Most cloud migration cost references are written from a single-workload or per-server angle. Full data centre exits work differently. The unit of analysis is the programme, not the server, and the cost shape is dominated by fixed costs that scale weakly with workload count: wave governance, decommissioning logistics, real-estate disposition, and the long-tail discipline that decides which workloads simply will not move. This page is the cost reference for those programmes in 2026.
A typical mid-enterprise data centre houses 200 to 1,500 servers, runs 50 to 250 applications, and operates either in a customer-owned facility or in colocation space (Equinix, Digital Realty, CyrusOne are the dominant providers in the US and UK). The all-in monthly run cost of a 500-server facility typically lands at $250K to $500K when counting power, cooling, space rent or amortisation, network connectivity, hardware refresh amortisation, and operations staff. This is the bill that the migration is trying to remove.
The cost of removing that bill is rarely a single migration project. It is a 18 to 30 month programme with eight to twenty-five waves, dependency-mapped across the application portfolio, with a wave-by-wave decision on rehost, replatform, refactor, retire, or retain. The cost shape concentrates in five buckets, in roughly the proportions shown below for a 500-server programme.
Data centre exit cost composition, typical mid-enterprise programme
| Cost bucket | Share of total | Notes |
|---|---|---|
| Migration labour (in-house + partner) | 40 to 55% | The largest single bucket. Partner-led typically 60 to 80% of this. |
| Parallel running infra (cloud + DC) | 20 to 30% | Compute, storage, network in both estates during cutover waves |
| Programme management and governance | 8 to 12% | PMO, architecture, change management, communications |
| Decommissioning and asset disposal | 2 to 5% | Hardware, data destruction certificates, logistics, vendor fees |
| Contingency and overrun reserve | 8 to 15% | Typically held against the long-tail workload problem |
Wave migration is the dominant pattern for any data centre exit above 100 workloads. The AWS Cloud Adoption Framework, Microsoft Cloud Adoption Framework, and Google Cloud Adoption Framework all describe variations of the same pattern: break the application portfolio into batches based on dependency clusters, plan each wave with a defined scope and timeline, execute the wave with a clear cutover gate, validate, and move to the next wave. The pattern works because it constrains complexity at any single point in time.
Wave sizing has a material cost impact. Waves of 5 to 10 workloads carry too much fixed overhead per wave (governance, change advisory, cutover coordination, post-cutover validation) and inflate per-workload cost by 20 to 35 percent. Waves of 50 to 100 workloads carry too much risk per wave; a single dependency miss can cascade across the entire wave and require rollback. The sweet spot for most mid-enterprise programmes is waves of 15 to 30 workloads on a 4 to 6 week cadence. At 500 workloads this implies roughly 20 waves over an 18 to 24 month programme.
Wave economics rule of thumb
Most enterprise data centres run in colocation space rather than in owned facilities. The colocation contract is typically a multi-year commitment with cancellation fees, ramp-down windows, and per-cabinet or per-megawatt billing. Exiting the contract requires coordinating the migration timeline with the contract renewal cycle, which often means aligning the final wave with the contract expiry to avoid a year of empty colocation space being paid for.
The three common colocation exit patterns. First, full exit aligned with contract expiry, which requires the migration programme to complete on a fixed timeline; this maximises cost recovery but adds programme risk if the migration slips. Second, partial exit with ramp-down billing, where the colocation provider agrees to reduce billing as cabinets vacate; this provides programme flexibility but typically carries an exit fee or rate increase. Third, transfer of the contract to a downstream user (acquisition, hosting provider, or DR site for another organisation), which recovers the full contract value but requires a willing counterparty.
For a 500-server estate occupying roughly 40 racks of colocation at typical $1,500 to $3,000 per rack per month, the steady-state colocation cost is $60K to $120K per month. A 24-month programme that exits at contract expiry recovers the full $1.4M to $2.9M annualised. A programme that exits six months past contract expiry pays a $360K to $720K penalty. The cost of misaligned exit timing frequently exceeds the cost of decommissioning the entire physical estate.
Physical hardware removal is a small line item by share of total cost but a significant operational discipline. The dominant categories are server hardware (typically 60 to 80 percent of asset count), storage arrays, network switches and routers, and edge infrastructure (UPS, PDU, KVM). Each needs assessment for residual value, data destruction, physical removal, and certificate of destruction generation for audit.
