McKinsey Warns Rising AI Costs Could Wipe Out $850B

By Daniel Rao | 25 October 2026

Two of the world's most influential research firms have just priced the same AI boom, and they cannot both be comfortable reading. McKinsey says applying AI to work that already exists could be worth $2.52 trillion a year. Gartner says the industry is spending roughly $7 trillion to build the machines, and the revenue to pay for them falls nearly $5 trillion short. One of these numbers is a prize. The other is a bill. The gap between them is where the next three years of boardroom arguments will live.

McKinsey's October 2026 report, "The AI value curve: The economics of applying AI to existing work", is not another hymn to adoption. Its central complaint is blunt, and it is the sentence every executive should have to read twice: "Many organisations, however, are still measuring AI adoption instead of value." Companies can show how many staff use a chatbot, how many pilots launched, how many licences bought. What most cannot show is whether any of it made money after paying for the technology, the supervision, the checking and the rework. That is the question the report tries to answer, and its answer comes with a warning label attached.

Photo: A technician works on a server rack at the National Energy Research Scientific Computing Center (NERSC). McKinsey's calculation counts human supervision, verification and rework in AI's true cost, not just the software bill. Credit: Derrick Coetzee, CC0, via Wikimedia Commons.

The $2.52 trillion prize, and the fine print

Start with the headline figure, because it is genuinely enormous. McKinsey models AI generating $2.52 trillion in annual global economic value by automating and supporting existing work. Note the two verbs. This is not only about replacing workers with software. A large part of the prize, on McKinsey's account, comes from supporting people who do skilled, well-paid work — giving a doctor, a lawyer, an engineer or a teacher a capable assistant — rather than simply automating large volumes of lower-wage tasks. That distinction matters, because it is the opposite of the cheapest version of the AI story, in which value means headcount removed.

Then comes the fine print, and it is doing a lot of work. McKinsey is explicit that these estimates represent a modelled scenario rather than a forecast. Nobody at the firm is predicting that $2.52 trillion will arrive on a particular date. The figure is what the maths produces if AI is applied to existing work at scale, under stated cost assumptions. Change the assumptions and the prize changes — which is exactly what the report then demonstrates, in a way that should sober up anyone who has been quoting the $2.52 trillion figure without the caveat.

How a 30% cost rise eats a third of the value

Here is the number that gives this story its sting. If the effective cost of AI rises by 30%, McKinsey's modelled global value pool falls from $2.52 trillion to $1.67 trillion. That is a drop of about $850 billion, or roughly 34%. Read that again: costs up by less than a third, value down by more than a third. The asymmetry is the whole point, and it is not a rounding error.

It happens because economic value in this model is what is left after costs are paid. Push costs up and the thinnest use cases — the ones that were only just worth doing — stop being worth doing at all. They do not shrink gracefully; they drop out. McKinsey's calculation includes technology costs, human supervision, verification and rework, and every one of those lines has been moving in the wrong direction for heavy AI users: dearer compute, pricier power, more expensive specialist staff to check machine output, and rework when the output is wrong. A 30% rise across that bundle is not a disaster scenario. In an industry where electricity contracts, chip prices and engineering salaries all bite at once, it is a perfectly ordinary bad year.

This is why the report's recommendation is so unglamorous, and so easy to ignore. Businesses, it says, should assess changing AI costs as they go and plan to redeploy employees as work changes. No grand vision, no moonshot. Just unit economics, checked repeatedly, and a workforce plan for the jobs that move rather than vanish. The firms that do that drudge work will keep most of the $2.52 trillion. The firms that bought AI as a fashion accessory will discover that fashion is expensive.

Why America has the most to lose

The United States dominates both sides of this ledger. Under the same 30% cost-rise scenario, McKinsey puts the estimated American annual value pool at $640 billion, down from $960 billion — a $320 billion fall, the single largest national hit in the model. That is the price of being first and biggest: the US has the most AI deployed, the most data centres under construction, and therefore the most value exposed when the cost line moves.

For American readers, this is not an abstract global story. The cost pressures McKinsey models are the same ones showing up in US power bills, in the queue for grid connections, and in the salaries commanded by engineers who can keep AI systems honest. Washington's AI debate tends to run on geopolitics — who leads, who lags, who controls the chips. McKinsey's report drags it back to accountancy: leadership that cannot pay for itself is not leadership, it is a subsidy looking for an ending.

