AI Data Center Boom Carries Huge Default and Demand Risks
“How does the digital economy exist?” asked John Medina, a senior vice president at Moody’s, who specializes in assessing infrastructure investments. “It exists on data centers.”
New investments in data centers to power Artificial Intelligence (AI) are projected to reach $3 trillion to $4 trillion by 2030, according to Nvidia. Other estimates suggest the investment needed to keep pace with AI demand could be as high as $7 trillion by 2030, according to McKinsey. This massive spending is already having a significant economic impact, with some analysis indicating that AI data center expenditure has surpassed the total impact from US consumer spending on GDP growth in 2025.
U.S. data center demand, driven largely by A.I., could triple by 2030, according to McKinsey. That would require data centers to make nearly $7 trillion in investment to keep up. OpenAI, SoftBank and Oracle recently announced a pact to invest $500 billion in A.I. infrastructure through 2029. Meta and Alphabet are also investing billions. Merely saying “please” and “thank you” to a chatbot eats up tens of millions of dollars in processing power, according to OpenAI’s chief executive, Sam Altman.
- OpenAI, SoftBank, and Oracle pledging to invest $500 billion in AI infrastructure through 2029.
- Nvidia and Intel collaborating to develop AI infrastructure, with Nvidia investing $5 billion in Intel stock.
- Microsoft spending $4 billion on a second data center in Wisconsin.
- Amazon planning to invest $20 billion in Pennsylvania for AI infrastructure.
Compute and Storage Servers within an AI Data Center. Photo credit: iStock quantic69
The spending frenzy comes with a big default risk. According to Moody’s, structured finance has become a popular way to pay for new data center projects, with more than $9 billion of issuance in the commercial mortgage-backed security and asset-backed security markets during the first four months of 2025. Meta, for example, tapped the bond manager Pimco to issue $26 billion in bonds to finance its data center expansion plans.
As more debt enters these data center build-out transactions, analysts and lenders are putting more emphasis on lease terms for third-party developers. “Does the debt get paid off in that lease term, or does the tenant’s lease need to be renewed?” Medina of Moody’s said. “What we’re seeing often is there is lease renewal risk, because who knows what the markets or what the world will even be like from a technology perspective at that time.”
Even if A.I. proliferates, demand for processing power may not. Chinese technology company DeepSeek has demonstrated that A.I. models can produce reliable outputs with less computing power. As A.I. companies make their models more efficient, data center demand could drop, making it much harder to turn investments in A.I. infrastructure into profit. After Microsoft backed out of a $1 billion data center investment in March, UBS wrote that the company, which has lease obligations of roughly $175 billion, most likely overcommitted.
Some worry costs will always be too high for profits. In a blog post on his company’s website, Harris Kupperman, a self-described boomer investor and the founder of the hedge fund Praetorian Capital, laid out his bearish case on A.I. infrastructure. Because the building needs upkeep and the chips and other technology will continually evolve, he argued that data centers will depreciate faster than they can generate revenue.
“Even worse, since losing the A.I. race is potentially existential, all future cash flow, for years into the future, may also have to be funneled into data centers with fabulously negative returns on capital,” he added. “However, lighting hundreds of billions on fire may seem preferable than losing out to a competitor, despite not even knowing what the prize ultimately is.”
It’s not just Silicon Valley with skin in the game. State budgets are being upended by tax incentives given to developers of A.I. data centers. According to Good Jobs First, a nonprofit that promotes corporate and government accountability in economic development, at least 10 states so far have lost more than $100 million per year in tax revenue to data centers. But the true monetary impact may never be truly known: Over one-third of states that offer tax incentives for data centers do not disclose aggregate revenue loss.
Local governments are also heralding the expansion of energy infrastructure to support the surge of data centers. Phoenix, for example, is expected to grow its data center power capacity by over 500 percent in the coming years — enough power to support over 4.3 million households. Virginia, which has more than 50 new data centers in the works, has contracted the state’s largest utility company, Dominion, to build 40 gigawatts of additional capacity to meet demand — triple the size of the current grid.
The stakes extend beyond finance. The big bump in data center activity has been linked to distorted residential power readings across the country. And according to the International Energy Agency, a 100-megawatt data center, which uses water to cool servers, consumes roughly two million liters of water per day, equivalent to 6,500 households. This puts strain on water supply for nearby residential communities, a majority of which, according to Bloomberg News, are already facing high levels of water stress.
“I think we’re in that era right now with A.I. models where it’s just who can make the bigger and better one,” said Vijay Gadepally, a senior scientist at the Lincoln Laboratory Supercomputing Center at the Massachusetts Institute of Technology. “But we haven’t actually stopped to think about, Well, OK, is this actually worth it?”
