37signals cut its cloud bill from $3.2 million to $1.3 million a year by moving seven apps off Amazon and onto its own gear [3][5]. That move won back almost $2 million a year. The firm now expects to save $10 million over five years [2][3].

The big picture:

Cloud prices charge extra for demand you cannot guess.

When an app runs flat and steady, you still pay for room you never use. Andreessen Horowitz looked at one software firm worth a billion dollars. Its cloud spend reached 81% of the cost of sales, and 75% to 80% is common in the field [1].

Bulk deals do not fix this gap. Long-term deals cut server costs by only 30% to 50%, and Amazon still keeps a profit margin near 30% after them [1].

What strikes me here is the sheer size of the base spend. Flexera’s 2026 survey of 753 buyers shows 76% of large firms spend more than $5 million a month on public cloud. Wasted rented space hit 29% [4].

This pressure hurts firms with huge bills and steady loads. It helps the server hosts and bare-metal shops that charge by the physical rack.

One rule forces many of these moves: data fees. Moving data out of a cloud is a metered charge that grows with traffic. Cloudflare prices its own storage with no fee to move data out [6].

By the numbers

  • $3.2 million to $1.3 million — Yearly cloud bill: The fall after pulling seven apps off Amazon’s cloud is a saving of almost $2 million a year [3][5].

  • 81% — Cloud sales cost: The share of sales cost eaten by public cloud at one billion-dollar software firm [1].

  • 29% — Wasted rented space: The share of rented servers buyers admit goes unused in Flexera’s 2026 survey [4].

  • $0.015 per gigabyte-month — Free-transfer storage: Cloudflare’s monthly list price for storage that charges nothing to move data out [6].

The Cloud Bill Break-Even: When Renting Servers Stops PayingFigures as stated in this article’s own numbers section (verified figures, %)

Cloud sales cost81%([1])

Wasted rented space29%([4])

What I’d watch:

The teams running these moves skip the debate over whether the cloud is good or bad. I’ve been watching them sort apps by how wild their traffic swings, hunting for the exact fees that grow with use.

  • The use profile: A job running flat all day acts differently than a spiky web app. The flat job pays back owned servers, while the spiky one needs rented room.

  • The discount ceiling: A bulk deal buys only 30% to 50% off the server sticker price [1]. Teams weigh any deeper offer against the hard cost of buying physical racks.

  • The transfer line: Data fees grow fast as traffic spikes, unlike fixed server costs. Storage with free data transfers lists at $0.015 per gigabyte-month [6]. I would check that exact price against any massive cloud bill.

  • The saving pattern: 37signals cut its costs by half to two-thirds [2][3]. Dropbox told investors it saved $75 million in the two years before it went public [1].

The catch

Moving off the cloud is not free money. Owned gear wears out and needs a strict plan to replace it.

Someone still has to manage the physical room, cooling, and power limits. 37signals got its massive cut partly because it squeezed new servers into racks and power it already leased [3].

That 29% waste figure is a broad survey average, not a direct read on your own bill [4]. A team dealing with wild demand swings will easily beat physical servers on total cost. Dropbox needed two full years and a large staff to land its $75 million win [1].

My read: treating any single price as a strict law is a mistake. What holds across the sources is narrower. Renting steady servers carries a built-in fee that bulk deals never fully erase.

At a glance

  • The Big Shift: 37signals cut its cloud bill from $3.2 million to $1.3 million a year by moving seven apps onto its own gear [3][5], projecting roughly $10 million in savings over five years [2].

  • Why It Matters: Bulk deals only cut server costs by 30% to 50% [1], leaving firms paying a massive fee for unused room on steady apps.

  • What I’d Watch: How tech teams sort apps by traffic swings to find the exact break-even point.

  • Use profile: Whether an app runs flat enough to pay back physical servers within a standard planning cycle.

  • The discount ceiling: The 30% to 50% limit on server deals [1], which sets the exact price floor physical ownership must beat.

  • The transfer meter: Data fees that grow fast with traffic, compared to free-transfer storage priced at $0.015 per gigabyte-month [6].

  • The Catch: Owned servers wear out and demand strict power planning, and the best savings rely on fitting new machines into existing rack space [3]. The 29% waste metric [4] is an industry average, not a sure thing.

Related reading

Sources

[1] Andreessen Horowitz, “The Cost of Cloud, a Trillion Dollar Paradox” — https://a16z.com/the-cost-of-cloud-a-trillion-dollar-paradox/ [2] 37signals, “Cloud Exit” — https://basecamp.com/cloud-exit [3] David Heinemeier Hansson, “Our cloud-exit savings will now top ten million over five years” (17 October 2024) — https://world.hey.com/dhh/our-cloud-exit-savings-will-now-top-ten-million-over-five-years-c7d9b5bd [4] Flexera, “2026 State of the Cloud Report” — https://info.flexera.com/CM-REPORT-State-of-the-Cloud [5] Data Center Dynamics, “37signals claims it saved almost $2m last year from cloud repatriation” (19 October 2024) — https://www.datacenterdynamics.com/en/news/37signals-claims-it-saved-almost-2m-last-year-from-cloud-repatriation/ [6] Cloudflare, “R2 pricing” — https://developers.cloudflare.com/r2/pricing/