Where the money sits: banks, classrooms, hospitals

Not every sector shares equally in the modelled prize. Financial services leads, at an estimated $530 billion a year, followed by education at $410 billion, technology, media and telecommunications at $360 billion, and healthcare at $310 billion. Together those four account for $1.61 trillion of the $2.52 trillion total; the rest is spread across every other corner of the economy.

The ordering is instructive. Finance sits on mountains of documents, compliance checks and customer queries — exactly the kind of existing work that AI can chew through, provided a human still checks the answer before money moves. Education's second-place showing is the more surprising one, and it fits the report's emphasis on supporting skilled professionals rather than replacing them: a tutor that makes one good teacher more effective is worth more, in this maths, than a bot that replaces a call centre. Healthcare, at $310 billion, is the cautionary case — huge potential, but verification is not optional when a mistake harms a patient, so its cost line is the hardest to keep down.

None of this means a bank, a school or a hospital will bank those sums. These are sector pools, not company profits, and they assume competent deployment. A hospital that deploys AI sloppily will not collect its share of $310 billion; it will collect lawsuits. The sector table is a map of where the value could be, not a promise that anyone will arrive there.

Gartner's colder arithmetic: the $4.9 trillion gap

Now set McKinsey beside Gartner, because the two firms arrive at the same $7 trillion from opposite directions — and confusing them is the easiest mistake a reader can make this week. McKinsey estimates around $7 trillion could be invested in AI infrastructure by 2030, requiring approximately $760 billion in annual profit pools to justify the investment. That is an annual-return test: every year, the infrastructure needs to throw off roughly $760 billion in profit to earn its keep.

Gartner's October 14 report, "Prepare for AI Market Consolidation", runs a different, cumulative test. It puts AI infrastructure spending at nearly $1 trillion in 2026 alone, plus about $6 trillion more between 2027 and 2030 — roughly $7 trillion in total. To earn a 12% return on invested capital on that outlay, Gartner calculates the industry would need $8.5 trillion in cumulative revenue. Its forecast for enterprise and consumer AI revenue flowing to those data centres is no more than $3.6 trillion. The difference between what is needed and what is forecast is $4.9 trillion. That is not a predicted loss, and Gartner does not present it as one. It is the distance between the return investors would want and the revenue the market is currently on course to deliver.

Photo: The exterior of a Verizon data centre in Washington, D.C. (file photo; not a McKinsey or Gartner facility). The buildings behind the AI boom are being financed now, while the revenue to pay for them is still arriving. Credit: Tony Webster, CC BY 2.0, via Wikimedia Commons.

The two $7 trillion figures therefore agree on the scale of the build and disagree on what it takes to justify it. McKinsey asks for $760 billion a year in profit pools. Gartner asks for $8.5 trillion cumulatively at a 12% return. Both cannot be waved away, and neither rescues the other: a world that meets McKinsey's annual profit test could still fall short of Gartner's cumulative return test, and vice versa. What both tests agree on is that the current pace of spending needs an extraordinary payoff, and the payoff is not yet visible in the revenue lines.

The neocloud shakeout Gartner sees coming

Gartner's pain, it says, will not land evenly, and here its forecast gets specific in a way executives should not ignore. Hyperscalers such as Microsoft, Google and Amazon have cloud, advertising and retail cash flows to absorb years of AI losses. Neoclouds — the specialist GPU-as-a-service operators that exist to rent out chips and little else — do not. Gartner expects more than 80% of independent neoclouds to cease operating independently by 2030, swallowed by acquisitions, folded into partnerships, or simply shut down. More than $3.5 trillion of spending is at risk of impairment by 2030, rising past $4 trillion by 2032.

Will Sommer, a senior director analyst at Gartner, has put the underlying problem plainly in coverage of the report: although AI infrastructure investment is expanding rapidly, revenue from enterprises and consumers alone is unlikely to generate sufficient returns on all the planned capital spending. The Gartner report itself warns product leaders in terms worth quoting exactly: "Product leaders should not view AI capital write-downs as an infrastructure problem alone." Provider consolidation, price swings, model discontinuations and service rationalisation will hit any company that depends on AI supply, whether or not it owns a single server.

