The multiplying risks of financing data centres

The AI revolution may change the world, but it comes with a 13-figure price tag. By 2030, tech companies are expected to pour $7tn into data centres — enough money to feed every person in China for three years.

So great are the sums involved that even Big Tech groups with tens of billions of dollars in cash need to fund the infrastructure with debt.

This is presenting major lenders with a complex set of calculations as they stretch themselves to finance, insure and underwrite an entirely novel asset class.

The possible returns are enormous, convincing some of the world’s biggest and most sophisticated private investors to provide Nvidia with $500bn in financing just this month, and offering what some in the insurance industry regard as a potential bonanza.

But the risks are also great, and growing. Firstly, the question of longevity: the possibility that the data centres and the chips that equip them will not retain their value for the lifetime of the financing that funds them.

“It’s like you’re financing a fax machine and then someone invented email,” says Carlos Mendez, co-founder at Crayhill Capital, which has repeatedly turned down debt deals secured by the value of chips that lacked structural protections.

There is also the risk of a growing popular backlash against the centres, with their mammoth appetites for water and electricity at a time when AI is increasingly controversial.

“As long as developers have the permits they need, they can continue borrowing. But local opposition or the loss of a permit could stop that midway through construction,” says Eric Klar, a debt finance partner at White & Case. “For lenders who have already put money into the project, the question is: what happens then?”

Then there is the additional difficulty of laying off risks to other parts of the financial system. Lenders and project developers are eager to insure against everything from natural disasters and power outages to a slump in demand for computing power. But insurance companies have been reluctant to provide billions of dollars’ worth of coverage for the biggest projects, concerned about catastrophic risk as well as concentrated exposure to these giant sites.

And that is without taking into account the most fundamental gamble of all — that the AI revolution will produce unprecedented productivity and profit gains, that the frontier models that require the most extensive data centre infrastructure will emerge the winners, and that the individual companies investors are backing will prevail.

“Everyone has decided they want to be in the railroad business,” says author Liaquat Ahamed, whose Pulitzer Prize-winning history of the 1873 railroad bubble has been cited as an essential text for navigating the current moment by Microsoft chief Satya Nadella. That company alone has announced some $175bn in data centre investment this year.

The problem with such expenditures, Ahamed adds, is that “they’re not taking into account that all the other tech companies are doing the exact same thing”.

Given all that, major lenders are looking to shift their riskiest exposures elsewhere, including to corners of the financial system that could be more vulnerable to stress — and in the process enveloping yet more of the wider economy in the AI bubble.

The data centre build-out has become an all-consuming feature of the American economy.

Stijn Van Nieuwerburgh, an economist at Columbia Business School, estimates that the AI build-out will account for some 2.8 per cent of economic output in the coming years, an even greater share of US GDP than railways represented at their 19th-century height.

The four so-called hyperscalers alone — Amazon, Microsoft, Alphabet and Meta — have signed leases in excess of $1.5tn since the AI boom began, including commitments that have yet to take effect.

The frenetic pace of activity across capital markets has been startling. Starting late last year, the hyperscalers have tapped the $11.7tn US corporate debt market repeatedly, so much so that they have to begin raising debt in foreign currencies to broaden the investor pool.

The financing requirements for AI are stretching Wall Street’s resourcefulness — and its resources.

It took six powerhouses of finance coming together to provide Nvidia’s half-trillion-dollar funding needs earlier in August, with BlackRock joined by Blackstone, Apollo Global, KKR, Brookfield and Goldman Sachs. Larry Fink, BlackRock’s chief executive, called the combination the “future for financial engineering”.

Large banks that invested heavily in data centre projects are seeking avenues to offload these concentrated risks.

JPMorgan Chase, Morgan Stanley and SMBC have all explored using synthetic risk transfers — where lenders offload risk to investors through a financial guarantee while the loans remain on their balance sheet — to reduce their exposure to data centres. Banks led by Morgan Stanley are also planning to offload $15bn of debt they lent to a Google-backed data centre in Texas, the FT previously reported.

Some new financing structures have relied on special-purpose vehicles, which keep the projects off the balance sheets of tech companies and shift the risk to external investors.

Chipmaker Broadcom’s deal with Blackstone and Apollo uses the semiconductors it develops with Google as collateral for potentially hundreds of billions of dollars in financing capacity.

Richard Myers, global head of private capital markets at Morgan Stanley, which advised on Broadcom’s chip financing package, says there is potential for these privately placed deals to be moved on to the public market and retail investors.