References:
What Wall Street Sees in the Data Center Boom – The New York Times
Will billions of dollars big tech is spending on Gen AI data centers produce a decent ROI?
Gartner: AI spending >$2 trillion in 2026 driven by hyperscalers data center investments
Analysis: Cisco, HPE/Juniper, and Nvidia network equipment for AI data centers
Cisco CEO sees great potential in AI data center connectivity, silicon, optics, and optical systems
Networking chips and modules for AI data centers: Infiniband, Ultra Ethernet, Optical Connections
Superclusters of Nvidia GPU/AI chips combined with end-to-end network platforms to create next generation data centers
Lumen Technologies to connect Prometheus Hyperscale’s energy efficient AI data centers
7 thoughts on “AI Data Center Boom Carries Huge Default and Demand Risks”
Leave a Reply to Anonymous Cancel Reply
U.S. government’s new data-center estimates (in its revised 2Q25 GDP report) show companies spent $40.4 billion at a seasonally adjusted annual rate on data center infrastructure in the second quarter, up more than fourfold from $9.5 billion at the start of 2020, the first year measured.
Steadily increasing investment in data centers “speaks volumes about why the U.S. remains the premium destination for foreign investment,” said Joe Brusuelas, chief economist at RSM. “It’s part of the rich mosaic of American growth” which will affect the direction and depth of U.S. economic output in the years to come.
https://www.wsj.com/economy/u-s-gdp-grew-stronger-than-estimated-in-second-quarter-5ab397ad
Massive infrastructure investment from the likes of Microsoft (MSFT), Alphabet (GOOG), Amazon (AMZN) and Meta (META) has been a primary source of fuel for the AI rally. Their spending has caused revenue at chipmakers like Nvidia (NVDA), Broadcom (AVGO) and Micron (MU) to explode, and underpinned the narrative that AI demand is insatiable.
The hyperscalers will all update investors on their capital expenditure plans when they report quarterly results in late October and early November. Alphabet and Meta each lifted their capex forecasts in their most recent earnings reports, while Microsoft and Amazon said they would continue to invest heavily in infrastructure throughout the year.
https://finance.yahoo.com/news/ai-stocks-fueled-bull-market-194514237.html
Neoclouds are emerging as a distinct category of cloud providers built specifically for AI. Their core attributes include GPU-as-a-service models, ultra-low-latency networking, and software stacks optimised for large-scale AI workloads. Unlike traditional infrastructure-as-a-service or platform-as-a-service offerings, neoclouds provide capabilities tailored to AI, from training and inference to fine-tuning and full model hosting.
Global and European players are already establishing themselves in this space, bringing new competition to hyperscalers and regional providers. For the wider ecosystem, the implications are clear. Neoclouds introduce new types of tenants, push facilities toward higher-density requirements, and require earlier supplier engagement than conventional cloud builds.
According to Kristina Lesnjak, EMEA Research Manager, “AI is rapidly reshaping the foundations of cloud infrastructure, driving demand for GPU scale and ultra-low latency that neoclouds are uniquely positioned to deliver. With record levels of investment flowing into the UK’s digital infrastructure and leading technology firms aligning around AI, the momentum behind this transformation has never been stronger.”
https://www.dcbyte.com/news-blogs/ai-neocloud-uk-infrastructure/
It seems a week doesn’t go by that a massive AI deal isn’t announced. OpenAI (OPAI.PVT), Nvidia (NVDA), AMD (AMD), Oracle (ORCL), Broadcom (AVGO), and a raft of other companies are spending billions investing in each other and each other’s hardware.
OpenAI is the company making the most waves in the dealmaking space of late. The ChatGPT developer signed a massive contract in September with Nvidia that will see the chipmaker invest up to $100 billion in the company in exchange for OpenAI purchasing upwards of 10 gigawatts of GPUs over several years. Nvidia said that the partnership will allow OpenAI to deploy “at least 10 gigawatts” of compute capacity from the chipmaker’s AI systems to train and run the ChatGPT maker’s next generation of artificial intelligence models.
The first phase of the partnership kicks off in the second half of 2026, when Nvidia will begin deploying its next-generation Vera Rubin superchips.
One month later, OpenAI unveiled an agreement with Nvidia rival AMD that will see the Sam Altman-helmed AI firm purchase shares equivalent to roughly 10% of the chip company in exchange for upwards of 6 gigawatts of GPUs.
As with the Nvidia deal, AMD will supply OpenAI with its next-generation MI450 AI chips beginning in the second half of 2026.
OpenAI wasn’t done making moves just yet, though. This week, it let fly with its latest news, a 10 gigawatt transaction with Broadcom through which the two will co-develop custom chips to run OpenAI’s AI models and services.
