Lex in depth: Will the AI data centre boom become a $9tn bust?
The biggest groups splashing their cash may not make their money back, but will almost certainly live to tell the tale
Every so often, the human race comes down with building fever. Investors’ temperatures soar, lavish projections of demand go viral and spending swells. Once the fever breaks, there comes a long and painful convalescence.
Railways, early automobiles, telecoms networks, shale oil wells and Chinese apartments have all been through the cycle. Now it is the turn of data centres that host AI services. Can the multitrillion-dollar investment in these air-conditioned electronic warehouses — perhaps the biggest peacetime investment project in history — buck the historical trend of booms ending in busts? Some powerful people believe so, at least for them.
Among them are Meta Platforms boss Mark Zuckerberg, and his counterparts at Google parent Alphabet, Microsoft, Amazon and Oracle. The five are forecast to deploy $4tn of capital expenditure over five years, according to analyst estimates gathered by Visible Alpha, most of it on data centres they hope will reshape their businesses.
The spending is already reshaping their balance sheets. The formerly debt-light Google recently borrowed $32bn from the bond markets. Meta issued $30bn last November, and is also taking on off-balance-sheet commitments for massive data centre projects. If this proves just another case of investment delirium, there is much to lose.
Big tech, bigger ambitions
For all the hype, it is surprisingly hard to ascertain how many data centres are actually being built and at what cost. Researchers at McKinsey have crafted a figure of $5.2tn for AI-related computing facilities by 2030, based on 125 gigawatts at roughly $40bn per gigawatt. Data centres are measured by the power they require at peak times rather than the amount of data they crunch. For comparison, the UK’s peak electricity usage in 2025 was 46GW.
The McKinsey sum already looks light. A year ago, you could budget $25bn for chips and hardware, and $15bn on land, power and other inputs for a gigawatt of data centre power. Those numbers are now more like $35bn and $20bn, according to people familiar with data centre development. That takes the theoretical bill between now and 2030 to $6.9tn.
The McKinsey total may also not fully reflect the giant future investment by so-called “hyperscalers” such as Google, Meta and Amazon. Add the equivalent of half the $4tn analysts expect them to deploy and it ratchets up the number to a mountainous $9tn.
In real terms, that is roughly what China spent on residential real estate between 2016 and 2021, before its property market started to slump. And it is more than double the value of US investment in computing equipment in the five years preceding 2000s dotcom crash, according to data from the Federal Reserve Bank of St Louis.
The quest for profit
Assuming these colossal investments actually happen, it’s worth standing back to look at what returns would be required to make them earn their keep. One way is to take the minimum profit needed to match the rate of return, or cost of capital, that investors see as an acceptable minimum.
What is that rate, though? For infrastructure assets such as toll roads and power stations, the blended cost of debt and equity capital can be quite low, perhaps 7 or 8 per cent annually. Data centres are riskier than that. Professor Michael Roberts at the Wharton School at the University of Pennsylvania suggests an asset with cash flows correlated to the business cycle could require a return of 15 per cent or more.
Say, then, that an aggregate $9tn investment in data centres will require a 10 per cent return, or $900bn of profit per year after operators have accounted for expenses such as energy and depreciation. Assuming a profit margin of about one-third implies a need for $2.7tn of revenue. That is not far off what the US spent on software last year, according to official data.
For the hyperscalers, there are two places such revenue can come from: sales of products and tools they cook up in their AI labs, or renting their chips and servers to others. Microsoft does both. Amazon and Google do the latter. Meta, with no cloud business, has no direct revenue from data centres: everything it builds and commissions is for its own use.
Drill down, though, and business models get more fuzzy. Google and Meta say their internally generated AI models are souping up their ad sales businesses and keeping users hooked for longer. Microsoft charges for its AI-enhanced office tools but hopes to lure consumers and software programmers with free services too.
There are various ways to size up the revenue opportunity. Wells Fargo analysts posit, for example, that if internet advertising is growing by 15 per cent across the industry, Meta’s 25 per cent expansion suggests that roughly 10 percentage points of growth — or roughly $20bn in a year — is down to AI. Meta itself remains mum on its own assumed returns.
Investors’ faith, perhaps unsurprisingly, waxes and wanes. Meta shares plunged 11 per cent in October after it raised its capital expenditure forecasts, but rose 10 per cent in January when it did so again. Microsoft stock fell 10 per cent after its latest quarterly earnings, despite beating forecasts. What’s clear is that as cash gets funnelled into capital expenditure, there’s a lot less left in the near term for shareholders.
Executives are doing their best to project confidence, of course. Microsoft’s Satya Nadella argues AI should “bend the productivity curve”. OpenAI’s Sam Altman predicts the creation of “universal extreme wealth”. And Meta this month, in a gesture of supreme self-belief, issued top executives stock options exercisable only in the event that its shares soar, in some cases to six times their current price.
What if it all comes unstuck?
In the short term, such confidence looks justifiable. Between their own needs and those of customers, there is excess demand for all they can build and more. Those who rent out cloud capacity report dramatically escalating customer orders; Microsoft, Google and Amazon’s revenue backlog doubled last year, Goldman Sachs analysis shows.