Hardware decommissioning cost by category
| Asset category | Per unit cost | Notes |
|---|---|---|
| Standard rack-mount server | $50 to $150 | Data wipe to NIST SP 800-88, physical removal |
| Server with regulated data (HIPAA, PCI) | $150 to $300 | Certified destruction with certificate |
| Storage array | $300 to $1,200 | Larger, often drive-level destruction required |
| Network switch / router | $40 to $120 | Configuration wipe, residual value often recoverable |
| UPS / PDU / rack infrastructure | $200 to $800 | Removal typically professional services, no resale |
| Mainframe | $5,000 to $25,000 | Specialised decommissioning required |
Vendor-funded recycling programmes (Dell Asset Resale, HPE GreenLake Lifecycle, Lenovo Asset Recovery) cover the lower end of the cost band and often produce small residual value cheques for hardware under five years old. Independent asset disposal vendors (Iron Mountain, Sims Recycling, Apto Solutions, Iron Mountain) handle larger estates with a full chain-of-custody audit trail. For regulated environments the audit trail is mandatory and the vendor selection criteria are dominated by certification status (NAID, R2, e-Stewards).
Owned data centres often have meaningful real-estate value once the facility is vacated. Recovery options are: repurpose for office or other commercial use, sell the property outright, lease the space, or convert to a different use entirely (warehousing, data centre for a different organisation, edge computing facility). The cash recovery lands 12 to 36 months after the migration completes because the property has to be cleaned, remaining hardware removed, and the new use prepared.
For colocation customers the real-estate recovery is the contract value rather than property value. The same programme discipline applies: align exit with contract expiry where possible, negotiate ramp-down where contract terms allow, and avoid the trap of carrying empty colocation space at full rate during the long tail of the migration.
A representative cost build for a 500-server data centre exit, mid-enterprise scope, 200 TB of data, mixed destination (60 percent AWS, 30 percent Azure, 10 percent Google Cloud) and mixed strategy (50 percent rehost, 25 percent replatform, 15 percent retire, 10 percent refactor), 24-month programme over 20 waves of 25 workloads each.
Worked data centre exit cost build, 500 servers, 200 TB, 24 months
| Cost line | Low estimate | Typical estimate | High estimate |
|---|---|---|---|
| Assessment and discovery (Migration Evaluator + partner) | $200,000 | $400,000 | $700,000 |
| Wave planning and PMO (24 months) | $1,200,000 | $1,800,000 | $2,600,000 |
| Migration labour, 450 workloads, blended | $3,200,000 | $6,400,000 | $11,500,000 |
| Tooling (DataSync, DMS, Azure Migrate, M4VMs) | $120,000 | $300,000 | $650,000 |
| Snowball Edge / Data Box / Transfer Appliance (15 devices total) | $60,000 | $120,000 | $220,000 |
| Direct Connect / ExpressRoute / Interconnect (24 months) | $140,000 | $280,000 | $520,000 |
| Parallel running, 18 months blended (compute + storage) | $2,400,000 | $4,800,000 | $8,000,000 |
| Colocation cost during migration (24 months at typical rate) | $1,400,000 | $2,200,000 | $2,900,000 |
| Cutover and downtime contingency (per wave) | $300,000 | $600,000 | $1,400,000 |
| Premium support (combined across 3 clouds, 24 months) | $700,000 | $1,000,000 | $1,400,000 |
| Staff retraining (40 engineers, multi-cloud) | $160,000 | $280,000 | $480,000 |
| Security rework (IAM, network, encryption, multi-cloud) | $300,000 | $700,000 | $1,500,000 |
| Hardware decommissioning and asset disposal | $80,000 | $140,000 | $240,000 |
| Real-estate / colocation exit fees | $0 | $150,000 | $500,000 |
| Contingency at 15 percent | $1,600,000 | $2,950,000 | $5,170,000 |
| Partner funding credits (MAP + AMMP + RaMP combined) | ($1,200,000) | ($1,800,000) | ($2,800,000) |
| Net total estimate | $10,660,000 | $20,320,000 | $33,980,000 |
The typical-column number, $20.3M for 500 servers over 24 months, works out at roughly $40,600 per workload. That is materially higher than the per-workload number for single-workload migrations because the wave overhead, PMO cost, multi-cloud security rework, and 24-month parallel running cost all add fixed-cost layers. The cost discipline that brings the typical-column number toward the low column is the wave structure: each wave shaved off the timeline saves roughly $400K to $800K of parallel running and governance overhead.
