From a rooftop in Kathmandu, the cloud has always felt like weather — something that happens elsewhere and arrives here as bandwidth, existing prices, and other people's outages. But the enterprises that spent a decade treating "cloud-first" as gravity are now doing something that would have been heresy in 2015: hauling workloads back onto their own steel. I went through the economic case, the regulatory case, and the actual scale of this shift against primary sources before writing this, because "cloud repatriation" has become one of those phrases repeated so often it's stopped being checked.
The economics that make repatriation pay
The foundational document here is Andreessen Horowitz's May 2021 essay "The Cost of Cloud, a Trillion Dollar Paradox," by Sarah Wang and Martin Casado, and it's worth reading in the original rather than through secondary summaries, because its central claim is specific: repatriation results in one-third to one-half the cost of running equivalent workloads in the cloud. The essay estimates this dynamic had already erased roughly $100 billion in market value across 50 top public software companies at the time of writing, and more than $500 billion across the broader software sector. Its most concrete evidence is Dropbox's own S-1 filing, which disclosed $75 million in cumulative savings over the two years prior to its IPO after repatriating infrastructure.
The clearest recent case study, though, is 37signals — the company behind Basecamp and HEY — because its founder David Heinemeier Hansson has published the actual numbers repeatedly and specifically. The company's 2022 cloud bill exceeded $3.2 million, a figure DHH has described as "horrified" him at the time. After exiting the cloud, 37signals' bills fell by roughly $2 million a year, with about $700,000 in Dell hardware costs recouped within a single year of 2023, and the company now projects more than $10 million in savings over five years. Even after repatriating compute, the company disclosed in 2025 that it still carries roughly $1.5 million a year in S3 storage costs — and that AWS waived $250,000 in egress fees as part of the exit, a detail worth including precisely because it shows repatriation isn't a clean break so much as a renegotiation of which costs live where.
Verified Repatriation Savings
The Register (37signals); a16z, from Dropbox S-1
$0.0M/yr
37signals, post-repatriation cloud-cost reduction (The Register)
$0M
Dropbox, cumulative savings over 2 years pre-IPO (a16z, from Dropbox S-1)
Recalibration, not exodus
The scale of this trend matters as much as the individual case studies, and the clearest primary source is IDC. In its June 2024 report "Assessing the Scale of Workload Repatriation" (IDC Doc #US50903124), analyst Natalya Yezhkova found that about 80% of enterprises expected to see some level of repatriation of compute and storage resources within the next 12 months. Crucially, a separate IDC blog post from October 2024 found that only 8-9% of companies plan a full workload repatriation — meaning the dominant pattern is enterprises moving specific, steady-state workloads back on-prem while keeping bursty or experimental workloads in the cloud, not enterprises leaving the cloud altogether.
Barclays' CIO Survey tells a similar story from a different angle, and it's worth citing its two waves precisely rather than blending them: the April 2024 wave found 83% of CIOs planning to move at least some workloads back on-premises, and the second-half-2024 wave found that figure at a record 86% — both up sharply from a low of 43% in the second half of 2020. The trend line is the finding here, not either individual number: intent to repatriate at least something has been climbing steadily for four years.
IDC: Some Repatriation vs Full Exit
IDC Doc #US50903124 (June 2024); IDC blog (Oct 2024)
0%
Expect some repatriation within 12 months (IDC, June 2024)
0-9%
Plan a full workload exit from the cloud (IDC, Oct 2024)
Barclays CIO Repatriation Intent
Barclays CIO Survey, "2024 Outlook Sustained, 2025 Uplift"
Three separate survey waves, not interpolated — Barclays CIO Survey, "2024 Outlook Sustained, 2025 Uplift"
The sovereignty forces pulling in the same direction
Economics alone doesn't explain the acceleration; regulation does the rest. The EU Data Act has been directly applicable across the European Union since September 12, 2025, and it's pushed hyperscalers toward building infrastructure that can credibly claim to keep data within EU jurisdiction rather than simply within an EU data center owned by a non-EU company. Microsoft's response is the EU Data Boundary — and here's the correction worth making plainly: Microsoft's own announcement states the EU Data Boundary was completed on February 26, 2025, not February 2024 as sometimes reported. Getting a well-documented corporate milestone's date wrong by a full year is exactly the kind of error that undermines an otherwise well-sourced piece, so it's worth stating the correct date explicitly here. Microsoft followed that with its broader Sovereign Public Cloud and sovereign solutions announcement on June 16, 2025.
