In 2022, the CTO who built Ruby on Rails opened his annual cloud invoice and found a number he could no longer defend: 37signals spent $3,201,564 on cloud that year, running the kind of steady, predictable workloads that public cloud is supposed to make cheap. Two years later the same company had pulled Basecamp, HEY, and five other apps back onto hardware it owns, spent roughly $700,000 on Dell servers, and watched its cloud bill fall by about $2 million a year. The hardware paid for itself inside twelve months. The five-year savings projection now tops ten million dollars.
That is not a contrarian stunt. It is arithmetic that thousands of finance leaders are quietly re-running. Gartner expects worldwide public cloud spending to hit $723.4 billion in 2025, up 21.5% in a single year, and somewhere inside that number is a margin transfer most companies never modeled. Flexera's 2025 State of the Cloud report found organizations waste 27% of their cloud spend and overshoot their cloud budgets by 17%, with 84% naming cloud cost management as their single biggest cloud problem.
Here is the reframe most CIOs still have not made. Cloud-first was never a strategy optimized for your balance sheet. It was optimized for someone else's. The hyperscaler's gross margin is your operating expense, and at scale the line between renting and owning stops being a convenience question and becomes a capital allocation decision. The companies repatriating right now are not nostalgic. They did the math, and the math changed.
The Paradox Andreessen Horowitz Put on the Record
The intellectual cover for this shift did not come from a hardware vendor. It came from one of the most cloud-bullish venture firms on earth. In 2021, Sarah Wang and Martin Casado of Andreessen Horowitz published The Cost of Cloud, a Trillion Dollar Paradox, and the conclusions landed hard precisely because of who wrote them.
Analyzing 50 of the top public software companies, a16z estimated that cloud infrastructure costs were eroding roughly $100 billion of aggregate market value by weighing down margins. Their framing has aged into a maxim: you are crazy if you do not start in the cloud, and you are crazy if you stay on it. The reason is structural. Cloud is brilliant when you are small, unpredictable, and racing to product-market fit. The moment your workloads become large and predictable, you are paying a premium for elasticity you no longer need.
Their headline case was Dropbox. The company saved nearly $75 million over two years by moving most workloads off public cloud onto custom-built infrastructure in leased colocation facilities. The result showed up where it matters most: Dropbox gross margins climbed from 33% to 67% between 2015 and 2017, an improvement the company attributed primarily to infrastructure optimization. a16z's blunt estimate was that repatriation can cut a heavy cloud bill by roughly half.
This Is Not Fringe. The Survey Data Is Damning.
The easy dismissal is that repatriation is a handful of loud founders. The survey data says otherwise. According to a Barclays CIO survey from late 2024, 86% of CIOs planned to move at least some workloads from public cloud back to private or on-premises environments, the highest reading the survey had recorded.
IDC's numbers point the same direction. Its June 2024 research found that about 80% of organizations expected some repatriation of compute and storage within the following twelve months, while only 8 to 9% planned a full exit. Read that carefully, because it is the whole point. Repatriation is not a stampede out of cloud. It is surgical. Companies are pulling the specific workloads where the economics inverted and leaving the rest.
A Citrix survey of 350 US IT leaders found that 94% had been involved in a cloud repatriation project in the prior three years, and 42% had already moved or were considering moving at least half of their cloud workloads back. The top drivers were not nostalgia. They were unexpected security issues (41%) and expectations the cloud simply failed to meet (29%).
The Unit Economics, Side by Side
The argument lives or dies on real numbers, so here are real numbers from companies that published theirs. The pattern is consistent: predictable, steady-state workloads are dramatically cheaper to own than to rent, while spiky and uncertain ones still favor the cloud.
| Company / workload | Cloud cost | Owned-hardware cost | Net result |
|---|---|---|---|
| 37signals compute (Basecamp, HEY, 5 apps) | ~$3.2M/yr AWS spend | $700K one-time Dell hardware | ~$2M/yr saved, hardware repaid in year one |
| 37signals storage (18 PB, S3 exit) | ~$1.5M/yr on S3 | $1.5M Pure Storage kit, under $200K/yr to run | ~$1.3M/yr saved after payback; AWS waived $250K egress |
| Dropbox (Magic Pocket) | Public cloud at scale | Custom colocation infrastructure | ~$75M saved over 2 years; gross margin 33% to 67% |
| Ahrefs (850 servers, 2.5 yrs) | ~$447M equivalent AWS quote | ~$39.5M colocation | ~$400M avoided by never fully going to cloud |
The Ahrefs figure deserves a caveat the company's own analysts offered: the AWS comparison was a list-price benchmark, not a negotiated enterprise contract, so the real delta is smaller. The direction, though, is not in dispute. When you run hundreds of servers at high, steady utilization, owning the iron wins by an order of magnitude.
Three Myths That Keep Companies Overpaying
Myth 1: "On-prem means a return to racking servers in a closet."
It does not, and this is the assumption that keeps the conversation stuck in 2010. Modern repatriation rarely means building a data center. It means leasing space in a colocation facility and running managed Kubernetes or a tool like 37signals' open-source Kamal on hardware you own, with the same deployment ergonomics engineers expect from cloud. The operational model is cloud-like. The cost structure is not.
Myth 2: "You will need to hire a huge ops team."
The single most cited reason this fails is staffing fear, and the evidence cuts against it. 37signals executed its entire exit without adding any new operations staff. The headcount math that justified cloud a decade ago assumed bare-metal tooling that no longer exists. Infrastructure as code, immutable images, and modern orchestration collapsed the operational tax that made cloud feel like the only sane option.
Myth 3: "Repatriation is an all-or-nothing bet."
