On-demand pricing sells an option, and the premium is charged whether or not you exercise it. The dev.to analysis works the case on a database held at a steady 60% utilization every day for three years [7]. Three years of hourly billing is 26,280 hours [1]. The right to hand that capacity back was available in all of them and used in none.
Read the reported saving band as a return per dollar of that workload's cloud line: 30 cents at the low end, 60 at the high end [4]. Then set it against the items the invoice bundled into one number [8]. At the low end, each dollar of owned-side cost you discover after the move needs $3.33 of that workload's annual cloud spend to offset it; at the high end, $1.67 [2]. A workload too small to carry an operator fails that test no matter how flat its utilization curve is.
Where I would push on the band is its definition. The article calls it lower infrastructure spend and does not say whether the organizations reporting it counted operating headcount inside that measurement or beside it [14], while the same piece names engineering time as the most expensive unlisted cost [8]. If those figures came off invoices, the band is gross, and the number you would actually book sits below it.
37signals remains the most-cited case: roughly $7 million over five years after leaving the public cloud [3], or about $1.4 million a year [3]. The article publishes no breakdown of what the total includes [15]. For that figure to transfer you need their load shape and a team that was already going to run hardware, which makes it evidence that the sign can be positive at that scale rather than a coefficient for your own model.
The placement rule in the piece is the part that survives contact with a spreadsheet. Steady round-the-clock compute, large stable datasets whose egress charges exceed their compute, GPU and ML inference at volume, and data under hard residency rules are the candidates [12]. Spiky or seasonal traffic, pre-product-market-fit work, and managed services that stand in for a full-time operator stay where they are [11]. Most teams land hybrid, with the steady and the sensitive on owned or colocated hardware and the unpredictable left where elasticity is worth its margin [10].
That narrows the answerable question. For one workload, over the life of the hardware you would buy, does 30 to 60% of its current cloud line cover the machines, the space, the spares and the people who would run them. A no there is also information: it says the premium was buying capacity you would otherwise have had to staff, which is what it was sold as. The survey numbers cannot answer that question for you, because intent aggregates across a company and savings do not.
