AWS told the UK Competition and Markets Authority (CMA) that customers can move workloads from public cloud back to on-premises infrastructure, arguing that this wider choice constrains cloud providers. Critics do not generally dispute that some workloads move. They question whether those examples are numerous or large enough to show that repatriation is a substantial competitive threat to AWS.
What AWS told the CMA
In a hearing summary dated July 2, 2024, AWS argued that cloud services compete with on-premises IT and the wider IT-services market. It said customers do move workloads back to on-premises systems, and rejected the idea that cloud adoption makes a customer’s choice irreversible. AWS’s argument was that the CMA should assess cloud in the context of those broader alternatives, rather than treating public-cloud infrastructure as a market with no meaningful outside options. AWS’s July 2024 hearing summary
AWS made similar points at a CMA hearing on March 12, 2025, saying that moving applications between IT providers has never been seamless but that cloud has made it easier than traditional data-centre migrations. AWS’s March 2025 hearing summary
That is both a claim about customer behaviour and an argument about market definition. If on-premises infrastructure is a close substitute for public cloud, it may affect how competition and market power are assessed. But demonstrating that one workload can be moved does not, by itself, show that on-premises infrastructure constrains cloud providers across the market.
#1 Best Overall
What critics say AWS may be overstating
The criticism is not necessarily that AWS invented customer moves. It is that isolated examples may overstate their frequency, scope or impact. Computer Weekly reported Gartner analyst Ed Anderson’s view that enterprise repatriation occurs but is too limited to call a broad market trend. ITPro also reported criticism of AWS’s framing, including from smaller cloud provider Civo. Computer Weekly’s account of the criticism · ITPro’s account of the debate
- Frequency: A handful of customer examples cannot establish how common repatriation is.
- Scope: Moving a database, storage tier or one application is not the same as leaving AWS.
- Competitive effect: Even real workload moves may not offset switching barriers, committed-spend arrangements or dependence on cloud-specific services.
What counts as cloud repatriation?
The term can describe materially different decisions. A full exit from public cloud is only one of them; calling every cost reduction or hybrid design “repatriation” blurs the distinction.
| Change | What it means | Does it show an AWS exit? |
|---|---|---|
| Full cloud exit | An organization moves its public-cloud estate to owned or privately operated infrastructure. | Yes, if the estate was on AWS and the move is complete. |
| Workload move | A particular application, database or storage system moves to a data centre, private cloud, colocation facility or hosted bare-metal service. | Only for that workload; the organization may retain substantial AWS use. |
| Hybrid placement | Some components run in cloud and others near users, equipment or data in private infrastructure. | Not necessarily. It may reflect architecture rather than a cloud departure. |
| Cloud-to-cloud migration | A workload moves from AWS to another public-cloud provider. | No. It is a provider switch, not a move back on-premises. |
| Cloud cost optimization | Teams resize instances, change commitments, use autoscaling or redesign services to reduce spend. | No. AWS consumption may fall without a change in hosting location. |
Colocation, owned data centres, private cloud and hosted bare metal also differ in who owns the hardware, operates the facility and carries the capacity risk. They should not be treated as interchangeable simply because none is a conventional public-cloud service.
What the evidence can—and cannot—show
A useful measure of repatriation needs a clear denominator. The number of organizations that moved anything back, the share of workloads moved, the share of cloud spending removed and the number of organizations that exited cloud entirely are different statistics. Net cloud growth is different again: new cloud adoption can outpace workload departures.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Rank #2
Uptime Institute reported that 6% of respondents to its 2022 data-centre survey had abandoned public cloud altogether. That is a survey result about complete exits, not a finding that the other 94% never moved any workload. Nor does it establish a current global exit rate or the proportion of cloud spending repatriated. Uptime also reported cost as the leading stated driver among organizations moving workloads back, and cited storage expense in connection with moves such as 37signals’. Uptime Institute’s discussion of its survey
Case studies and survey answers can establish that particular moves occurred, but they do not automatically measure the market-wide trend. A convincing assessment would distinguish workload type, amount of spending moved, destination, timing and whether the workload later returned to cloud. The available figures cited here do not establish that repatriation is large enough by itself to discipline AWS’s market power.
Why organizations move some workloads back
Predictable, continuously used capacity
Cloud’s ability to scale quickly is valuable when demand is uncertain. For a steady workload running at high utilization, dedicated hardware may be less expensive over several years—especially for an organization that already has facilities and infrastructure staff. The comparison changes if a company must build those capabilities from scratch or buy enough capacity to cover occasional peaks.
Storage and data movement
Large data stores can carry recurring storage, backup, replication and network-transfer costs. Data-heavy systems may be candidates for a different location when those charges are material and access patterns are predictable. The relevant calculation is not storage price alone: moving the data, maintaining copies and connecting systems also have costs.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteRank #3
Performance, control and compliance
Some applications need predictable local performance, specialized hardware or low-latency links to industrial equipment. Organizations may also choose private infrastructure for operational control, data-residency obligations or sector-specific compliance. Those motivations are not necessarily evidence that public cloud was too expensive; they may reflect requirements that vary by workload and jurisdiction.
Accelerated computing
GPU and other accelerator workloads depend on hardware availability, power, utilization and depreciation. Cloud can provide access without a long procurement cycle, while owned capacity may suit sustained use if it can be kept busy. Neither model is automatically cheaper across different utilization and availability conditions.
