Who’s paying for the AI build-out? Probably you

What was the cost of the Great Pyramid of Giza, as a percentage of the Fourth Dynasty’s GDP? It is unlikely that Pharaoh Khufu thought in those terms, and the accounts have been lost over the intervening 4,600 years. The question is topical nonetheless, as America is building a 21st-century equivalent: an AI data centre network that dwarfs all of its previous infrastructure efforts.

Stijn Van Nieuwerburgh of Brookings estimates the AI infrastructure boom will cost the equivalent of 3.6 per cent of US GDP a year over the eight-year period to the end of 2032 — equalling the annual spending on the railroad boom of the late 19th century, plus the interstate highway projects of the mid-20th century, plus about two-thirds of the electrification spending of the early 20th century. Total bill: $10.3tn, an estimate including only tangible expenditures, not intangibles such as software.

Who pays? Substantially all of the profits of some of the world’s most profitable companies — the so-called hyperscalers — have been staked. Add to this the new debt these mega-companies are taking on directly or guaranteeing, supplemented by more speculative debt and equity financing for smaller companies. Almost every big pool of capital, from every corner of the developed world, has been tapped. In short, if you have invested savings, you are financing the AI build-out.

The details are not easy to nail down, but hyperscalers’ direct investments are the most transparent. Alphabet, Amazon, Microsoft, Meta, Oracle and SpaceX are expected to generate $760bn in cash flow this year. All of this and more will be invested, meaning “free cash flow” will be negative, a situation Wall Street expects to persist for several years, even as profits at the companies are projected to grow at a breakneck 30 per cent a year.

Equity investors can console themselves with the idea that a big chunk of the hyperscalers’ investment converts into profits at semiconductor companies. But remember that these companies are investing heavily, too: capital expenditures at the 30 largest are set to pass $200bn this year, double the level of three years ago, and to grow from there.

There is another layer of financing flowing from the hyperscalers and chipmakers: equity investments in AI labs and other AI start-ups. Nvidia, for example, carries $94bn in private and public equity investments on its balance sheet, including stakes in OpenAI, Anthropic and CoreWeave. In a gold rush, it is often said, the real profits are made selling picks and shovels. In the AI gold rush, the shovel companies are investing part of their profits to buy mines.

The six hyperscalers and the big global chipmakers together are worth $31tn, or a staggering 20 per cent of the world’s stock market capitalisation. American families hold a particularly big portion of this equity risk: stocks account for almost a third of household assets, not including the equities backing pension entitlements.

Turning to debt. So far this year, $466bn in AI debt has been issued, according to Morgan Stanley. This figure may not be comprehensive, given the rainbow of asset-backed securities, convertible bonds and preferred shares that have come to market. Despite the intrinsically speculative nature of investment in a new technology, the debt supporting data centres is mostly rated investment grade — either because it carries guarantees from the highly rated hyperscalers, or is secured against the GPU chips. Is it a bit worrisome that a group of companies that are devoting all of their profits to a set of projects is also guaranteeing a lot of the debt financing of those same projects? And that much of the rest of the debt is secured against physical assets that will become less valuable if demand for AI services disappoints?

It is. But there is no way for AI infrastructure to be built without the investment-grade bond market. Over $2tn of US IG bonds were issued over the past year. Whatever needs to be done to secure investment-grade ratings will be done. This is the same reason why the hyperscalers are issuing bonds in the UK, Canada, Switzerland and Australia.

The significance of the investment-grade ratings, from the point of view of savers, is that they allow AI debt a bigger role backing pensions, annuities and life insurance policies. If you own any of these things, you are increasingly invested in AI infrastructure.

The Great Pyramid, after millennia of tourist revenue, has almost certainly produced a good return on the initial outlay. American savers, all of whom are invested in the AI build-out, have a somewhat shorter time horizon.

robert.armstrong@ft.com