The promise was lower costs, infinite scale, and zero maintenance headaches.

For a lot of companies, the bill finally came due.

For more than a decade, cloud-first was almost a religion. Questioning Azure in a boardroom was roughly equivalent to questioning gravity. But something shifted. The companies that moved early now have years of billing data behind them, and that data is telling a different story than the one the hyperscalers originally pitched.

This doesn’t mean public cloud is bad. The history simply shows us predictable workloads carrying unpredictable price tags, and a growing number of CIOs doing the math out loud.

Here’s what they found.

1. The Cost Shock Case: 37signals (Basecamp/HEY)

This is the story that broke the spell.

When David Heinemeier Hansson, creator of Ruby on Rails and CTO of 37signals, published Why We’re Leaving the Cloud in October 2022, the technology industry’s reaction ranged from dismissal to outrage.

He was paying $3.2 million annually on AWS. He bought Dell servers for approximately $700,000, moved the compute workloads, and cut the bill by $2 million a year. The hardware paid for itself within the first year.

Then he came back for the storage.

He moved 18 petabytes off Amazon S3 to Pure Storage arrays for roughly $1.5 million. That infrastructure now costs less than $200,000 per year to operate, saving another $1.3 million annually. Total infrastructure spend is projected to fall from $3.2 million to well under $1 million, without hiring additional engineers.

His conclusion was simple:

"Cloud can be a good choice in certain circumstances, but the industry pulled a fast one, convincing everyone it's the only way."


2. The "We Did Everything Right and Still Got Burned" Case: GEICO

Scale doesn’t actually insulate you. It just makes the mistake bigger.

GEICO isn’t a startup that rushed a migration without a plan. It’s a Fortune 500 insurer with a mature IT organization that spent a full decade moving more than 600 applications to Azure. After all the planning, investment, and execution, costs reportedly came in 2.5 times over budget. Reliability declined. Vendor dependency deepened.

GEICO is now reversing course, repatriating at least 50% of those workloads to a private cloud environment while targeting 50% savings per compute core and 60% savings per gigabyte of storage.

If one of the world’s largest insurers can spend ten years getting cloud economics wrong, it’s worth asking whether you have enough data to know whether your own workloads are in the right place.

3. The Early Mover Validation: Dropbox

Dropbox was doing this before anyone called it cloud repatriation.

Between 2013 and 2016, Dropbox migrated the majority of its customer data away from AWS through an internal project called Magic Pocket, moving to proprietary colocation facilities. The project resulted in almost $75 million in savings over two years, while giving Dropbox significantly greater control over its storage platform.

The timeline is what makes this example so important, not just the savings.

This predates the current trend by almost a decade. When today’s wave of cloud repatriation gets dismissed as another technology cycle or a contrarian point of view, Dropbox is the answer.

This has always been a math problem, not a movement.

4. The Smaller Company Signal: Paddle, Honeybadger, Plausible Analytics

37signals and Dropbox are easy to dismiss because their scale is unusual.

The more practical signal is what’s happening further down the market.

Paddle cut approximately $400,000 a year after repatriation. Honeybadger reported a 60% reduction in infrastructure costs. Plausible Analytics reduced infrastructure costs by approximately 75%.

These aren’t household names with armies of infrastructure engineers. They’re companies that look a lot like yours. The pattern that emerges across all three is predictable workloads, growing data volumes, and cloud bills that gradually shifted from being convenient to becoming structural overhead.

When a bootstrapped analytics company can reduce infrastructure spend by three-quarters, the question becomes simple.

Does your workload profile match the ones where this works?

5. The Reliability Driver: CrowdStrike / Azure – July 2024

Cost is the loudest argument, but it isn’t the only one.

In July 2024, a faulty CrowdStrike update triggered a cascading failure across the Azure ecosystem that affected airlines, banks, broadcasters, and millions of endpoints around the world. The incident quickly became a boardroom conversation about what centralized dependency really means when something goes wrong.

No infrastructure architecture eliminates operational risk. But consolidated failure risk, where one vendor’s bad update can affect your organization and thousands of others simultaneously, is very different from distributed failure.

Even if hybrid and private cloud environments don’t make you immune to outages, they can give you greater control over the blast radius.

6. The AI Economics Driver

This is the emerging argument, and it may ultimately prove to have the longest lifespan.

As organizations run more AI workloads, whether that’s inference pipelines, GPU-accelerated applications, or large-scale model training, the economics of public cloud GPU access are pushing the math further toward private infrastructure. Equivalent GPU performance can cost 50–70% less on owned infrastructure than comparable workloads running in AWS or Google Cloud, depending on utilization.

Public cloud remains genuinely valuable for burst compute and irregular AI training jobs where you don’t want idle hardware sitting unused.

Continuous inference workloads are different.

They’re beginning to resemble the same predictable utilization patterns that drove companies like 37signals to rethink their cloud strategy. You’re paying for elasticity that may never actually be exercised.

What This Actually Means

The companies getting ahead of this are making infrastructure decisions now, before the GPU bill becomes the cloud bill.

Cloud repatriation is not a complete rejection of public cloud itself, but a rejection of the idea that public cloud is always the right answer.

The companies in this article didn’t leave the cloud because they failed at it. Dropbox built one of the largest cloud-native storage systems in the world before deciding to own its own infrastructure. 37signals ran HEY on AWS through a 10x growth spike before concluding that ongoing elasticity wasn’t worth the ongoing cost. GEICO completed one of the largest enterprise cloud migrations in the insurance industry before deciding half of it shouldn’t be there.

These are mature decisions made with real data, not reactions. The question worth asking isn’t “Should we leave the cloud?” It’s “Do we have enough data to know which workloads belong where?” If the answer is no, that’s where to start.

The CIOs making these decisions aren’t chasing headlines. They’re running the numbers and making infrastructure decisions based on what the data tells them.

If your cloud spend has become increasingly difficult to predict, it may be time to take a closer look at where your workloads belong. At CenterGrid, we help organizations evaluate their infrastructure, identify where public cloud continues to make sense, and where private or hybrid cloud may deliver better long-term value.

Talk to our team about building an infrastructure strategy that fits your workloads – not someone else’s.