Artificial intelligence could transform the economy without delivering the returns expected by everyone financing it. As governments and technology companies seek vast amounts of capital, higher borrowing costs are sharpening the distinction between a useful technology and a profitable investment.
The artificial intelligence boom could fulfil its technological promise and still disappoint some of the investors paying for it. Cheaper computing would help businesses adopt the technology. It could also reduce the income earned by infrastructure bought at yesterday’s prices.
That tension becomes harder to manage when money costs more. Data centres and processors must earn enough to cover electricity, maintenance, equipment replacement and financing. Growing demand helps, but the return depends on the price customers pay and the capital required to serve them.
On September 14, the benchmark US Treasury yield for bonds maturing in 10 years crossed 5 per cent. Reuters attributed the rise to renewed inflation concerns, heavy debt issuance and expectations of continued economic resilience. Investors financing uncertain future earnings now have a more remunerative alternative.
Governments are seeking substantial funding too. The question is how the price of capital will shape the industry being built, and who captures the gains when AI becomes cheaper and more widely used.
An expensive route to cheaper computing
A building and the processors inside it have different economic lives. A site with reliable power and good connections may remain valuable for decades. Equipment can remain operational while losing ground to machines that deliver the same service at lower cost.
For customers, that progress is welcome. Lower prices could open new applications and encourage existing users to consume more. If demand expands sufficiently, operators can increase revenue even as the price of a unit of computing falls.
But the adjustment need not be comfortable for every investment. An operator that paid more for equipment may face competition from a newer installation with lower costs. It must meet its financing obligations while deciding whether to keep running existing machines or commit further capital to replacements.
The result depends on utilisation, pricing and operating efficiency. Rapid technological improvement can increase the economy’s productive capacity while shortening the period in which a particular investment earns an attractive return.
Higher required returns make that period more demanding. They reduce the present value of future income and raise the threshold a new project must clear. Existing fixed interest rates and hedging can provide protection, but they do not secure cheap financing for the next expansion.
Governments enter the same market
Technology companies including Alphabet, Amazon and Meta had issued almost $220bn of bonds by mid August, compared with $108bn throughout 2025, according to LSEG figures reported by Reuters. Separately, the US federal deficit reached $1.97tn in the first 11 months of its fiscal year.
Gross corporate issuance and a government deficit measure different things and cannot simply be added. Nor is all technology borrowing destined for AI. They nevertheless illustrate the substantial financing demands reaching investors simultaneously.
McKinsey estimates that meeting global demand for AI computing could require $5.2tn of investment in data centres and associated equipment between 2025 and 2030. Including conventional computing takes the estimate to about $6.7tn. These are demand scenarios, rather than committed expenditure or sums that must all be borrowed.
There is no fixed global pot of savings. Higher returns attract capital, and successful investment creates income. Government borrowing does not mechanically determine yields, which also reflect inflation, monetary policy and growth.
But the alternatives available to investors matter. A project promising uncertain receipts must offer sufficient compensation to compete with other assets. Businesses with substantial earnings can finance expansion internally; others must persuade shareholders or lenders that their projected returns justify the commitment.
Consider a wholly hypothetical project costing $1bn, financed with $600mn of debt and $400mn of equity. Assume it generates $80mn a year after operating costs and an allowance for equipment replacement, but before interest and repayment of principal.
At an illustrative interest rate of 6 per cent, annual interest would be $36mn, leaving $44mn before principal repayments, tax and shareholder distributions. At 8 per cent, interest would rise to $48mn, leaving $32mn. If cash generation also fell by a quarter to $60mn, only $12mn would remain.
These invented figures describe no particular company or transaction. They assume the debt remains fully outstanding for the year, with no fees or hedging. They show how a project can continue serving customers while producing much less cash for its investors. Whether it can meet all its obligations also depends on its repayment schedule and reserves.
Who receives the income?
The financing structure determines how that income is shared. Shareholders receive what remains after obligations are met. Lenders receive contractual payments, sometimes supported by assets, guarantees and restrictions on distributions.
Those protections can improve recovery prospects. They cannot ensure that equipment will earn the revenue originally expected. If operating income falls short, shareholders may receive less, additional capital may be needed, or lenders may face losses depending on the circumstances and their contractual position.
Customer commitments can reduce uncertainty. Long contracts provide visibility, prepayments can help fund construction, and staged investment limits the amount committed before demand is established. Their value depends on payment terms, performance conditions and the customer’s ability to pay.
Risk can pass through banks, funds and insurers to the investors behind them. The institution arranging a transaction need not retain the exposure. Understanding who ultimately holds it requires looking beyond the size of the financing announcement.
For infrastructure investors, revenue growth is therefore only part of the test. Cash must arrive after operating costs and in time to meet payments and replacement expenditure. An expanding market does not remove that discipline.
Europe wants the productivity dividend
The financing search already extends beyond America. An August European Central Bank study found that US hyperscalers accounted for close to a tenth of new euro denominated borrowing by nonfinancial companies, although their share of outstanding bonds remained much smaller. The ECB said their growing presence could affect investment allocations and other borrowers’ access to capital.
In a September 14 speech, ECB president Christine Lagarde argued that Europe would bear some of the financing costs of the American AI expansion through connections between markets. She also cited estimates that rapid adoption could lift European productivity by up to 4 per cent over a decade.
That is the strongest economic counterargument to a pessimistic reading of the spending boom. Higher output and taxable income could help governments service the debt now competing with AI for capital.
The difficulty is timing and distribution. Construction comes before the broader productivity gains, which depend on businesses learning to use the technology effectively. Some benefits may appear as lower customer costs or higher profits elsewhere, rather than income for the infrastructure owner.
An economy can therefore earn a substantial dividend from AI without every supplier earning the return its investors anticipated.
The price of time
Energy prices complicate the timetable. Brent traded above $108 a barrel on September 14 as attacks on Saudi infrastructure compounded disruption from the Middle East war, the Guardian reported.
Oil prices do not translate directly into equivalent changes in every data centre’s electricity bill. Local generation, contracts and hedging determine the exposure. The wider effect comes through construction costs, inflation and the interest rates investors demand.
Persistent inflation could delay monetary easing. An eventual slowdown could pull demand and rates lower. Investment plans must accommodate that uncertainty rather than depend on an early return to cheap money.
The case for continued expansion remains substantial: growing demand, improving efficiency and established businesses able to fund development from earnings. Borrowing can sensibly match expenditure today with receipts tomorrow.
The financial test is whether those receipts cover the full cost of supplying the service. Utilisation, renewal prices, cash collection and equipment replacement will reveal more about that than spending announcements alone.
AI may become cheaper for its users while remaining expensive to build. The gains could spread throughout the economy. The debts must still be serviced by the businesses that incurred them.