Are the colossal sums being poured into artificial intelligence (AI) a ruinous waste of money, or an investment that will transform the world?
Few doubt the technology’s potential. The real question is whether the spending that the technology requires is economically viable. On this point, there is no more important question facing the global economy today, such is the dependence that global growth has on a financially and economically feasible future for AI.
The AI boom has triggered one of the costliest building sprees in history. Over the past three years, leading tech firms have committed more capital expenditure toward AI data centres, chips and energy than the United States spent building the entire interstate highway system over four decades, even adjusted for inflation.
Unlike highways, however, no one can say what these investments will do for global productivity. Moreover, no one knows when, or even if, the companies funding the investments will earn their money back.
What does this spend look like?
This year alone, America’s “hyperscalers” (to use the ghastly lexicon) — Amazon, Google, Meta and Microsoft — have pledged to spend a record R5.5-trillion ($320-billion) on capital expenditures (capex), much of it on building massive data centres packed with specialised AI chips known as “AI accelerators”.
This is a 40% jump from last year’s record-setting R4-trillion ($230-billion), and about two-thirds of all inflation-adjusted capital spent on laying the telecom fibre-optic cables during the 1990s, built over a decade and which continue to underpin the entire internet. Now, four firms are spending nearly that much in a single year.
An example is Meta’s Hyperion, an enormous AI data-centre complex in Louisiana with planned capacity of 2 GW, about the size of Manhattan, and with an estimated cost of R864-billion ($50-billion). That’s almost the size of South Africa’s entire R1-trillion national infrastructure budget over the next three years for roads, rail, energy, water and social projects combined.
The once-dull business of chips and servers has become a multihundred-billion-dollar battleground where Silicon Valley giants compete on spending commitments and increasingly silly sci-fi names.
In addition to Zuckerberg’s plans, OpenAI’s Sam Altman has announced “Project Stargate”, an R8.6-trillion ($500-billion) project 10 times the size of Hyperion. Bloomberg estimates that collectively tech CEOs have floated plans requiring at least R17-trillion ($1-trillion) of data centre investment over the next five years.
At the core of this trillion-dollar spending is one single goal: even without any idea of what the business case for the investment is, you must build as many AI-ready data centres as possible before the competition beats you to it. The closest historical parallel may be the Cold War arms race for nuclear weapons, when the US and USSR stockpiled countless nuclear warheads knowing there was no point as a fraction of them would destroy the world. They were spending simply to outspend their rival.
Where it could all come unstuck
There are several problems with this.
First is the financing. Technology firms with vast funding need to pay for their AI ambitions are striking blockbuster debt deals at the fastest pace in years, taking advantage of near-insatiable investor appetite to lock in financing for projects whose ultimate payoff remains uncertain.
This year alone they have raised about R2.7-trillion ($157-billion) from bond issuances, up roughly 70% from the same period in 2024, according to Bloomberg data.
Investors appear willing to bet that future demand for AI will justify this borrowing, trusting the companies to be able to generate enough cash to service the bonds decades from now.
This heavy reliance on debt means the AI boom is not just a stock market story with downside risks for shareholders. Any significant correction in valuations would spill over into bond markets and threaten the integrity of the entire global financial system.
Second is the incestuous nature of the financing and deal making happening between the handful of protagonists. Last week’s blockbuster tie-up between chipmaker Nvidia and OpenAI is a case in point.
The crux of the deal is simple. Nvidia will sell chips to OpenAI for their enormous data centres, in which it will train and then host AI models such as the recently launched GPT-5. In return, Nvidia will buy R1.7-trillion ($100-billion) of OpenAI’s unlisted stock over time, adding to the small stake it already has. That’s comfortably more than the R1.7-trillion ($72-billion) OpenAI has raised over its entire 10-year life.
This “you-scratch-my-back, I-scratch-yours” arrangement sounds a bit like the kind of vendor financing common during the early 2000s telecom frenzy. Or critics point out that it even resembles a Ponzi scheme, with sellers funding buyers without either really fully knowing what the endgame or who the ultimate customer is.
The revenue gap
Finally, spending is outpacing earnings by a huge margin. Consulting company Bain estimated in their latest Global Technology Report that by 2030 AI firms will need R34-trillion ($2-trillion) in annual revenue, roughly five times what they are currently making, to cover the costs of the capex. Yet they project the companies will fall about R13.8-trillion ($800-billion) short as efforts to monetise products like ChatGPT lag the vast infrastructure costs.
The economics of AI just don’t add up. Even OpenAI’s Sam Altman admits the company is not breaking even on compute costs, despite charging R3,400 ($200) per month for its premium tier. According to Bain, America’s largest software companies — including OpenAI, Salesforce and Adobe — have only generated R172-billion to R345-billion ($10-billion to 20-billion) in AI-related revenue this year, a fraction of their multihundred-billion-dollar investments.
This mismatch between colossal capex and uncertain returns feels familiar. Silicon Valley veterans warn that enthusiasm for AI has turned into a bubble that has increasingly worrying echoes of the mania around the internet’s infrastructure build-out in the late 1990s.
Then, telecom companies laid vast fiber optic networks ahead of demand that turned out to be much slower than anticipated. Although the companies collapsed, at least they created the infrastructure for what has become the internet. The difference is that data centres are not like underground cables. They are simply assets that must be depreciated over time. Like any PC’s CPU, they become obsolete.
Few doubt that AI will change the world. The question is whether companies can grow revenues and monetise the technology fast enough to close the gigantic mismatch between capex spending and revenue.
Should growth falter and companies fail to deliver on expectations, markets are primed for a dotcom-style bust. The stakes could not be greater. DM