Why public cloud remains useful
Repatriation is not proof that cloud was a mistake. Public cloud can avoid buying capacity before demand is known, speed up experimentation and provide global regions, managed databases, analytics, security services and access to specialized hardware. It also shifts much of the physical data-centre operation to the provider and offers a large ecosystem of skills and partners.
AWS’s own cost guidance says that a meaningful comparison should examine service choices and total cost of ownership, including operations and management rather than infrastructure prices alone. AWS Well-Architected guidance on service selection and cost analysis
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsFor some applications, the value of deployment speed, managed services or elastic capacity outweighs the potential savings from dedicated infrastructure. For others, stable demand and high data-transfer costs may make another placement worth considering. The decision belongs at workload level, not at the level of a slogan about all cloud or all on-premises infrastructure.
How the CMA’s findings change the debate
The CMA’s UK investigation provides context, not a ruling on whether AWS’s repatriation examples are representative. Its provisional findings, published January 28, 2025, said UK customers spent £9 billion on cloud services in 2023, with spending growing by more than 30% annually at that point. The CMA said AWS and Microsoft each had up to 40% of UK customer cloud spending, with Google substantially smaller. These are UK-specific estimates, not global market shares. CMA provisional findings
The CMA identified concerns including limited provider choice and technical and commercial obstacles to switching or using multiple providers. Its case page says the investigation closed on July 31, 2025, and records a recommendation to consider strategic-market-status investigations for AWS and Microsoft. CMA cloud-services market investigation
The two positions can therefore both contain truth. On-premises infrastructure may be a real alternative for some workloads, as AWS argues, while barriers may still make it difficult or costly for many customers to switch providers or divide work across them. The CMA’s work concerned competition in public-cloud infrastructure services; AWS argued that the market should be considered in the broader context of IT services. That difference in market definition matters independently of whether a particular company found a workload cheaper elsewhere.
Best Value
What makes switching difficult
Leaving or diversifying away from a cloud provider can involve more than copying data to new servers. Technical dependencies, commercial terms and business risk all contribute to switching costs.
- Data transfer: Egress and inter-region charges can add cost, while large transfers take time and require a migration plan.
- Application dependencies: Proprietary managed databases, messaging, analytics and security services may need replacement or redesign.
- Infrastructure integration: Identity and access controls, monitoring, automation, network architecture and database compatibility may be provider-specific.
- Commercial commitments: Committed-spend agreements and licensing arrangements can affect the economics of moving or splitting workloads.
- People and risk: Staff retraining, compliance recertification, downtime and operational differences increase the effort and uncertainty.
The CMA’s investigation examined egress fees, committed-spend agreements, licensing, technical barriers and multi-cloud use. AWS’s counterargument is that cloud has made provider changes easier than traditional data-centre migrations, not that they are effortless.
Does repatriation threaten AWS’s growth?
Its effect depends on the net change, not simply on whether a move occurs. Repatriation can reduce AWS revenue from particular workloads, improve customers’ bargaining positions and make colocation or smaller cloud platforms more credible alternatives. But the effect on the provider as a whole depends on how much spending leaves, what remains, and whether new cloud workloads grow faster than departing ones.
A customer might move a predictable, storage-heavy system while continuing to use AWS for elastic applications, managed services or new projects. A move to private infrastructure may also be reversed if requirements change. Amazon CEO Andy Jassy’s 2025 shareholder letter offers company context on AWS and AI expansion, but corporate growth statements do not measure the scale of repatriation. Amazon’s 2025 shareholder letter
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →How to compare cloud with private infrastructure
Finance and technology teams should compare realistic alternatives for a specific workload over a consistent planning period. A three-to-five-year model may be appropriate for a hardware decision, but its result depends on the workload, contract terms and refresh assumptions; it is not a universal rule.
- Define the workload: Record utilization, demand peaks, data volumes, performance targets, availability requirements and dependencies on managed cloud services.
- Model all cost components: Include compute and accelerators; storage, backup and replication; network ingress, egress and inter-region transfer; software and licensing; and facilities, power, cooling, connectivity and rack costs.
- Include operations and risk: Add hardware purchase, depreciation, maintenance and refresh; security, monitoring, incident response and compliance; engineering and operations labor; disaster recovery; and migration, retraining and exit costs.
- Test demand and capacity: Model average and peak use, headroom, seasonality, procurement lead times and the cost of idle capacity. Compare like-for-like resilience and availability.
- Compare credible destinations: Evaluate AWS, at least one alternative cloud where appropriate, and a colocation or private-infrastructure option. Account for each destination’s service dependencies and migration effort.
- Challenge the assumptions: Separate flexible on-demand prices from reserved or committed rates, document utilization and depreciation assumptions, and have procurement or finance independently validate vendor-produced estimates.
- Revisit after deployment: Compare forecast with actual usage and costs, then optimize placement as demand, hardware and business requirements change.
Common errors include comparing a cloud compute bill with a server purchase price, omitting staffing and facilities, ignoring migration engineering or backup, and assuming an application will run unchanged outside AWS. Cloud cost optimization can reduce spend without proving that the cloud model itself is uneconomic.
AWS Cost Explorer can help inspect AWS costs, but it is not a neutral cross-cloud TCO model. AWS publishes its Cost Explorer pricing and API charges separately; teams should check the current terms and assess whether the tool covers their decision. AWS Cost Explorer pricing
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
