum commitment and renewal terms negotiated before signature rather than discovered after. The data center lease guide covers what to look for. 
  4. Hardware, weeks eight to sixteen for bare metal, months four to twelve for owned. Bare metal provisions in days once contracted. Owned hardware is a procurement cycle. 
  5. Migration in waves, three to six months for live workloads. Non-critical first, production once stability is confirmed, with both environments running in parallel for each wave. The data center migration guide covers the mechanics. 
  6. Cloud commitment reduction. Reserved instances and committed-use contracts are reduced as workloads leave, not before. Cutting the commitment early is the most common expensive mistake. 

What this page cannot tell you

It cannot tell you which of your workloads are steady, because that depends on utilization data you have and we do not. It cannot tell you what your egress line is, because that is on your bill. It cannot tell you what dedicated infrastructure would cost for your specific footprint in your specific markets, because that depends on operator pricing that is contracted, not published. And it cannot tell you where your company lands inside the 30 to 50 percent range, because that depends on all three.

Those four things are the analysis. The framework above is what the analysis uses. The difference between the two is the difference between knowing that repatriation usually works for companies like yours and knowing whether it works for you.

Frequently Asked Questions

If they run at steady utilization, almost certainly. Rented GPU capacity is the most expensive line in cloud, and a training or inference cluster that runs around the clock is the clearest repatriation case there is. The move is to owned rack-scale hardware in a liquid-cooled colocation facility, which at 140 to 160 kW per rack is a wholesale lease, not a cabinet agreement. The hard part is not the economics, which are usually decisive. It is finding a facility that can serve the density on the timeline you need, which is a short list and getting shorter.

Bare metal: weeks to provision, three to six months to migrate live workloads. Colocation with owned hardware: nine to fifteen months from decision to production, because hardware procurement and facility build-out run in series. Most enterprise moves take six to eighteen months end to end, with the saving arriving in waves as each workload lands. The analysis that precedes the decision takes four to six weeks.

Bare metal is a dedicated server you rent; the provider owns the hardware and the facility. Colocation is space and power you rent for hardware you own. Bare metal is faster and needs no capital; colocation is cheaper over three years and gives you control of the asset. Many companies start on bare metal and move to colocation once the workload is proven on dedicated infrastructure.

For some workloads, egress alone does. At 5 to 9 cents per gigabyte, a business sending 500 terabytes a month outward pays $25,000 to $45,000 in egress before anything else. On dedicated infrastructure that line falls to a transit commit and a port fee. Real-time bidding, ad serving, content delivery and high-volume APIs are the workloads where egress is the deciding factor, and AWS’s RTB Fabric product exists because they were losing exactly those customers.

Yes, and the analysis is what gives you the leverage. A documented cost for the alternative is what makes a cloud provider move on price. Some companies run the analysis, take it to their account team, and get 15 to 25 percent off without moving anything. That is a fine outcome. The comparison has to be against optimized cloud, not list price, or the repatriation case is overstated.

Bare metal improves EBITDA directly, because the cost is still operating expense but at a lower rate. Colocation with owned hardware improves it more, because the recurring cost drops further, but part of the cost moves to depreciation below EBITDA and there is a capital outlay. For companies where infrastructure sits in cost of revenue, the saving improves gross margin as well. Which treatment you want depends on what the balance sheet and the investors prefer, and hardware leasing lets you choose.

Usually a managed service provider rather than new hires, with the colocation operator handling remote hands and hardware replacement. It is a cost line in the model, typically 5 to 10 percent of the dedicated infrastructure spend, and it is the line in-house comparisons most often leave out. Bare metal providers include hardware maintenance in the rental, which is part of why it is the simpler first step.

For bursty, variable workloads, yes. For steady workloads running 24/7, almost never over three years. The cloud vs on-premise cost comparison usually favors dedicated infrastructure by 30 to 50 percent on the movable share once utilization is above roughly 60 percent, because cloud is priced for the ability to scale down that a steady workload never uses. Colocation is the practical form of on-premise for most companies, since it provides the facility without the burden of running one.

No. The hybrid architecture that most companies land on keeps cloud connectivity through private interconnect, and workloads can move back if their profile changes. Bare metal contracts are typically one to three years; colocation contracts three to five. Neither is a permanent commitment, and the analysis includes what it would cost to reverse if the assumptions turn out wrong.

Related reading

Find out where your company lands

If the question is a P&L across scenarios at seven figures of annual spend or more, tell us what you are dealing with. It comes to the principal directly and you will have a first read within 24 hours.

If your footprint is smaller, start with the calculator, then the assessment. We come back within 72 hours with what equivalent dedicated infrastructure would cost, and whether the move is worth pursuing. No cost.

Contact Us
Contact