AWS has built a parallel product: the AWS European Sovereign Cloud, announced in October 2023 and launched in 2025, structured around German legal entities and EU-resident staff specifically to address sovereignty requirements AWS's standard European regions didn't satisfy. France has taken a more explicit localization approach, requiring US cloud providers to partner with local firms — Microsoft's tie-up with Capgemini and Orange being the clearest example — while Oracle has built EU sovereign regions in both Spain and Germany. Underneath all of this sits a genuine tension that Microsoft's own French subsidiary has acknowledged under testimony: it cannot fully guarantee data sovereignty against the reach of the US CLOUD Act, which allows US authorities to compel US companies to produce data regardless of where it's physically stored. That admission is arguably the single strongest piece of evidence for why "sovereign cloud" products exist at all — they're an attempt to engineer around a legal exposure that a data center's physical location alone can't fully close.
This piece is deliberately narrower than my earlier look at the AI inference cost curve, which is about the economics of running inference workloads specifically; this one is about where a workload legally and physically lives, for reasons that are as much about jurisdiction as they are about the monthly bill. It connects more directly to the splintering of sovereign compute I covered separately and to the broader inference-and-sovereignty reckoning facing hyperscalers — repatriation and sovereign-cloud buildouts are two responses to the same underlying pressure.
Sovereign-Cloud Milestone Timeline
blogs.microsoft.com; Reuters/TechInformed; European Commission
- Oct 2023AWS European Sovereign Cloud announcedGerman legal entity, EU-resident staff
- Feb 2025Microsoft EU Data Boundary completedFeb 26, 2025 — corrects a widely repeated 2024 date
- Jun 2025Microsoft Sovereign Public Cloudbroader sovereign solutions announcement
- Sep 2025EU Data Act directly applicableacross the European Union
The view from Kathmandu
What strikes me most about this story, watching it from outside the countries actually doing the repatriating, is how little of it is about rejecting the cloud as an idea. IDC's own numbers make that explicit: 80% expecting some repatriation against only 8-9% planning a full exit is not a company like 37signals proving the cloud was a mistake — it's the market recalibrating which workloads belong where, the same way a hydropower grid learns which loads are worth serving locally and which are worth exporting. The sovereignty angle adds a second axis to that calculation that has nothing to do with cost at all: increasingly, "the cloud" isn't one thing enterprises are choosing in or out of, it's a jurisdiction question layered on top of a cost question, and 2025's regulatory deadlines made the jurisdiction question impossible to defer any longer.
Sources
- a16z.com — "The Cost of Cloud, a Trillion Dollar Paradox," Sarah Wang and Martin Casado, May 27, 2021 (repatriation cost savings; Dropbox S-1 $75M figure; market-value estimates)
- The Register — 37signals cloud-exit coverage, Oct 21, 2024 (2022 bill,
$2M/yr savings, $700K hardware recoup) and May 9, 2025 ($1.5M/yr S3 bill, $250K AWS egress waiver); Data Center Dynamics — 37signals five-year savings projection - CIO.com; IDC — "Assessing the Scale of Workload Repatriation," Natalya Yezhkova, IDC Doc #US50903124, June 2024 (~80% some repatriation); IDC blog, Oct 28, 2024 (8-9% full exit)
- Barclays — 2H24 CIO Survey, "2024 Outlook Sustained, 2025 Uplift" (86% 2H24 reading; 83% Apr 2024 reading; 43% 2H20 low)
- European Commission — EU Data Act, directly applicable since Sept 12, 2025 (via TechInformed)
- blogs.microsoft.com — EU Data Boundary completion, announced Feb 26, 2025; Sovereign Public Cloud and sovereign solutions, June 16, 2025
- Reuters; TechInformed — AWS European Sovereign Cloud, announced Oct 2023, launched 2025
- Data Center Dynamics — France localization partnerships (Microsoft/Capgemini/Orange); Oracle EU sovereign regions (Spain, Germany)
- The Register — Microsoft French subsidiary testimony on CLOUD Act exposure, June 2025
Written by Abhishek Kushwaha, founder and writer at Global Tech Search, based in Kathmandu, Nepal.