The final 5 to 10 percent of workloads in a data centre exit are disproportionately expensive to migrate. These workloads usually share three characteristics: no active development team that understands the application, vendor or third-party dependencies that have not been certified for cloud, and dependency on hardware or network features that do not have direct cloud equivalents. Specific examples that recur in cohort programmes: a payroll application from a vendor that went out of business in 2018, a manufacturing control system that depends on a specific serial connection card, a legacy ERP customisation that the original consulting firm took with them when the engagement ended.
Most programmes carve these workloads out into a separate sustain plan rather than forcing migration. The sustain plan typically maintains a minimal on-premise or colocation footprint at $30K to $150K per month per long-tail workload, with a multi-year plan to refactor or retire them as opportunities arise. The cost is real but it is bounded; the alternative of forcing a $500K to $2M refactor on a $40K-revenue workload rarely passes business case scrutiny.
A data centre exit is the largest, longest, and most complex shape of cloud migration. The per-workload cost is structurally higher than single-workload migrations. The discipline that produces an on-budget outcome is the wave structure, the alignment of colocation exit with the final wave, and the willingness to carve out the long tail rather than force every workload to migrate. The cost of forcing the long tail almost always exceeds the cost of sustaining it.
A. Mid-enterprise data centre exits of 500 to 1,000 servers typically cost $8M to $25M over 18 to 30 months, all-in including parallel running, partner consulting, equipment decommissioning, colocation exit fees, and contingency. The cost per server in a full data centre exit usually averages $15,000 to $25,000, materially higher than partial workload migrations because the wave planning, dependency mapping, and last-workload-out cutover discipline carry fixed costs that small migrations avoid.
A. Wave migration breaks the data centre estate into batches of 10 to 50 workloads per wave, with each wave moving over a 4 to 8 week cycle. Wave-based approaches add roughly 10 to 20 percent to the per-workload cost compared with a single large cutover, but reduce risk dramatically. Most enterprise programmes run 8 to 25 waves over an 18 to 30 month programme. The wave overhead is the cost of optionality: each wave can pause, adjust, or roll back without compromising the entire programme.
A. Yes, but slowly. Owned data centres can be repurposed (often to office or other commercial use), sold, or leased. The cash recovery typically lands 12 to 36 months after the migration completes because the property has to be cleaned, the hardware physically removed, and the new use prepared. Colocation contracts typically have 12 to 36 month exit windows; teams that plan the contract exit in parallel with the migration save 10 to 20 percent of the colocation cost compared with teams that exit after the migration completes.
A. Hardware decommissioning costs $50 to $300 per server depending on the certification requirements. Government and regulated estates require NIST SP 800-88 compliant data destruction with certificates, which costs roughly $150 to $300 per server. Commercial estates with no regulatory requirement can use vendor-funded recycling at $20 to $80 per server. Bulk asset disposal vendors (Iron Mountain, Sims Recycling, Apto Solutions) typically handle the logistics, certificate generation, and asset tracking for fees in the same band.
A. Small data centres of under 50 servers exit in 6 to 12 months. Mid-enterprise data centres of 200 to 500 servers exit in 18 to 24 months. Large enterprise data centres of 1,000+ servers run 24 to 48 months. The timeline is driven primarily by application complexity (the long tail of legacy workloads often takes longer than the bulk of modern workloads) and by regulatory or compliance constraints that limit the cutover speed.
A. The last 5 to 10 percent of workloads in a data centre exit are disproportionately expensive to migrate. These are typically legacy applications without active development teams, vendor-supported applications where the vendor has not certified cloud equivalents, or specialised hardware-dependent workloads (specific FPGA or appliance hardware). The cost per workload in the long tail can be 5 to 10 times the average, and many programmes carve these workloads out into a separate sustain plan rather than forcing migration.
On-prem to AWS ->
MAP credits, EC2, Snowball
On-prem to Azure ->
Hybrid Benefit, FastTrack
On-prem to GCP ->
Sustained-use, BigQuery
100-server worked scenario ->
Mid-market reference build
Lift-and-shift per server ->
$3K to $8K band decomposed
Refactor vs replatform ->
3-year TCO difference
VMware to cloud ->
Post-Broadcom licence flight
Mainframe modernisation ->
$5M to $50M range
10 hidden costs ->
Detailed playbook
Updated 2 May 2026