The Great Cloud Repatriation: Why Some Companies Are Bringing Workloads Back On-Premise

The Cloud-First Era Meets Its Counterargument

For the better part of a decade, the technology industry operated under a simple orthodoxy: move everything to the cloud. In 2019, Gartner predicted that 80% of enterprises would shut down their traditional data centers by 2025. AWS, Azure, and Google Cloud grew into a combined infrastructure services market exceeding $210 billion annually, with total cloud spending topping $750 billion in 2025. CIOs who did not have a “cloud-first” strategy were viewed as dinosaurs.

Then David Heinemeier Hansson — DHH, creator of Ruby on Rails and co-founder of 37signals (Basecamp, HEY) — published a series of blog posts in late 2022 that landed like a grenade in the infrastructure discourse. 37signals was leaving AWS. They had spent $3.2 million on cloud services in 2022 and calculated they could run equivalent infrastructure on owned hardware for a fraction of the cost. By mid-2023, the compute migration was complete. By October 2024, DHH reported annual compute savings of nearly $2 million after investing $700,000 in Dell servers. Then in May 2025, 37signals completed the final phase: migrating off AWS S3 entirely, spending $1.5 million on 18 petabytes of Pure Storage equipment that costs under $200,000 per year to operate. AWS even waived a $250,000 data egress bill for the departure. DHH’s updated projection: cumulative savings will top $10 million over five years, with the total annual infrastructure bill dropping from $3.2 million to well under $1 million. The company plans to delete its AWS account entirely.

DHH was not alone. Dropbox had quietly completed one of the largest cloud repatriations in history years earlier, moving 90% of its storage infrastructure from AWS to three dedicated colocation facilities in California, Virginia, and Texas starting in 2015. The company invested over $53 million in its own infrastructure and estimated it saved $75 million over two years. Ahrefs, the SEO analytics company, runs its multi-petabyte infrastructure on owned hardware in a Singapore colocation facility and claims savings of $400 million over three years compared to equivalent cloud costs — estimating that one server in the cloud would cost 11.3 times more than its collocated setup. A 2021 Andreessen Horowitz report — “The Cost of Cloud, a Trillion Dollar Paradox” by Martin Casado and Sarah Wang — found that cloud costs consume an average of 50% of the cost of goods sold (COGS) for the 50 public SaaS companies analyzed, and argued that the top 50 public software companies collectively lose $100 to $200 billion in market value due to cloud’s margin impact.

These are not fringe voices. A June 2024 IDC survey found that approximately 80% of respondents expect to repatriate some compute and storage resources within 12 months. A 2024 Barclays CIO survey reported that 86% of chief information officers plan to move at least some public cloud workloads back to private cloud or on-premises environments. Gartner predicts that 40% of enterprises will adopt hybrid compute architectures in mission-critical workflows by 2026.

When Repatriation Makes Financial Sense

Cloud repatriation is not universally correct, and its advocates do not claim it is. The financial case for leaving the cloud depends on specific workload characteristics, and understanding these characteristics is more valuable than any ideological position.

Predictable, steady-state workloads are the strongest candidates for repatriation. When a company knows it needs 500 servers running 24/7/365, the cloud’s core value proposition — elasticity, pay-for-what-you-use — provides no benefit. You are using all of it, all of the time. In this scenario, cloud pricing includes a hefty margin (AWS posted operating margins of approximately 35% in fiscal year 2025, peaking at 39.5% in Q1) on top of the actual infrastructure cost. Buying or leasing equivalent hardware from Dell, Lenovo, or Supermicro and running it in a colocation facility eliminates that margin. 37signals estimated that their owned hardware cost roughly 60% less than equivalent cloud instances over a five-year depreciation cycle.

Data gravity creates another repatriation trigger. Companies with large datasets — multi-petabyte and above — face punishing cloud egress charges when they need to move data out of the cloud provider’s network. AWS charges $0.09 per gigabyte for the first 10 TB of monthly data egress (with tiered discounts at volume, and a recently expanded free tier of 100 GB per month). For a company transferring 1 PB per month — common for large-scale analytics, media processing, or backup operations — that is $90,000 monthly in pure network fees. Dropbox’s repatriation was driven substantially by the economics of storing and serving exabytes of user files. Notably, AWS now waives egress fees for customers migrating off the platform entirely — a concession that effectively acknowledges repatriation as a legitimate trend.

AI workloads have become the most significant new driver of cloud repatriation in 2025-2026. Model training, inference at scale, video encoding, financial modeling, and scientific simulation all require sustained access to expensive GPU hardware. Cloud GPU pricing still carries substantial premiums — an NVIDIA A100 80GB instance costs roughly $1.40 to $5.00 per hour depending on provider (down from higher figures in previous years as competition from Lambda, CoreWeave, and RunPod has pressured pricing). The same GPU can be purchased for $15,000-$20,000 and amortized over three years at an effective cost well under $1.00 per hour. For organizations running GPU workloads at high utilization rates, the economics overwhelmingly favor owned hardware. This is why AI infrastructure is now the leading catalyst for enterprises evaluating their on-premise options.

When the Cloud Still Wins

The repatriation narrative, however compelling, has a selection bias problem. The companies that leave the cloud publicly are the ones for whom it makes financial sense — and they are systematically different from the broader market. Most companies are not 37signals. IDC data confirms that only 8-9% of organizations plan full workload repatriation; the vast majority are selective, moving specific workloads while keeping others in the cloud.

Variable and unpredictable workloads remain the cloud’s strongest use case. A retail company whose traffic spikes 10x during Black Friday, an event platform that scales from 1,000 to 1 million concurrent users for major broadcasts, a startup whose growth trajectory is genuinely unknown — these organizations would be foolish to buy hardware for peak capacity that sits idle 90% of the time. The cloud’s elasticity is not just convenient; it is economically superior when utilization is low and variable.

