The trillion-dollar question: how much value must AI create to pay for its build-out?
Spending on AI data centres is approaching $1 trillion a year. We size what would have to be true for it to pay off and show why returns will hinge less on whether the value arrives than on who captures it.
Autheurs
The build-out of AI infrastructure is one of the largest private investment programmes in modern history. Spending on AI data centres, which covers the chips, networking, land, buildings and power behind services like ChatGPT, is around $1 trillion in 2026 on our estimates. This is towards the upper end of published figures and rising fast. Nvidia, the largest supplier into the build, expects annual spending of $3-4 trillion by 2030.
Whether all this will pay off has become the defining debate in markets, usually conducted by analogy: optimists cite electrification and the railways, sceptics the telecoms fibre bubble of 2000. Analogies cannot settle it, because people pick whichever one suits their conclusion.
There is a more answerable question lying underneath though: how much value would AI have to create for the spending to make sense? That can be sized with simple arithmetic and a handful of transparent assumptions. The answer does not tell you whether the cycle succeeds. It tells you how high the bar is, where the value would have to come from, and what to watch as the evidence arrives.
One dollar of capex, a chain of hurdles
AI capital expenditure (capex) sits inside a value chain with three layers. The data centre owner, typically a hyperscale cloud provider, earns infrastructure revenue. Companies in the model and software layer, such as OpenAI and Anthropic, add a markup, while the end customer pays the full bill, and must extract enough value from AI to justify it. Each layer has a return hurdle, and the hurdles compound.
Three assumptions drive the arithmetic: the data centre owner targets a 15% return on capital at a 30% operating margin, in line with mature cloud businesses; the model and software layer charges 1.5 times the underlying infrastructure cost, conservative against current closed-model pricing though a growing open-weight mix could compress it over time. The customer requires a 15% return on their total cost of adoption, including implementation. None of these are facts; they are dials, and the conclusions are reasonably robust to sensible alternatives.
Work through the chain and each $1 of annual capex requires around $0.63 of annual infrastructure revenue, which becomes a customer bill of around $0.94 after the markup, which must in turn generate around $1.19 of annual value for the customer. That $1.19 is a gross figure, measured before the customer pays the AI bill and the cost of putting AI to work.
The rule of thumb: every $1 trillion of annual data centre capex needs around $1.2 trillion of annual customer value
The same hurdle can be expressed over the life of the asset rather than in any single year. Assume a blended data centre life of roughly ten years: shorter for the chips, much longer for the buildings and power. On that basis, and allowing for depreciation along the way, each dollar of capex must ultimately be matched by roughly $6 of customer value. That is the real test: a dollar spent on AI infrastructure today is not justified by one year of demand, but by whether it supports enough valuable use over its life, before the assets it bought lose their economic relevance.
Each $1 of capex needs $1.19 of customer value a year
From one year's spending to the whole build-out
The framework scales linearly, so the hurdle compounds with the capex. We grow spending at 30% a year from the $1 trillion 2026 base, which lands at roughly $2.9 trillion by 2030, the lower end of Nvidia's estimate, but by no means guaranteed.
The hurdle compounds with the capex (30% a year growth assumed)
The chart above asks what each year's new spending must earn. The bigger question is what everything built across 2026-2030 must earn, all at once: cumulative capex of around $9 trillion. So, when should that estate be judged? Not in 2030, when the assets are on average around two years old and customers are still part way through deploying them; demanding a full return then would be like judging a factory on its first year of output.
The fair test comes once the estate is mature and earning, which the lag we describe below puts at around 2033-34. By then, depreciation has reduced the net capital base from nearer $7 trillion at the end of 2030 to roughly $4.5 trillion, and the whole estate must generate around $5.4 trillion of customer value every year.
To put that number in context: on our estimates it is roughly 20% of today's global corporate operating profit and around 4.5% of world GDP; against the larger economy of the early 2030s, the ratios shrink. Demanding, but not impossible.
Where could $5.4 trillion a year come from?
If the AI bill simply came out of existing corporate profits, the cycle would be a transfer, and someone would be paying for it. But the corporate pie is not fixed. There are three ways the bill can be covered by value AI itself creates: cost substitution (AI replaces existing labour and process cost), revenue acceleration (companies grow faster than they otherwise would), and entirely new value pools that do not exist yet.
These pools can be sized. The activities AI can plausibly address today, from software development and IT services to customer support and back-office knowledge work, gross to more than $26 trillion. Apply plausible capture shares and the value AI could realistically extract is around $2.2-3.7 trillion a year, in line with published estimates of generative AI's potential. The pools we can name today fund roughly half the requirement.
