The usual rule says owned servers only pay off above 60–80% utilisation. At 2026 prices the line sits nearer 52%, and most established firms are paying for flexibility they never use.
Every few months someone publishes a spreadsheet showing that leaving AWS would save them millions, and someone else replies that the spreadsheet forgot about on-call rotations, hardware failures and the cost of being wrong about demand. Both sides are usually arguing about the same thing without saying so: whether the price premium the cloud charges is worth what it buys.
That is an old question in corporate finance. Leasing versus buying, capacity investment when demand is uncertain, the value of waiting before committing capital: all of it has decades of theory behind it. For my master's thesis I built a decision model for the cloud question on that basis, calibrated it only to public 2026 prices, and then checked it against the two companies that left the cloud and published their numbers, Dropbox and 37signals.
This post walks through what it found. The short version:
The model compares two full cost stacks for the same workload over ten years and discounts both at the firm's cost of capital. What makes it more than a TCO spreadsheet is the two layers on top: demand that moves randomly instead of following a forecast, and an option to wait before exiting rather than deciding once.
The unit of capacity is 32 vCPU with 128 GiB of memory, which is eight AWS m7i.xlarge instances. Rented on three-year terms it costs about €5,500 a year. Owned, it is €3,600 of server plus €320 a year of colocation. The reference firm needs 500 units (16,000 vCPU, a mid-sized software company), grows 10% a year, has a 14% cost of capital and pays Bulgarian corporate tax.
I tilted the calibration against ownership on purpose. The cloud gets its cheapest price; the owned fleet gets triple the operations staff and a quarter of its capacity left idle as a buffer. If ownership still wins under those terms, the result is not coming from favourable assumptions.
It still wins. Over ten years, serving the workload from the cloud costs €24.1M in present value; owning it costs €15.6M. The difference, €8.5M, is the 54.6% premium.
The useful question is not whether ownership wins at one utilisation, but where it stops winning. Pull the slider to change the cost of capital and watch where each curve crosses zero.
The crossing barely moves. At an 8% cost of capital the breakeven is 49%; at 20% it is 55%. Discounting shrinks the prize but never changes the sign, because the two strategies spend their money on similar enough schedules.
The intuition is simple once the premium is on the table. If renting costs 1.55 times what owning costs, you can buy nearly twice the capacity you need and still come out ahead. A fleet that sits idle half the time already beats committed-price cloud.
Growth is a different matter. The advantage is almost flat between zero and 25% annual growth, then collapses: €2.0M at 40% and −€9.3M at 50%. A firm growing that fast has to buy hardware for demand it does not have yet, carries it idle for years, and pays more for it than it would a year later. Hypergrowth companies belong in the cloud, and the reason is the arithmetic of buying ahead of demand rather than volatility.
The deterministic answer assumes you know your demand. The stochastic layer drops that: demand follows a random walk with 20% annual volatility, the fleet is sized in advance on the forecast, and anything above capacity bursts to on-demand cloud at the on-demand price. Forecast errors fall on the owner. Run 5,000 times:
The average advantage falls from €8.5M to €7.4M, which is the cost of being wrong about demand. The spread is wide, but the cloud comes out cheaper on fewer than one path in a hundred. Volatility has to rise far beyond what a stable business sees before it becomes a coin toss: at 35% ownership still wins 87% of the time, and at 50% (startup-grade uncertainty) it wins 65% of the time, though by then the expected gain of €2.0M is small for the commitment involved.
The best argument for paying the premium is that it buys an option: the right to scale up, scale down or walk away without owning anything. Options have a price, and the price rises with uncertainty. So the right test is to ask what premium a firm with a given volatility should rationally pay, and compare it with what it does pay.
The shape is what option theory predicts: the more volatile the demand, the more flexibility is worth, by about 35 percentage points between 5% and 50% volatility. The level is the surprise. The fair premium is negative across the whole range. Even a firm with startup-grade volatility would need the cloud to be cheaper than owning before renting everything made sense.
That is because pure cloud is not the alternative to pure ownership. Own the stable base, rent the peaks on demand, and you capture most of the option's value without paying the premium on all of your capacity. Against that hybrid, the full premium buys very little. The workloads where it is defensible are the ones that might disappear entirely, where the valuable option is to abandon rather than to scale.