This is the costliest myth because it frames a reversible, workload-by-workload decision as an identity crisis. The IDC data is explicit: only 8 to 9% of companies pursue full repatriation. The smart pattern is hybrid by design. Keep the bursty, experimental, and globally distributed workloads in cloud, where elasticity earns its premium. Pull the steady, high-utilization, data-heavy workloads onto owned hardware, where that same premium is pure margin leakage.
How the Bill Quietly Inverts
The reason this catches sophisticated teams off guard is that no single line item screams. The cloud bill grows the way a frog boils. A few forces compound at once.
- Egress and data-transfer fees. Moving your own data out of the cloud is metered, and at scale it becomes a structural tax. The fact that AWS waived a $250,000 egress charge for one departing customer tells you how large these fees get.
- Always-on premium for predictable load. Elasticity is priced into every hour, even the hours your load never varies. If your traffic is flat, you are paying for an insurance policy you never claim.
- Waste you cannot see. 27% of cloud spend is wasted on idle and oversized resources, and the org chart that creates it (every team self-provisions) is the same one that makes it nearly impossible to claw back.
- The reliability tax. GEICO's infrastructure leadership reported that a decade into its cloud journey, the company still had not migrated everything, its bills had risen 2.5x, and reliability challenges grew alongside them. Complexity is not free, and the cloud did not eliminate it. It relocated it onto your invoice.
The AI Bill Makes Repatriation Urgent, Not Optional
The cleanest case for owning hardware in 2026 is the most expensive line item on the modern P&L: GPU compute. The rent-versus-own gap that took 37signals five years to prove out shows up on a GPU server in a single quarter, because the markup on cloud accelerators is not subtle. An 8x H100 server bought outright and run around the clock amortizes to roughly $2.07 per GPU-hour over three years once power, cooling, and depreciation are counted. The same H100 on AWS lists at $7.50 and up per GPU-hour, and on Azure's ND H100 v5 it runs closer to $12 per GPU-hour. That is a three-to-six-times premium on the single most capital-intensive workload most enterprises will ever deploy.
The break-even is not a coin flip, either. Independent TCO modeling puts the crossover for an owned 8x H100 build at roughly 53 percent sustained utilization across the depreciation window, or about 18 months of near-continuous use when you include the data center costs. Anything below that line, rent. Anything above it, you are paying the cloud a multiple to hold hardware you could have bought.
And the spend is not slowing. Hyperscaler capital expenditure is forecast to clear $600 billion in 2026, a 36 percent jump over 2025, with about 75 percent of it tied directly to AI infrastructure. Data center capex itself rose 57 percent in 2025 (the top four US cloud providers grew theirs 76 percent) and is projected to pass $1 trillion in 2026, with high-end accelerators the primary driver. Every dollar of that buildout is a dollar the cloud intends to rent back to you at a margin.
Here is the structural point operators keep missing. Training is bursty and unpredictable, the workload cloud rental was built for. Inference is the opposite: a steady, high-utilization, always-on workload that runs every time a user hits the product. That is exactly the profile that favors owning. As AI moves from pilots to production, enterprises are pulling inference back on-premises for both cost control and low-latency edge delivery, because unbounded GPU bills and over-subscribed cloud queues do not survive contact with real revenue traffic.
AI did not make the cloud-versus-own math obsolete. It made the gap wider, because the most expensive new workload in the enterprise is also the most predictable. Anyone planning large-scale inference on rented GPUs without first modeling owned hardware is repeating the 2015 cloud-first mistake, only this time the invoice has a comma in a place it did not used to.
The Operator's Framework: When to Move What
The decision is not philosophical. Run every workload through four questions, and the answer falls out on its own:
- Is the load predictable? Flat, steady utilization is the strongest signal to own. Spiky and seasonal load favors cloud.
- How data-heavy is it? Storage-dominant and egress-heavy workloads punish you hardest in the cloud and reward repatriation fastest.
- What is the real total cost of ownership? Model three years, not three months. Include hardware, colocation, power, and the actual incremental headcount (which, per the evidence, is often near zero).
- Where is the regulatory and latency gravity? Data residency, sovereignty, and low-latency requirements increasingly tilt toward infrastructure you control.
Anything that scores predictable, data-heavy, cost-sensitive, and compliance-bound is a repatriation candidate today. Anything experimental, bursty, or globally distributed should stay where elasticity earns its keep. The mistake is not choosing cloud or choosing metal. The mistake is choosing once, in 2015, and never re-running the model as your workloads matured.
The Signal Everyone Else Is Ignoring
The companies moving workloads home are not retreating from modern infrastructure. They are refusing to let a default decision made years ago keep quietly transferring their margin to a vendor's income statement. a16z put the number at $100 billion of market value sitting trapped in that paradox, and the most disciplined operators are the ones reclaiming their share of it one workload at a time. Cloud-first was a strategy optimized for someone else's balance sheet. Cloud-smart is optimized for yours, and the difference, at scale, is measured in the millions of dollars 37signals, Dropbox, and Ahrefs have already booked.
The reckoning is not whether cloud was a mistake. It was not. The reckoning is whether you are still paying cloud prices for workloads that stopped needing the cloud years ago, and whether you have the nerve to re-run the math that everyone else is finally running.
Strategia-X is the senior operator that re-runs the unit economics of your infrastructure before the invoice does it for you, mapping which workloads to own, which to rent, and what the three-year balance sheet actually says: strategia-x.com.
-Rocky
#CloudRepatriation #FinOps #CloudCosts #ITStrategy #HybridCloud #VendorLockIn #EngineeringDreams #StrategiaX