Global distribution is another domain where the cloud’s built-in infrastructure is nearly impossible to replicate. Serving users with low latency across North America, Europe, and Asia requires points of presence on three continents. Building and operating your own global network is a billion-dollar endeavor reserved for the largest technology companies. Even Ahrefs, with its $400 million in claimed savings, still relies on AWS for frontend hosting across multiple global regions. For everyone else, cloud regions and CDN services provide global reach at a fraction of the cost and complexity of doing it yourself.

Speed of iteration matters too, and it is harder to quantify. A startup that can deploy a new microservice in minutes using managed Kubernetes, connect it to a managed database, and scale it automatically is moving faster than one that needs to provision hardware, configure networking, and manage capacity planning. This velocity advantage compounds over time, particularly in competitive markets where product development speed is a strategic differentiator. The cloud’s hidden value is not just infrastructure; it is the managed services ecosystem — databases, message queues, ML platforms, observability tools — that would require dedicated teams to replicate on-premise.

The Hidden Costs of Coming Home

The most common mistake in repatriation analysis is comparing cloud bills to hardware costs while ignoring the full cost of running on-premise infrastructure. DHH is transparent about 37signals’ approach: they employ a dedicated infrastructure team, they run their hardware in colocation facilities (not self-owned data centers), and they have deep institutional expertise in managing servers built up over 20 years. Most companies do not have this.

Staffing is the biggest hidden cost. A competent infrastructure team — systems administrators, network engineers, security specialists, on-call rotation — requires 3-8 full-time employees depending on scale, at fully loaded costs of $150,000-$250,000 per person in the US market. That is $450,000-$2,000,000 annually in labor costs before touching hardware. Companies that have outsourced infrastructure management to cloud providers for years may need to rebuild this expertise from scratch, with all the recruitment costs, ramp-up time, and institutional risk that implies.

Hardware lifecycle management adds ongoing costs that cloud providers abstract away. Servers depreciate over 3-5 years and must be replaced. Storage arrays need expansion. Network equipment reaches end-of-life. Each hardware refresh requires capital expenditure approval, procurement cycles, physical installation, and migration of workloads — operational complexity that cloud consumption neatly avoids. Warranty management, spare parts inventory, and vendor relationship management all consume time and attention that could go toward product development.

Colocation costs are not trivial either. Enterprise colocation in major markets runs $150-$300 per kW per month. A modest 20-rack deployment drawing 100 kW costs $15,000-$30,000 monthly in facility charges alone — before power consumption, which is billed separately. Add redundant network connectivity from two or more ISPs, and the facility cost baseline is $25,000-$50,000 per month. These costs are predictable and typically lower than equivalent cloud spend, but they are not zero, and they must be included in any honest comparison.

The “Cloud-Smart” Compromise

The emerging consensus among pragmatic infrastructure leaders is neither “cloud-first” nor “cloud-exit” but “cloud-smart” — a workload-by-workload evaluation of where each application should run based on its specific characteristics. This is less catchy than either extreme but more intellectually honest and financially sound.

In practice, cloud-smart means running steady-state, predictable workloads on owned or leased hardware (colocation or private cloud), using public cloud for variable workloads and managed services that would be expensive to replicate, and treating cloud provider regions as deployment targets rather than strategic commitments. It means evaluating total cost of ownership — including staffing, maintenance, and opportunity cost — rather than comparing monthly bills to hardware invoices.

Several vendors are building for this hybrid reality. Oxide Computer Company, founded by former Joyent CTO Bryan Cantrill, builds rack-scale servers designed to bring cloud-like operational simplicity to on-premise hardware. The company’s momentum is telling: Oxide raised $100 million in a Series B round in July 2025, followed by a $200 million Series C in February 2026, with customers including Lawrence Livermore National Laboratory and Idaho National Laboratory. Tailscale and Nebula provide software-defined networking that connects on-premise and cloud infrastructure seamlessly. Crossplane, which graduated from the Cloud Native Computing Foundation in November 2025, allows teams to manage both cloud and on-premise resources through a single Kubernetes-based control plane. The tooling for hybrid infrastructure is maturing rapidly, reducing the operational overhead that historically made on-premise management so painful.

For technology leaders evaluating their infrastructure strategy in 2026, the question is not whether to use the cloud — virtually every organization will use cloud services for something — but how much of the cloud to use and for which workloads. Gartner’s 2019 prediction that 80% of enterprises would abandon traditional data centers by 2025 did not materialize; survey data shows roughly 43% of workloads still run in corporate data centers. The answer requires honest financial analysis, realistic assessment of operational capabilities, and a willingness to reject both the “cloud-first” dogma of the 2010s and the “cloud-exit” counter-narrative that has emerged in response. The boring, correct answer is usually somewhere in between.

Frequently Asked Questions

Sources & Further Reading

• Our Cloud-Exit Savings Will Now Top Ten Million Over Five Years — DHH / 37signals
• 37signals On-Prem Migration to Save Millions, Abandon AWS — The Register
• Dropbox Saved Almost $75 Million Over Two Years by Building Its Own Infrastructure — GeekWire
• How Ahrefs Saved US$400M in 3 Years by NOT Going to the Cloud — Ahrefs Tech Blog
• The Cost of Cloud, a Trillion Dollar Paradox — Andreessen Horowitz
• Storm Clouds Ahead: Missed Expectations in Cloud Computing — IDC
• Cloud Trends 2025: Repatriation and Sustainability Make Their Marks — InfoWorld
• Oxide Closes $200M Series C to Scale On-Premises Cloud Computing — Intel Capital
• CNCF Announces Graduation of Crossplane — CNCF
• AWS Egress Costs in 2025: How to Reduce Them — nOps