The Downtime Stat That Fails Its Own Arithmetic [2026]

Oct 2

The Pipe

Predictive Maintenance vs. the Breakdown Economy: Nobody Measured the Before

The business case says the sensors paid for themselves. Against what baseline?

Written by Scot Free

Every predictive maintenance business case in the world is the same multiplication: hours of downtime avoided × cost per hour of downtime. Put sensors on the asset, catch the bearing before it seizes, multiply the hours you didn't lose by what an hour costs, and the number is large enough that nobody asks a second question.

This piece asks the second question. Twice, once for each term.

Term two: cost per hour

Start with the easier one, because it's the one everybody quotes and almost nobody sources.

Here is what the trade press reports as the cost of an hour of unplanned downtime in manufacturing, all published within the last year, all presented as established fact:

Figure	Attributed to	Scope as stated
$125,000/hr	ABB, Value of Reliability survey, 2023	Median across industrial sectors
$260,000/hr	Aberdeen Research — though several outlets credit it to Siemens 2024 instead	Average across all manufacturing sectors
$532,000/hr	A maintenance-software vendor citing "Siemens and Aberdeen Group research"	Average across all sectors
$50,000–$260,000/hr	"Widely cited industry research"	Range by sector
$2.3M/hr	Siemens, True Cost of Downtime, 2024	Large automotive plant

Figures as published by the outlets listed in Sources. The point is the spread, not any one number.

"Average across all manufacturing sectors" appears three times in that table attached to $260,000, $532,000, and a range starting at $50,000. Those are not three measurements of the same thing. A better-than-4x spread on an identically scoped claim means the scopes are not actually identical, and the labels aren't telling you how they differ.

Two more things are worth noticing in the same literature:

The population being averaged can't measure the thing. One benchmark roundup notes, in passing, that a majority of manufacturers cannot state their true downtime cost per hour. Another observes that unplanned downtime is a cost most plants can name in general terms but rarely quantify precisely. If most respondents can't compute it, the industry average is an average of estimates, carrying every estimator's incentive with it.
Independent confirmation gets manufactured. One vendor page argues that Aberdeen's independent benchmark of about $260,000 an hour sits alongside its own per-hour estimate, so two separate methods land in the same range. The vendor's own number is being offered as corroboration of the figure it was calibrated against. That is not two methods agreeing. That is one number with a second name.

The headline number fails its own arithmetic

The most-repeated statistic in this entire category is Siemens' True Cost of Downtime figure, and it usually appears in one sentence with two halves:

The world's 500 largest companies lose approximately $1.4 trillion annually to unplanned downtime — equivalent to 11% of their total revenues.

Check it against Fortune's own published aggregate. The Fortune Global 500 generated $41.7 trillion in combined revenue on the 2025 list, and $43.1 trillion on the 2026 list.

$1.4 trillion ÷ $41.7 trillion = 3.4%, not 11%.
11% of $41.7 trillion = $4.6 trillion, not $1.4 trillion.

Both halves of that sentence cannot be true of the Fortune Global 500. One of them is describing a different denominator — most likely a surveyed subset of heavy industrial firms rather than the full list, which would be a defensible finding stated plainly and a misleading one stated as "the world's 500 largest companies."

The failure isn't Siemens'. A survey of industrial plants that reports 11% of those firms' revenue is doing honest work. The failure is in the repetition: by the time the stat has been through a dozen trade articles, the subset has become the Global 500, the percentage has been welded to the dollar figure, and nobody down the chain has divided one by the other.

This is the first thing to do with any number in a business case: divide it by its own denominator. It takes nine seconds and it is the single highest-yield audit step in the category.

Term one: hours avoided

The second term is worse, because it's unobservable by construction.

Avoided downtime is a counterfactual. You are claiming credit for a failure that didn't happen. There is no sensor reading for the outage you prevented, and the only way to know whether the hours actually fell is to compare against a measured before-state.

Three questions decide whether that comparison is real:

Was the baseline measured or estimated? If the before-state downtime figure was produced during the business case, by the people proposing the project, it is not a baseline. It is an input to a sales argument. A real baseline comes from the CMMS work-order history, the production system's stop records, or the OEE availability data, pulled for a defined period before anyone proposed anything.
Did anything else change in the same window? Predictive maintenance projects rarely arrive alone. They come with a new CMMS, a reliability engineer, a PM schedule refresh, and a round of management attention on the asset. Any one of those reduces downtime. Attributing the whole delta to the model is the oldest attribution error in operations.
Is the alert actually acting as a sensor, or as a reason to look? This one is uncomfortable and usually true. Much of the early value of a predictive program is that somebody finally inspects the asset regularly. That is real value. It is not model value, and it does not require the model to sustain it.

If the baseline was estimated, the window had three other changes in it, and the mechanism was attention rather than prediction, the ROI number isn't wrong exactly. It's unfalsifiable, which is worse.

What the breakdown economy actually costs

Here is the turn, and it cuts the other way.

