The artificial-intelligence investment boom is entering a more difficult phase. For two years, companies, investors and governments have treated data centres, advanced chips and power infrastructure as the foundation of a new productivity cycle. Now markets are asking a less comfortable question: will the revenue generated by AI be large and reliable enough to justify the debt financing the build-out requires?
That tension became visible in October as borrowing costs rose for AI-linked companies and investors reassessed the durability of their valuations. Deloitte reported that credit-default-swap prices and bond yields for AI-related borrowers had climbed, signalling greater concern about repayment risk. Gartner has separately forecast that 80% of “neocloud” providers could cease operating independently by 2030 because capital expenditure may outpace returns. Deloitte
Why the spending is so large
AI systems require unusually expensive infrastructure. Training and running advanced models depends on specialised processors, electricity, cooling systems and high-speed networks. The largest technology companies are therefore committing tens of billions of dollars to data-centre construction and equipment, while smaller cloud providers borrow to compete for customers seeking computing capacity.
The investment has an economic logic. Supporters argue that AI can raise productivity, automate routine work, improve scientific research and create new services. Strong demand for chips, construction and electricity infrastructure has already helped sustain activity in several economies. The International Monetary Fund raised its 2026 global growth forecast to 3.1%, while identifying AI investment as a near-term source of momentum. International Monetary Fund outlook
For technology companies, moving quickly also has strategic value. A firm that delays investment risks losing access to scarce computing capacity or falling behind rivals whose products improve faster. That creates a race in which defensive spending can appear rational even when the returns remain uncertain.
The argument against an endless boom
The central risk is not that AI has no useful applications. It is that infrastructure may be built ahead of dependable demand. Cloud providers can sign large contracts, but those agreements do not guarantee that end users will pay enough to cover the cost of servers, power and financing. If prices for computing fall faster than utilisation rises, highly leveraged operators could face pressure to restructure, consolidate or sell assets.
Gartner’s forecast reflects that possibility, while Deloitte pointed to the widening gap between capital expenditure and likely returns. A market correction would not necessarily end AI development, but it could redistribute power toward companies with strong balance sheets and away from independent providers. That would make the sector less competitive and potentially leave customers more dependent on a small number of dominant platforms.
Investors have already shown sensitivity to this issue. A sell-off in AI-related shares spread through global markets in October, with the Nasdaq falling 4.5% over five sessions and South Korea’s KOSPI suffering sharper declines that triggered trading halts. Market data Such moves do not prove that the technology is overvalued, but they show that enthusiasm is no longer insulating the sector from conventional concerns about cash flow, debt and interest rates.
Why the consequences extend beyond technology
AI’s infrastructure needs are becoming an energy and public-policy issue. Data centres compete for electricity with households and manufacturers, and their construction can require new transmission lines, gas generation or renewable projects. In countries already facing high energy prices, policymakers must decide how much infrastructure to subsidise and who should bear the cost.
There is also a distributional question. The benefits of automation may accrue first to large firms and highly skilled workers, while communities hosting data centres absorb pressure on power grids, water supplies and land. A rapid build-out could therefore increase productivity while widening regional and income inequalities—unless governments link incentives to local investment, grid upgrades and workforce development.
What comes next
The next phase is likely to be less about headline spending and more about proof. Investors will scrutinise customer retention, revenue per unit of computing capacity and whether AI products reduce costs or generate measurable new income. Lenders will distinguish between firms with contracted demand and those relying on continual refinancing.
Governments face a parallel choice. They can support research and infrastructure without guaranteeing every private project, impose clearer disclosure requirements for AI-related financial risks, and ensure that energy and environmental costs are visible in project decisions. A more disciplined market would slow some construction, but it could also direct capital toward applications with stronger economic value.
The AI boom is therefore unlikely to end in a single dramatic reversal. More plausibly, it will divide into winners with durable demand and losers whose business models depend on ever-rising valuations. The technology may still transform the economy; the unresolved question is whether the financial system can fund that transformation without turning optimism into a debt crisis.