Read the two reports back to back and the mood whiplash is the story. McKinsey's prose is opportunity with a caveat. Gartner's title does not bother with subtlety. One firm models what AI could be worth if costs behave. The other models what happens to the people who built the infrastructure if revenue does not. Both can be right at once, and that is precisely why this week matters.

What this means for businesses buying AI now

Here is the uncomfortable read, and I will state it plainly: the industry has spent two years celebrating adoption and roughly no time proving value, and the bill for that botched priority is now arriving in two separate envelopes. McKinsey's envelope says a third of the prize vanishes if costs rise 30%. Gartner's envelope says the infrastructure bill needs revenue that does not yet exist. Any board still signing multi-year AI cheques on the strength of a usage dashboard, rather than a costed case per workflow, is volunteering to be the cautionary tale in somebody's next report.

The practical lessons are unglamorous but clear. First, price every AI workflow fully — compute, power, supervision, verification, rework — and re-price it quarterly, because the 30% scenario is one bad procurement cycle away. Second, favour deployments that make expensive professionals better over deployments that merely remove cheap labour; McKinsey's sector maths rewards the former. Third, avoid long, non-cancellable commitments to single-focus compute providers while Gartner's consolidation wave is still coming; today's bargain GPU contract is tomorrow's stranded supplier. Fourth, keep a human in the loop where errors cost real money — finance, healthcare, education — because verification is not overhead to be cut, it is the thing that keeps the value pool real.

We checked McKinsey's figures against three independent carriers of the October 24 ANI report and compared Gartner's numbers across four separate write-ups of its October 14 analysis before writing this article, and the cross-check held: the two firms' $7 trillion totals are real, but they measure different things, and any article that adds them together or swaps one for the other is misleading its readers. The honest summary is narrower and more useful. AI applied to existing work could be worth trillions — if, and only if, the people buying it keep its costs under control while the people building it wait for revenue that has not yet arrived.

For further context on the spending boom behind both reports, see our earlier coverage of UBS's warning that AI spending has reached 2% of US GDP, and on the data-centre build-out itself, Australia's $70 billion data-centre bet. Readers wondering whether this rhymes with a past mania should read what the Nasdaq's 78% fall after 2000 says about today's AI market.

Frequently asked questions

Is the $2.52 trillion a forecast?
No. McKinsey describes its figures as a modelled scenario, not a forecast. The $2.52 trillion is what AI applied to existing work could be worth annually under the report's cost assumptions. If costs rise, the modelled value falls — to $1.67 trillion under a 30% cost increase.

What counts as the "effective cost" of AI?
More than the software licence. McKinsey's calculation includes technology costs, human supervision, verification of machine output, and rework when output is wrong. All four must be paid before any net value exists.

Why does a 30% cost rise cut value by 34%?
Because value here is what remains after costs. When costs rise, the most marginal AI uses stop paying for themselves entirely and drop out of the pool, so the pool shrinks faster than costs grow. That leverage effect is the report's central warning.

How do McKinsey's and Gartner's $7 trillion figures differ?
They describe similar total infrastructure investment by 2030 but test it differently. McKinsey asks for roughly $760 billion in annual profit pools to justify the spend. Gartner asks for $8.5 trillion in cumulative revenue to earn a 12% return on invested capital, against a forecast of no more than $3.6 trillion. Same build, different yardstick — and a $4.9 trillion gap on Gartner's test.

Which sectors stand to gain most?
On McKinsey's modelling: financial services at $530 billion a year, education at $410 billion, technology, media and telecommunications at $360 billion, and healthcare at $310 billion. These are sector-wide pools assuming competent deployment, not profits any single company can claim.

Sources

McKinsey, "The AI value curve: The economics of applying AI to existing work" (October 2026), as reported by ANI on 24 October 2026 and carried by The Economic Times, The Tribune and AI Affairs; Gartner, "Prepare for AI Market Consolidation: Capital Write-Downs Could Exceed $4 Trillion by 2032" (14 October 2026), as covered by Crypto Briefing, Startup Fortune and XenoSpectrum. Sector, cost-scenario and infrastructure figures are attributed to McKinsey or Gartner respectively in the text; neither firm's figures have been combined into a single total.

Daniel Rao covers AI economics and enterprise technology for TechPulse USA. He writes about what AI costs, who pays for it, and whether the returns ever arrive.