Cloud Repatriation in 2026: What the Surveys Say and Which Customer-Data Workloads Move First

Cloud repatriation in 2026 means moving selected workloads from public cloud back to private cloud, colocation or on-premise infrastructure; it rarely means leaving public cloud altogether. Two 2026 surveys report that 50% and 75% of respondents have already moved something back, but both come from vendors with something to sell, and IDC found only 8-9% planning full repatriation. For a customer data platform (CDP), identity graphs, consent stores, inference on personal data and long-retention event stores move first.

Who this is for: architects and data leaders deciding where customer data should run. Part of the guides hub. If you are already planning a move, pair this with the cloud exit strategy template.

What "cloud repatriation" means, and what the headline numbers measure

Repatriation is a placement decision for individual workloads, not a verdict on cloud. Most organisations that do it keep public cloud for elastic, bursty or globally distributed work and bring back steady, data-heavy or tightly regulated systems. IDC describes it as a growing trend but not a wholesale migration: most organisations move specific elements such as production data, backup and compute 3.

That distinction explains why survey headlines look contradictory. A claim that 75% have moved some workloads back and a claim that 8-9% plan full repatriation can both be true, because they measure different things.

Cloud repatriation statistics: what the surveys and case studies actually say

Cloudian and Centiment. The survey was fielded in February 2026 among 212 senior IT decision-makers at organisations with significant public cloud use. Its drivers are consistent with the Broadcom data: 84% exceed cloud storage budgets, egress fees (46%) and premium prices for residency-compliant regions (43%) lead the cost complaints, 55% say cloud cannot consistently meet AI inference latency, and 82% call sovereignty a primary or significant driver 2. Cloudian's own framing is that this is not enterprises souring on cloud: about 30% of respondents are still expanding their cloud presence 2.

IDC. IDC is the clearest counterweight. Its survey data say only 8-9% of companies plan full repatriation, larger organisations repatriate more than smaller ones, and cost overruns are a recurring theme 3. In October 2026 IDC's European analysts argued that AI will run across public cloud and on-premises deployments, shaped by where data must reside and by power availability, and that sovereignty is moving from debate to deployment without a single definition 7.

37signals. The most detailed public cost case. The company reported a 2022 cloud budget of US$3.2M, bought about US$500,000 of Dell servers, and projected savings of at least US$1.5M a year, with an operations team that did not change in size 4. In January 2026 it described moving about 5 PB of unique data (on the order of five billion objects) off S3 in under ten days of transfer, using a negotiated 90-day window to qualify for AWS's egress-fee waiver 5. It is one firm with steady workloads and strong in-house operations skill; treat it as an existence proof, not an average.

Gartner. We found no Gartner repatriation percentage we could verify. Its 2026 sovereign cloud forecast measures spending on sovereign cloud IaaS, not workloads moving back, so treat it as a demand signal for control 6. The European Commission's own proposal adds context: EU providers' share of the European cloud market fell from 29% in 2017 to 15% in 2022, and three non-EU hyperscalers hold over 70% 8.

How to read vendor-sponsored survey numbers

Use this checklist on any repatriation statistic, including the ones above:

Why customer data moves first

Customer-data workloads share three traits that favour repatriation: they are steady rather than bursty, they hold regulated personal data, and they grow with every event. Four candidates stand out.

Workloads that usually stay in public cloud: elastic campaign rendering, experimentation sandboxes, edge collection close to users, and third-party SaaS destinations you do not operate. A customer data platform on AWS, Azure or GCP in your own account is often the first step, before any hardware; the complete guide to self-hosted and private-cloud platforms compares the four deployment models.

A decision framework: cost, sovereignty, latency, egress, skills

Score each criterion 0 (stay), 1 (neutral) or 2 (move) per workload. Three or more twos on a workload make it a candidate; sovereignty or latency at 2 with skills at 0 means "move to a vendor-operated private cloud" rather than "self-host".

Egress arithmetic, illustrated. Moving a 200 TB event store at an assumed US$0.05 per GB costs about US$10,000 (200,000 GB x US$0.05). The inputs are assumptions for illustration only; use your contract rate. The sharper risks are time and conditions: 37signals negotiated a 90-day window because the waiver depended on completing the move 5. Under the EU Data Act, providers of data processing services may charge only cost-based switching charges until 12 January 2027 and none after that date 9. That does not make egress free: scope, multi-cloud traffic and carve-outs are still disputed, so read the contract. Detail in EU Data Act and cloud switching.

Twelve-point repatriation readiness checklist

• 1. Inventory customer-data workloads by data class (identity, consent, events, features, models) and jurisdiction.
• 2. Measure steady-state versus peak demand for each; note the peak-to-median ratio.
• 3. Compute 3-year cloud cost, including egress, support and marketplace fees.
• 4. Cost the owned or private alternative, including people, spares, power and refresh.
• 5. Quantify egress: volume, contract rate, waiver or credit conditions, Data Act switching terms.
• 6. List regulatory triggers: residency, exit-plan duties, sector rules.
• 7. Review provider-access exposure: key custody, support access, sub-processors, telemetry.
• 8. Set latency budgets and decide where inference runs.
• 9. Map dependencies: destinations, ESPs and APIs in the same region that break or slow down.
• 10. Plan a parallel run with dual writes, replay and reconciliation of identity and consent state.
• 11. Check commitments: committed spend, credits, notice periods, termination timing.
• 12. Define success metrics and a stop rule, and write down what deliberately stays in cloud.

The exit strategy template turns items 5 to 11 into a document you can test.

---