The pools we can name today fund roughly half the requirement
The remainder comes from revenue acceleration, which is additive to these pools, from capture shares rising as capability improves, and from categories that do not exist yet; the pools also exclude consumer AI spending entirely, so the base could be conservative. A useful stress test is to ask what each channel would have to deliver if it carried the entire bill alone. Put that way, the numbers sound heroic: cost substitution alone would mean displacing work equivalent to around 11-12% of the global knowledge workforce; revenue acceleration alone would mean growing global corporate revenue by an extra 16% or so within a decade.
Expressed as rates, however, the numbers shrink: the full revenue ask works out at around 1.5 percentage points of faster global corporate revenue growth, sustained into the mid-2030s, and half that if substitution carries half the bill. History also suggests the third channel, the pools nobody can name yet, may end up the biggest. The late-1990s fibre build-out was justified on web pages and email; but the demand that eventually filled it, smartphones, streaming and cloud, was largely unimaginable in 1999. That is a reminder rather than a forecast. No single channel has to carry the bill alone, and the blend brings each ask back inside what past technology revolutions actually delivered. What none of them delivered, as the final section returns to, is anything like the speed required here.
Bills are payable now; value arrives later
One structural feature of the cycle deserves more attention than it gets: the long, unavoidable lag between a dollar of capex and a dollar of customer value. A data centre takes 18-24 months from groundbreaking to power-on. Once live, capacity fills almost immediately, and mostly not with mature enterprise demand. Model training and consumer AI services absorb capacity from day one, and the fastest-growing category is tools that workers and teams adopt directly: AI coding is already one of the largest workloads in the system, increasingly in agentic form, where the AI plans and executes whole tasks. The lag applies to the deeper deployments that must ultimately carry the framework, where AI is wired into a company's processes: a year or two to deploy and change ways of working, longer still for benefits to reach financial statements. Stacked together, a capex dollar spent in 2025 may not appear as customer value until 2028 or later.
The cycle will therefore repeatedly look unjustified before it can possibly look justified, even in scenarios that work. Bills are payable now; value materialises later. Investors who do not internalise this will misread the data on the way through.
The number to watch
All of this can be reduced to one observable test. We estimate customers today pay roughly $300 billion a year for AI services across cloud AI, the model layer and AI-native software. The framework requires that bill to reach around $4.2 trillion at maturity: growth of roughly 45% a year, sustained for seven years. The estimate of today's bill is necessarily imprecise, because disclosure is patchy, but the conclusion barely moves across any sensible range.
We have written previously about where AI revenue surfaces across the technology stack, and what the early signs of payoff look like; the framework here sizes what those revenues ultimately have to become.
Current growth clears the bar, though less comfortably as the base broadens, and unlike long-dated capex forecasts, fresh evidence arrives every reporting season. If AI services spending keeps growing at or above roughly 45% a year, the bill is catching up with the build; if services growth decays while capex keeps compounding, the gap is widening. Composition matters as much as pace: revenue paid out of customers' measured returns counts for more than revenue financed by the ecosystem that is building the infrastructure.
So, is the spending rational?
The framework does not answer yes or no. What it shows is that the bar is high, rising, and yet not without precedent. The scale of investment, productivity gain and labour reallocation implied all have analogues in electrification, the railways and the computing revolution, each of which reshaped a comparable share of the economy over a couple of decades. The one element genuinely at the edge of precedent is speed: the revenue ramp required would be among the fastest large-scale technology revenue ramps on record. The counterpoint is that diffusion has accelerated with every technology generation, and AI is the first to arrive on infrastructure that already exists: the devices, networks and habits were built in the internet era.
Three conditions need to hold. Capex growth must eventually moderate, so the hurdle stops compounding; if 30% growth ran another five years beyond 2030, the requirement would push toward half of today's global corporate operating profit. Capacity must get built and powered on schedule, since construction, power and semiconductor supply, not demand, are the binding constraints today. And customer value must arrive before the financing built around the assets needs to be rolled.
Which points to where the answer will actually come from. It will not be settled by forecasts, or by analogies to railways and fibre. It will be settled by a bill: whether customers keep choosing to pay it, year after year, at the pace the build-out requires. For now, they are.
For investors there is a sharper question underneath, because the aggregate can clear while individual industries shrink. AI will deflate revenue in some sectors even as it accelerates others, and some layers of the chain will earn far more than their hurdle, while others never reach it. The same evidence that tracks the cycle also discriminates within it: the framework sets out what each layer must deliver, so as the evidence accumulates, it shows which businesses are clearing their hurdles and which are falling short.
That is where this framework meets our team's philosophy of finding growth gaps: companies where the market's assumptions about future growth prove too low. Whether AI creates the value will decide if the build-out was rational. Who captures it will decide returns.
Subscribe to our Insights
Visit our preference center, where you can choose which Schroders Insights you would like to receive.