Big Tech is spending $725 billion on AI this year, and central banks are worried
The four biggest cloud companies are putting three quarters of a trillion dollars into AI infrastructure this year, and the world's central banks now list it as a financial risk. If your business runs on their AI, part of that risk is yours.
On June 28 the Bank for International Settlements, the institution that coordinates the world’s central banks, named the AI investment boom one of the biggest threats to global financial stability. That is a long way from tech-blog talk of a bubble. For a company that uses AI, the useful question is narrower: what happens to its tools and its costs if the money behind them dries up?
Three quarters of a trillion dollars in one year
Amazon, Alphabet, Meta and Microsoft plan to spend roughly $725 billion on capital expenditure in 2026, up about 77% from around $410 billion the year before, and nearly all of it goes to AI. Their guidance: Amazon near $200 billion, Microsoft around $190 billion, Alphabet $175 to $185 billion and Meta $115 to $135 billion. With Oracle, the five largest spenders will pass a trillion dollars across 2025 and 2026, and Goldman Sachs models around $7.6 trillion of cumulative buildout between 2026 and 2031.
The money used to go back to shareholders. Capital spending has pushed free cash flow toward negative territory for the first time in decades: Amazon is projected to go negative, and analysts expect a steep drop at Meta. Capex now runs at roughly a third of revenue for the big spenders, more than double the peak of the 1990s internet buildout. Some of it buys no extra compute at all. Microsoft’s finance chief put $25 billion of its budget down to rising memory and component prices. The buildout is driving up the price of its own inputs.
What the central bankers see
The BIS warning compared the buildout to canal mania, the railway bubble and the dot-com crash, and said disappointing returns could turn the boom into a protracted investment bust.
Two things worry regulators. One is the financing. Hyperscalers, chipmakers and AI labs are tied together through circular deals: a cloud provider takes an equity stake in an AI lab, and the lab commits to spending billions on that provider’s compute. Each company’s revenue looks strong partly because another company in the loop is paying it. The other is the gap between spending and income. Against roughly $700 billion of annual capex, end-user AI revenue in 2026 is estimated at between $50 and $150 billion, and the BIS found few companies reporting clear productivity gains once AI projects reach production at scale.
The optimists have numbers too. Google’s cloud revenue rose 63% year over year, and Fidelity points out that, unlike in the late 1990s, most of this is still paid for from earnings rather than debt. Nobody knows yet whether revenue will grow fast enough to justify the data centers already under construction, and anyone who tells you otherwise is guessing.
What it means if your business runs on AI
You can act on this without a view on the stock market.
Your AI is subsidised. Frontier model access is cheap partly because providers are spending ahead of revenue to win customers. If financing tightens, generous prices and usage limits are among the first things to go. Budget for per-token costs that can rise.
A correction would not stay in tech. A handful of AI companies make up an outsized share of the main stock indices, so a sharp fall would hit pension funds and consumer confidence. A Danish company is exposed through its export markets and its employees’ pensions, whether it uses AI or not.
European companies rent this capacity. Most of the AI compute European businesses use belongs to the same four American companies, even when the data center is in Europe. Their spending decisions, and any pullback, set prices and availability here.
Check the smaller suppliers. The hyperscalers can absorb a bad year. The contractors and smaller infrastructure firms with thinner balance sheets are more fragile. If you depend on a smaller cloud or GPU-rental provider, find out how it is financed before you build on it.
Measure your own return. A company that can show productivity gains from AI in production, past the pilot stage, is protected if sentiment turns, and it is good discipline anyway.
Electrification and the early internet both looked overbuilt before they paid off, and this buildout may too. But at $725 billion a year, the stability of the money behind the models matters to your business about as much as the models do.