“Pensions and insurance are obvious candidates for this type of stable and low-risk investment profile,” he says, referring to the long-term contracts and credit guarantees provided by high-grade companies.

The promise of long-term rent payments and stable streams of income has convinced lenders to offer ultra-low interest rates that are only achievable for the most creditworthy companies.

In many cases, these data centre projects have the same credit rating as the tech companies that lease them, despite the fact that the debt vehicle often has no direct claim on the physical assets.

But over the course of this year, investor interest in buying debt related to data centres has begun to fade. Order books on new deals — a key measure of investor demand as bankers tot up willing participants in a bond sale — have slimmed meaningfully in size.

When QTS was selling $4.6bn of bonds to fund its Fayetteville, Georgia, data centre in April, it attracted an order book almost three times the deal size, as investors clamoured over its long-term contract with the most creditworthy company in the world: triple A-rated Microsoft.

Many investors overlooked a caveat at the time: the debt will not be fully paid down before the initial lease ends, meaning that the data centre will have to find new tenants to stay afloat — and investors are sceptical of the long-term demand for AI compute.

“We don’t know what this data centre is going to be worth in 10 years,” says Edwin Wilches, co-head of PGIM Credit’s securitised products team, who passed on the deal. “The world got a rude awakening . . . Investors were like, that’s not what I thought it was, and they started to reprice the risk.”

QTS’s new debt issued in August was sold at over 7.2 per cent yield, similar to the borrowing costs of junk-rated companies, compared to only 5.7 per cent during its last bond sale in April.

“People just assumed everything Microsoft-backed is great,” says Wilches. “The market was a little bit lax in assessing these deal structures, especially when it came to reading all of the terms and conditions.”

Similarly, loans for data centres backed by AI chips rest on the assumption that the rapidly changing technology will not quickly lose value.

But Nvidia releases new, improved and more expensive versions of their chips every year, potentially making older silicon obsolete within a handful of years. For that reason, many lenders are now asking for such loans to be paid off entirely before the underlying lease ends.

Across America, construction workers are racing against tight deadlines to build the infrastructure of the AI revolution.

Data centre operators have set up camp wherever they have been able to identify a reliable source of power and water to help fuel and cool the semiconductor chips and other IT equipment to train and deploy AI models.

But the AI boom is running into some very human hurdles, as communities protest against the development of data centres in their backyards. Seven in 10 Americans oppose the construction of data centres in their local area, according to a March poll by Gallup.

Large projects, such as a Blackstone-backed facility in Virginia, have been cancelled after running into local opposition and lawsuits, while many local governments, including in New York State, are implementing moratoriums on the development of new data centres.

Investors are worried. A 1.4-gigawatt Oracle campus in Saline Township, Michigan, suffered financing trouble late last year following permitting delays and local pushback. The project received backing from Blackstone in April but investors demanded a higher risk premium and more credit protections.

Strict local regulations are adding to costs. The Public Service Commission of Wisconsin, the state’s utility regulator, in June slapped Oracle with a new $7bn collateral requirement to secure power, citing the tech giant’s weak creditworthiness. Oracle’s credit rating was downgraded by S&P later that month to triple B minus, just one notch above junk status.

In New Mexico, Oracle replaced plans for a potential gas plant with fuel cells from Bloom Energy to power its 1,400-acre Project Jupiter data centre, after the public raised concerns about pollution. But now the regulator has cracked down on plans for a gas pipeline to feed the fuel cells.

The “pool of friendly locations” is “shrinking a lot”, says Elaine Walsh, chair of law firm Baker Botts’ power practice group. “A lot of the hyperscalers are having to think twice about locations, it’s slowing down development.”

Financiers are trying to deal with political risk by requiring projects to have the permits, equipment and labour they need lined up before deals are signed, and implementing get-out clauses if these fall through. But lenders at the start of the financing chain can still be left exposed.

In the meantime, developers say that investor caution can prevent them from getting the permits they need in the first place.

“It can create a chicken-and-egg problem,” says Ben Alingh, chief executive of Monarch Energy, a developer that specialises in securing land and power for data centre sites. “Sometimes you need investment to enable the permits and it’s getting more difficult to match up the timelines.”

The data centre build-out promises a colossal reward for the insurance industry, especially at a time when its business is under pressure elsewhere. Joe Peiser, a senior executive at broker Aon, called data centres the industry’s single “biggest-ever opportunity”.