One thing this does not price: managed services. Renting a managed database is not the same product as renting a virtual machine, and part of what the premium buys is real operational work someone else does. The 37signals case below gives a sense of how large that part is.
If the premium is this large, why did firms only start leaving around 2021? Because it was not this large ten years ago. Cloud list prices have barely moved in a decade, while the hardware underneath kept getting cheaper.
S3 Standard has cost $0.023 per GB-month since December 2016. A 4-vCPU general-purpose instance rented for $0.200 an hour in 2015 and its same-shape successor rents for $0.202 today; the newest generation lists about 5% higher. Disk cost per gigabyte, meanwhile, fell about 9% a year from 2017 to 2026, and server compute improved at a similar pace. Some of that improvement reached cloud customers as faster chips at the same price, at most about 5% a year.
Put together, the premium drifted up by roughly 5–10% a year, which compounds to a factor of 1.6–2.6 over the decade. A workload for which renting and owning cost the same in 2016 faces a 60–160% premium today on price drift alone. The order in which firms left fits that drift: Dropbox first, at enormous scale, then mid-sized firms like 37signals as the line moved past them.
There is a caveat at the end of the series. The AI build-out has stalled hardware deflation. Disk prices have fallen under 4% a year since 2022; Western Digital said early in 2026 that its drive output for the year was essentially sold out to cloud customers; server memory rose about 15% between August and October 2026. If scarcity pricing lasts, 2026 may be the high-water mark of the ownership case for storage-heavy workloads. If the cycle clears the way previous ones did, the divergence resumes. Either way, the decade's evidence is that cloud prices do not follow hardware costs down.
Every input in the model has a range as well as a central value. The chart moves each one, alone, to both ends of its range.
No single parameter overturns the result. The two that matter most are the cloud unit price, which is where deep enterprise discounts would show up, and the ownership staffing ratio. They are also the least transparent number on the cloud side and the most judgement-based number on the ownership side. The things people argue about most (migration cost, salvage value, colocation rates) are second-order: even at the bad end of their ranges they cost €1.2M, under €0.1M and €1.7M respectively, out of an €8.5M advantage.
Move every parameter to its cloud-friendly end at once (the deepest discount, cheap managed operations, expensive in-house staff) and the cloud wins, by 34%. Move them all the other way and ownership wins by 298%. So the conclusion depends on conditions, and they are conditions a firm can measure from its own books.
The model never saw either company's data. It was calibrated on public price sheets and then compared with what the companies disclosed.
Dropbox moved over 90% of its user data off AWS onto its own storage fleet in 2015–16. Its 2018 IPO filing breaks down the 2016 change: payments to the cloud provider fell by $92.5M while the cost of the owned fleet rose by $53.0M. Replacing $92.5M of cloud with $53.0M of ownership is a 43% reduction. The model's prediction follows from the premium alone, since a premium of π means an exit saves π/(1+π) of the bill: 35%. It misses by seven points, on the conservative side, for a company that is storage-led, runs custom hardware at exabyte scale and looks nothing like the calibration basket.
37signals published the full decision arithmetic for leaving AWS in 2022–23: a $3.2M annual bill, about $0.7M of replacement servers for roughly 4,000 vCPU, and an estimate of $7M saved over five years, with no change in headcount.
| 37signals compute exit | Model | Disclosed |
|---|---|---|
| Fleet capex | $0.5–1.0M | ≈$0.7M |
| 5-year savings, undiscounted, pre-tax | $6.0M | $7M |
| 5-year savings, after tax, discounted at 14% | $3.4M | — |
| Implied cloud premium at their bill | 131% | — |
Two things came out of this that I did not expect. First, 37signals' actual bill per unit of capacity was 2.4 times the committed-price basket the model uses. The difference is managed databases and search, egress, and reservations that were not fully used: things real bills contain and pricing calculators leave out. The usual objection is that list-price comparisons are unfair to the cloud because nobody pays list. For a mid-sized firm the omissions outweigh the discounts, so the 54.6% premium is a floor.
Second, the headline number shrinks under proper accounting. $7M is an undiscounted, pre-tax difference in bills. After tax and discounting, the same facts give $3.4M. The exit is still clearly worth it, but at roughly half the advertised figure, and the same correction applies to vendor TCO studies in either direction.