That means that in the event of some kind of sudden reappraisal, hyperscalers can simply slow their rollout. This isn’t “if you build it, they will come”, unlike the telecoms-network boom and bust of the late 1990s, which left a glut of unused “dark fibre” that took a decade to find users.
But there’s no guarantee that what worked for last year’s investments will also work for next year’s, or that supply won’t start to outstrip demand. Zuckerberg admits he is spending to meet “the most optimistic cases”. Technology can change quickly, and with it, assumptions about how much capacity is needed and where. Nor is it clear how long expensive chips will last before they need to be replaced.
Another risk is simply that demand for AI products builds more slowly than expected. Around 95 per cent of AI projects in businesses currently fail, according to an oft-cited report last year by MIT. In the telecoms bubble, the fatal belief was that internet traffic was doubling every 90 days, whereas it was in fact doubling once a year. Timing matters enormously to financial returns — especially when there is debt to be serviced.
Some AI insiders are already warning of the risk that comes with grand projections. Dario Amodei, co-founder of Anthropic, has cautioned that if the real numbers go off course, big spenders could face bankruptcy as a result of “Yolo-ing” on capital expenditure, a reference to some of his peers’ you-only-live-once exuberance.
OpenAI might be the biggest wild card. Altman’s company, the inventor of ChatGPT, at one point wanted to commission 250GW of data centre capacity over time — potentially costing more than $10tn. Since OpenAI doesn’t hold data centres on its own balance sheet, that would have fallen to companies that specialise in building facilities for others such as CoreWeave and Crusoe, as well as chipmakers like Nvidia and Advanced Micro Devices.
Altman has subsequently moderated his plans. OpenAI had intended to spend $1.4tn of its own money on renting data centres over eight years; now it will spend $600bn over four. Altman is also discontinuing the company’s power-hungry video generator Sora, launched only a few months ago, in a sign that financial discipline may be becoming more of a focus. Lay-offs and bonus cuts at Meta tell a similar tale.
A much weirder variable may be OpenAI’s foundational “self-sacrifice” clause. Altman has committed to stop OpenAI’s march towards superhuman intelligence if a rival “comes close to” reaching that goal, and redirect its efforts to helping them out instead. What that means in practice is up for grabs. Investors in future data centres OpenAI might use will presumably hope it doesn’t happen.
Even without that, it’s hard to know exactly how much the hyperscalers themselves think they need. Meta gives guidance for its own investment plans but is separately agreeing to rent space from other cloud providers, which makes its potential consumption unclear. Given the fierce rivalry between AI model makers, the ambiguity is probably intentional.
If Zuckerberg’s plans change after he has invested a chunk of the $620bn analysts expect over the next four years, or if Meta agrees to leases it later realises it does not need, investors will not thank him for incinerating their capital. That said, since Zuckerberg controls his company through his super-voting shares, there’s not much investors can do besides grumble.
Even a massive climbdown would hardly be an existential issue for the Facebook owner, which last year made almost $200bn in advertising revenue. Google, Amazon and Microsoft too have real businesses to fall back on. OpenAI, in contrast, does not.
That, plus the fact that so much of their investment is funded out of cash flow, puts the more established Silicon Valley giants in a very different position to companies in previous manias, where massive outlays were financed by borrowings or equity issuance and the bursting of the bubble left participants with no business at all. Big Tech may not make its money back, but it will almost certainly live to tell the tale.
The AI countertrade
The tech giant that is conspicuously sitting out the data centre craze is Apple, which has retained an extremely lean balance sheet as its peers develop a taste for big debt issues and fixed assets.
While it has partnered with Google to power its own AI offerings, and dabbled unimpressively with “intelligence” features, the company behind the iPhone has eschewed the heavy lifting of developing models from scratch.
To see the difference, consider a somewhat old-fashioned but illustrative financial measurement known as “fixed asset turnover”. For each dollar’s worth of property and equipment on Apple’s balance sheet last year, it made more than $8 in revenue. That compares with $2 at Amazon and just over $1 at Meta. And at most of the hyperscalers this yield is declining rapidly.
For Apple boss Tim Cook, a lack of interest in covering the planet with servers is either a stroke of genius or a fatal miscalculation, something Carlyle’s head of strategy Jason Thomas likens to “a binary option”. If Apple were also to lose its grip over the device market — OpenAI, Google and Meta are all working on their own gadgets — it would be in serious trouble.
Another way of seeing this is that the iPhone maker can use its balance sheet for other things. Cook can choose his AI partners based on the best models available for a given task. Or he could do deals of an entirely different kind. Apple could, theoretically, buy Disney, creating a consumer media-and-tech juggernaut, with a little more than the $185bn Google has earmarked for capital expenditure this year.
Moreover, all that AI has to be used somewhere, and Apple still has an advantage when it comes to the computing devices on which hundreds of millions of users actually interact with large language models. That is not just phones: the launch late last year of OpenClaw, a customisable AI personal assistant that can run on a home computer, led to a rush of armchair tech buffs buying Apple’s dependable, user-friendly Mac Minis.
That speaks to another unknown that might work in Apple’s favour: the growing move towards “edge AI”, or models run on local devices. While Zuckerberg, Altman and peers drive towards godlike AI that sits in the cloud, many users may find their needs met by simpler models that reside on their laptop or phone, barely touching a data centre at all. If that’s the future, sitting out Big Tech’s generational spree could be a smart move.