But the vast scale of the projects is testing the limits of insurers’ underwriting capabilities and their balance sheets.

Construction costs for a single data centre now routinely exceed $10bn, with financing packages for hyperscalers stretching into the trillions.

By comparison, the largest natural catastrophe exposure last year at Munich Re, the world’s largest reinsurer, was an €8.5bn exposure to a hypothetical giant Atlantic hurricane, which represents aggregate losses from thousands of individually insured entities.

Even after dozens of insurers agree to back a data centre, brokers have struggled to squeeze out enough cover for basic catastrophe risks, let alone the more complex coverages, such as power outages, that developers are seeking.

Insurers also said that they lacked the data to model risks facing data centres, which has made them warier.

Aki Hussain, chief executive of commercial insurer Hiscox, says that the company has not yet decided whether to grow its data centre insurance business beyond its current “incidental” exposure.

“It’s very easy to jump in with both feet and then to regret it,” he tells the FT.

Tanguy Touffut, the chief executive of Descartes Underwriting, has sold data centres so-called “grid downtime” parametric insurance agreements covering them against the risk of power outages. Parametric agreements pay out automatically if certain triggers are met, such as the issuance of an emergency alert from a grid operator.

Similar coverage has been difficult for traditional insurers to provide at the necessary scale, he says, because “the level of uncertainty for insurers covering the collapse of a grid system is super high”.

Insurers’ wariness threatens to create a drag on the hyperscalers’ ambitions. Large investors including KKR and Blackstone are among lenders to have turned down data centre debt because of insufficient insurance, according to people familiar with the matter.

Hyperscalers have nonetheless pushed ahead with the projects, exposing themselves to billions of dollars of potential damages in the event of a catastrophe, grid outage or lawsuit.

Only a fraction of Meta’s $14bn new data centre in Texas is covered by insurance, leaving the project exposed to billions of dollars of potential losses in the event of disasters.

Meta declined to comment. A person familiar with the matter said no bidders under formal consideration withdrew from Meta’s partner selection process for its Texas site.

Big tech companies moving ahead with an expected $3tn of capital spending without full insurance coverage presents its own risks, with analysts at S&P warning it could ultimately dent the willingness of investors to bankroll projects.

In addition to construction and property insurance policies, developers have paid tens of millions of dollars in premiums for insurance against terrorism that will give them coverage against attacks worth hundreds of millions of dollars.

One of the largest data centre insurers told the FT that when they recently signed up to insure property risks at a Texas data centre, they had turned down a request for terrorism cover.

“Normally we’d cross-sell the terror [insurance],” the insurer says, “but if somebody chucks something over the fence perimeter, you’re looking at a big loss”.

Finally, there is the risk looming in the background that the data centres and the power plants being built to electrify them will one day be stranded assets.

The AI industry has not been immune to market ructions. Chinese lab DeepSeek’s R1 model triggered a sell-off early last year after it trained a model on significantly fewer resources than its western counterparts. The release raised doubts about the amount of computing power leading AI labs would need to build capable models.

Investors’ fears around the timeline by which AI-led returns might be achieved in July wiped more than $1tn from the market capitalisation of several leading semiconductor companies.

Planned IPOs from Anthropic and OpenAI this year will further test investors’ appetite for the technology and their willingness to fuel a capital cycle that relies on uncertain returns.

Advait Arun, senior associate for capital markets at the Center for Public Enterprise, warns that even with the possibility of government intervention, there was no playbook to unwind a troubled data centre.

“We really don’t know what a data centre bankruptcy restructuring process would look like,” he says.

Recommended

If the data centre bubble pops, the places where they have set up shop — states such as Ohio, Texas and Wisconsin — could be left with bruising economic consequences.

Communities could be left with vast white elephants in their backyards, the costs of which could filter into monthly utility bills for retail customers. Localities that offered data centres generous tax incentives could find a glaring hole in their coffers.

Even if data centre projects are not cancelled en masse, advancements in chip efficiency and shrinking demand could slow the spending boom.

According to Andy DeVries, who covers utilities at credit research shop CreditSights, the supply of AI computing capacity will begin to outstrip demand after 2029, leading to significant reverberations beyond the current bubble.

“That will trickle through the whole AI ecosystem,” he says. “Then the Main Street economy.”

Additional reporting by Eric Platt in New York. Data visualisation by Eva Xiao.