Accelerators change the arithmetic. An eight-GPU H100-class server costs around $285,000, about seventy times a general-compute unit, it is obsolete in three years, and it draws ten kilowatts. And there are two rental markets, not one.
| Renting 8× H100 from | $/GPU-hr | Breakeven |
|---|---|---|
| Specialist cloud, low | 2.29 | 69% |
| Specialist cloud, median | 2.70 | 58% |
| Hyperscaler, reserved block | 5.19 | 30% |
| Hyperscaler, on-demand | 6.88 | 23% |
| Hyperscaler, high | 10.98 | 14% |
Owned: $285k capex, 3-year refresh, 20% salvage, $24k/yr colocation at 10.2 kW, $10k/yr operations. Rental prices as of October 2026 (CloudZero; Ornn; Thunder Compute; Converge Digest).
Against specialist GPU clouds the breakeven is higher than for general compute, 58–69%, because a three-year refresh runs the capital clock nearly twice as fast. But the spread between rental markets is the real finding: the same GPU-hour costs $2.29 or $10.98 depending on who you rent it from. For AI workloads, which rental market you can get into matters more than whether to rent or own. Renting from a hyperscaler at list price is beaten by the specialist market at every utilisation, and by ownership above about a quarter.
The decade-long drift does not apply here either. AWS cut on-demand H100 prices by about 44% in mid-2025 as the next generation arrived, then raised its reserved GPU capacity rates twice in 2026, by about 15% in January and 20% in July. GPU prices move within months, in both directions, so the comparison has to be redone on the day of the decision.
The cost stacks scale linearly, so on paper a 50-unit firm faces the same 54.6% premium as a 500-unit one. Reality has a fixed cost the linear model hides: someone has to be on call. Three-quarters of an engineer cannot run a data centre. Putting a floor under the operations team gives:
| Units | ≈ vCPU | No floor | 2-person team | 3-person team |
|---|---|---|---|---|
| 25 | 800 | +€0.43M | −€0.31M | −€0.79M |
| 50 | 1,600 | +€0.85M | +€0.34M | −€0.14M |
| 100 | 3,200 | +€1.70M | +€1.64M | +€1.16M |
| 250 | 8,000 | +€4.25M | +€4.25M | +€4.25M |
| 500 | 16,000 | +€8.50M | +€8.50M | +€8.50M |
The threshold lands between roughly 30 and 60 units, or 1,000–1,900 vCPU of steady load. In money, that is about €0.2–0.4M a year of committed cloud spend. Below it, an exit cannot pay for its own operators. Above about 250 units the floor never binds, and the argument about "cloud economics at scale" stops being about scale at all. It becomes a question of utilisation and price.
That threshold is low. It rules out startups and small firms, as you would expect, but it takes in far more companies than the exabyte-scale examples in the public debate suggest. A firm spending half a million euros a year on committed cloud capacity is, at 2026 prices, probably overpaying.
The cloud debate has been run as a matter of engineering taste, when it is a lease-versus-buy problem under uncertainty with switching costs, and standard finance gives it quantitative answers you can test. Here they say three things. The utilisation needed to justify owning is lower than the industry believes. The premium is far above anything flexibility can justify for a firm whose demand is reasonably predictable. And the gap opened over the last decade because cloud prices stopped following their own costs down.
None of this makes leaving the cloud the right move for everyone. Hypergrowth, small scale, workloads that may vanish, and GPU work in a fast-moving market all point the other way, and the model says so. What it replaces is the argument by anecdote: with a breakeven line, a firm can put its own numbers in and see which side it is on.
Updated 5 October 2026. First published in May 2026. Every figure now comes from the current model and a calibration whose prices and sources were re-checked on 5 October 2026, including the latest GPU rental rates and the corrected hard-drive price series.
The thesis is a full draft, due in January 2027 and defended in February. A few inputs are still open before the calibration is frozen: European colocation rates (the current ones are North American), an Azure and GCP replication of the compute basket, and a re-based EUR/USD rate, which on its own would move the premium from 54.6% to 57.0%. If you run infrastructure and have numbers that disagree with these, I'd like to hear from you: teodor@azbouki.com.