Reactive maintenance has real costs that the comparison usually misses too, and they're missed because they live in other people's budgets:

Expedited freight on the part you didn't have, which lands in logistics, not maintenance.
Overtime to recover the schedule, which lands in production labor.
Scrap and work-in-process lost in the stop itself, which lands in quality or yield.
Spares carried against uncertainty. A plant that can't predict failures holds inventory it wouldn't otherwise need. That's working capital, and it sits on the balance sheet rather than the P&L, so it never appears in a maintenance comparison at all.
Planner and supervisor time consumed by firefighting instead of scheduled work.
The reliability tax on commitments. Plants that break down unpredictably quote longer lead times and hold more buffer stock. Customers pay for that, and so does the plant.

One trade analysis puts the fully loaded cost of downtime at two to three times the direct production loss that most plants use internally. Directionally, that matches what anyone who has traced the charges across departments would expect. It also means the honest finding in most predictive maintenance programs is not "the ROI was inflated."

It's that both sides of the comparison were estimated, and the actual gap between them was never measured. The claimed benefit was a counterfactual. The baseline it was measured against was an estimate. And the real costs of the status quo were sitting in four other cost centers where nobody added them up.

What a defensible business case looks like

Everything above is fixable, and none of it requires better models. It requires the measurement to be designed before the sensors go in, not reconstructed afterward.

Pull the baseline from a system nobody is trying to persuade. Work-order history and production stop records for twelve months prior, pulled before the proposal, by someone with no stake in it. Freeze it and write it down.
Define cost per hour from your own P&L, not a benchmark. Lost units × contribution margin, plus the four line items above traced to their actual cost centers. Your number will be smaller than $260,000 and it will be defensible, which is the trade worth making.
Leave a control group. Instrument eight of twelve identical assets. The four uninstrumented ones cost you nothing and convert the whole program from a testimonial into a measurement. This is the single step most programs skip and the only one that isolates the model's contribution from the attention effect.
Count the full run cost. Licenses, connectivity, the platform, and the engineer who maintains the pipeline when the firmware changes and the units shift. If you've read The Lakehouse at the Edge, you know that last one is not hypothetical.
Book the harvest, or don't claim it. Avoided downtime becomes money only if something changes: a shift not added, a line not duplicated, inventory actually reduced, a maintenance req not backfilled. Name the thing, name the owner, and check it at twelve months. Capacity that nobody harvests evaporates, and next year somebody asks where the savings went.

The Underground Take

You've sat in the room where these numbers get approved. The ones that survive aren't the biggest, they're the ones where somebody can say where the number came from. And the question that kills most of them isn't "is that too high?" It's "is that cost leaving the budget, or is that capacity?"

Sagan's rule was that extraordinary claims require extraordinary evidence, and the practical half of that rule is independent confirmation. A vendor citing an analyst who is cited by the vendor's own marketing is not independent confirmation; it's a circle with two labels on it. The benchmark literature in this category is full of those circles, and they are easy to spot once you start dividing numbers by their own denominators.

Deming's version is shorter and harder. You cannot improve what you have not measured, and a baseline produced by the people proposing the project is not a measurement. Predictive maintenance is real engineering, and the physics works. The arithmetic around it is where the discipline goes missing. Measure the before. Then believe the after.

Next in The Pipe: what actually converts the telemetry into something — the Lean 4.0 question.

Sources

Siemens / Senseye, The True Cost of Downtime, 2024 — $1.4 trillion across the world's 500 largest companies, stated as 11% of revenues, up from 8% in 2019; $2.3M/hour for a large automotive plant; ~$36K/hour FMCG; 25 incidents and 27 hours per month at large plants; 62% rise in downtime cost since 2019 as incident counts fell. As reported by the trade outlets below; the primary report was not independently obtained for this piece.
Fortune, Global 500 rankings: $41.7 trillion combined revenue (2025 list, FY2024) and $43.1 trillion (2026 list, FY2025).
ABB, Value of Reliability survey, 2023 — ~$125,000/hour median across industrial sectors, as reported by Acronis and TeepTrak.
Aberdeen Research — ~$260,000/hour, as reported by ReliaMag, Verdantis, and others. Note the inconsistent attribution of this figure across outlets.
Verdantis, Unplanned Downtime: Cost and Recovery Insights, 2026 — the "two separate methods land in the same range" framing discussed above.
iFactory, downtime benchmark articles, 2026 — $532,000/hour all-sector figure; the "fully loaded cost is two to three times the direct production loss" estimate; the $125,000/hour 2026 benchmark.
FactoryMetrics, Manufacturing downtime statistics 2026 — the $50K–$260K range, and the observation that a majority of manufacturers cannot state their true downtime cost per hour.

A note on these sources: nearly every figure above is a vendor or analyst estimate, several are repeated between outlets without independent verification, and the attribution chains do not always survive inspection. That is the subject of this article, not a defect in it. Where this piece cites a number, it is citing what was published, not endorsing the method behind it.

Corrections and additions: Scott@IoTunderground.com.

Scott Tower