Who actually moved back, and what it saved them

Start with the only repatriation figure in general circulation that a stranger can audit line by line, and note that it is not the figure everybody quotes.

Dropbox registered for its initial public offering on 23 February 2018. Buried in the management discussion is a section headed Recent Initiative, describing the migration of the vast majority of user data off third-party infrastructure and onto custom-built hardware in co-location facilities the company leases and operates. The migration completed in the fourth quarter of 2016.

Here is what the filing actually says happened to the money. Cost of revenue fell by $16.8m, or 4%, in 2016, driven by a net decrease of $39.5m in infrastructure costs. That net figure was a $92.5m fall in expense paid to the third-party datacentre service provider, offset by a $53.0m rise in depreciation, facilities and support costs on the equipment now sitting in Dropbox’s own racks. In 2017 cost of revenue fell a further $21.7m, of which $35.1m was infrastructure. Gross margin went from 33% in 2015 to 54% in 2016 to 67% in 2017.

The $39.5m and the $53.0m in the same sentence are the whole subject in miniature. Leaving a cloud provider does not delete a cost. It converts an operating expense billed monthly into a capital asset that depreciates, plus a facilities bill, plus a support function. The gross saving was more than twice the net one.

The famous number is a press aggregation

Add $39.5m and $35.1m and you get $74.6m, which is where the widely repeated claim that Dropbox saved $75m by leaving the cloud comes from. That total appears nowhere in the S-1. It is two consecutive year-on-year deltas summed by somebody else, and the two years are not comparable, because 2016 includes the cost of running duplicate copies of user data in both environments during the migration and 2017 does not.

More importantly, the filing tells you plainly that the programme was not only a migration. Throughout 2016 Dropbox also closed the accounts of long-inactive free users who had not responded to repeated emails, freeing storage capacity, and it made further usage optimisations in 2017. The company states that these efforts, together with the migration, are collectively referred to as its Infrastructure Optimization. The headline number therefore includes deleting data belonging to people who had stopped using the product. That is a perfectly sensible thing to do and it has nothing to do with where the servers are.

And Dropbox did not leave. The same document puts more than 90% of user data on its own racks, in three American co-location sites, with Amazon Web Services retained both for the balance of storage and for delivery, including in Europe. The canonical exit is a hybrid estate, described as such, under signature, on penalty of securities law.

The second checkable case published its invoices

37signals, which makes Basecamp and HEY, is the other organisation that has shown its working. In January 2023 it itemised a 2022 cloud bill of $3,201,564, or $266,797 a month, broken down by service: $907,838 on S3, $759,983 on EC2 and EKS, $519,959 on OpenSearch, $473,196 on RDS, $123,852 on Elasticache and $66,742 on CloudFront.

The company’s own account of what followed states that it ordered roughly $600,000 of Dell servers, completed the move in June 2023, and estimated conservatively that it would save $7m over five years, while keeping the same team operating the same applications.

Three things about that deserve stating rather than skipping. The $7m is a projection, not a realised saving, and it was published by the company that made the decision, whose chief technology officer also authored and promotes the deployment tooling built to carry the move. The $3.2m starting point was, by the same account, a bill that had already been optimised and haggled down repeatedly, and had once been twice as large. And the headcount claim is the one most likely to differ elsewhere: an organisation whose cloud operating model consists of four managed services and no on-call rota for hardware is not in the same position.

Even so, this is real evidence. A named company, dated figures, an itemised before, a stated after, and a method anyone can argue with.

The $400m that was never spent or saved

The third figure people cite is Ahrefs, and it belongs in a different category altogether. The company compared what it actually spent on 850 co-located servers with what the equivalent capacity would have cost on three-year reserved instances at AWS list prices in the Asia Pacific Singapore region, over roughly two and a half years from mid-2020, and put the difference at about $400m.

That is a counterfactual, and the article says so. Ahrefs never migrated to the cloud, so it never repatriated anything. The comparison uses published list prices rather than the negotiated rate a customer at that scale would be offered, and the workload is one of the least cloud-flattering in existence, which is a permanently loaded search index rather than a service with peaks and troughs. As a demonstration that some workloads are badly priced on rented infrastructure, it is useful. As evidence about repatriation it is worth nothing, and it is quoted as repatriation evidence constantly.

The number underneath all the others is an assumption

The claim that repatriation represents an enormous recoverable value in software originates in a single essay, published on 27 May 2021 by the venture firm Andreessen Horowitz. Its authors took roughly $8bn of aggregate cloud spend across fifty large public software companies, assumed a 50% saving from repatriation, and applied a gross profit multiple of 24 to 25 times, arriving at about $100bn of market value, which they then extrapolated to more than $500bn across a wider set of companies.

Read the arithmetic again. The finding is the assumption. Every pound of the $100bn is produced by the decision to write 50% into the model, and the piece’s supporting case study is the Dropbox number examined above. It was written by a firm whose business is owning equity in software and infrastructure companies, which is not a disqualification but is a disclosure the citations rarely carry.

Britain’s competition regulator looked at exactly this and would not use it

The most useful thing published on repatriation prevalence is not a survey. It is a regulator refusing to rely on the surveys.

In the final decision report of its cloud services market investigation, dated 31 July 2025, the Competition and Markets Authority had to decide whether on-premises IT belongs in the same market as public cloud. AWS argued that it does, and submitted repatriation research to prove it, including a survey finding that 65% of organisations that moved workloads away from hyperscalers moved them to their own on-premises services, a claim that between 24% and 36% of cloud customers have moved workloads back or plan to, and the Barclays survey of chief information officers reporting that 83% of respondents planned to repatriate at least some workloads in 2024.

The inquiry group placed limited evidential weight on all of it, and set out why in terms that transfer directly to any boardroom reading the same numbers. The word repatriation is imprecise, because it implies a permanent and absolute move when the shift may be temporary or cover a handful of workloads. Results are sensitive to which survey is used, a point the research submitted by AWS concedes itself. A binary question asking whether a customer switched to another cloud provider or to on-premises misclassifies anyone who did both. The word ever puts no bound on recency or frequency. And counting customers who moved any workload weights a firm that moved one job identically to a firm that moved everything.

Then the arithmetic the CMA did on AWS’s own figure: if 24% to 36% of cloud customers have migrated or plan to migrate workloads back, then 64% to 76% have never done so and do not plan to. From another of the submitted surveys it calculated that just 7.5% of infrastructure and platform customers had ever switched to an on-premises solution.

Note who was arguing which side. The party pressing the case that repatriation is widespread was the largest cloud provider in the world, because a wide market definition weakened the finding against it. Figures in this subject travel in whichever direction serves the person carrying them, and that is true of the challengers and the incumbents alike.

What would actually settle it

There is no reliable aggregate for the United Kingdom and this desk is not going to manufacture one. What exists is a small number of organisations that have published enough to be checked, and the standard they set is not difficult to state.

A named organisation. A dated document that carries a consequence if it is wrong, which means a filing, an audited account or a published bill rather than a conference slide. The workload described, because moving a data warehouse is not moving a customer-facing application. Before and after figures with the capital treatment made explicit, since depreciation, facilities and support are where the gross saving goes. The counterfactual stated, so a reader can tell an achieved saving from an avoided one. And the term of the comparison, because hardware bought in year one is still being depreciated in year four.

Two of the cases above meet most of that standard and the other two do not. Anyone who tells you what percentage of British enterprises have moved back is quoting a survey they have not read the method of. Why almost nobody ends up on one side or the other is the subject of hybrid stopped being the thing you settled for. The rest of this desk’s work sits under cloud and infrastructure.