OpenAI will design the accelerators and systems, and Broadcom will develop and deploy them.
All totald, OpenAI has signed deals worth up to 26 gigawatts of GPUs over the time of the three agreements. One gigawatt of electricity is enough to power 800,000 homes, according to Reuters. Those 26 gigawatts then work out to roughly 20.8 million homes.
OpenAI has also signed an agreement with Oracle as part of the company’s Stargate Project. Worth more than $300 billion, the deal, announced in July and fully revealed in September, is one of the AI giant’s most ambitious ventures and will give OpenAI an additional 4.5 gigawatts of GPUs over the next five years.
………………………………………………………………………………………………………………………………………….
Oracle isn’t just working with OpenAI. On Tuesday, the cloud computing company said it will purchase upwards of 50,000 AMD GPUs starting with its MI450 chips in the second half of 2026.
Oracle also reportedly signed an earlier deal to purchase $40 billion worth of Nvidia chips for OpenAI’s Stargate Project, according to the Financial Times.
https://finance.yahoo.com/news/nvidia-stock-jumps-on-100-billion-openai-investment-as-huang-touts-biggest-ai-infrastructure-project-in-history-171740509.html
Since 2024, private investors have spent $259B funding AI start-ups, according to Crunchbase. OpenAI is making an early bid to become a colossus in the AI age, but many of these new companies may not yet be on anyone’s radar, and some might not even exist. There are long-term risk tied to OpenAI, which underlies Oracle’s growth.
There’s still no clear answer as to where OpenAI will get the money for its Oracle cloud contract. Or how Oracle itself will finance the capital expenditures required to fulfill its end of the bargain. Perhaps most complicated: Where will the companies find the electricity to power these data centers?
Oracle’s long-term value relies on answers to these thorny questions.
https://www.barrons.com/articles/nobel-prize-ai-innovation-disruption-7cd620ce
From Nov 8, 2025 NY Times:
According to an investor offering sheet obtained by DealBook, Blackstone is on the cusp of closing a $3.46 billion commercial-mortgage-backed securities (C.M.B.S.) offering to refinance debt held by QTS, the biggest player in the artificial intelligence infrastructure market. It would be the largest deal of its type this year in a fast-accelerating market. (Blackstone declined to comment.)
The bonds would be backed by 10 data centers in six markets (including Atlanta, Dallas and Norfolk, Va.) that together consume enough energy to power Burlington, Vt., for half a decade.
Blackstone’s offering is part of the latest push in the A.I. infrastructure financing blitz. According to McKinsey, $7 trillion in data center investment will be required by 2030 to keep up with projected demand. Google, Meta, Microsoft and Amazon have together spent $112 billion on capital expenditures in the past three months alone.
The sheer scale of spending is spooking investors: Meta’s stock tumbled 11 percent after the company revealed its aggressive capital expenditure plans last week, and tech stocks have sold off this week on overvaluation fears.
Now, the tech giants are turning to financing maneuvers that may add to the risk. To obtain the capital they need, hyperscalers have leveraged a growing list of complex debt-financing options, including corporate debt, securitization markets, private financing and off-balance-sheet vehicles. That shift is fueling speculation that A.I. investments are turning into a game of musical chairs whose financial instruments are reminiscent of the 2008 financial crisis.
Big tech companies are looking for new sources of financing. While Meta, Microsoft, Amazon and Google previously relied on their own cash flow to invest in data centers, more recently they’ve turned to loans. To diversify their debt, they’re repackaging much of it as asset-backed securities (A.B.S.). About $13.3 billion in A.B.S. backed by data centers has been issued across 27 transactions this year, a 55 percent increase over 2024.
If investors want to buy data center A.B.S., they have two options, according to Sarah McDonald, a senior vice president in the capital solutions group at Goldman Sachs: They can invest in a data center that has one tenant, like a hyperscaler, or in a co-location data center, which has thousands of smaller tenants. The former is an investment-grade tenant with a long-term lease, but the risk is highly concentrated; the latter is most likely renting out to noninvestment-grade tenants with short-term leases, but the investment is extremely diversified.
Digital infrastructure “is something that investors have a huge appetite for,” McDonald said.
Despite the increase in popularity, data center securities are just a small slice of the A.B.S. market, which is dominated by credit card, auto, consumer and student loans.
Blackstone’s $3.46 billion C.M.B.S. offering may seem like small potatoes compared with some other debt-fueled deals, such as Meta’s $30 billion corporate offering to finance its data center in Louisiana. But it’s unprecedented for the C.M.B.S. market, where issuance for data-center-backed deals was just $3 billion for all of 2024.
“They realize how much cash they’re going to need, so they’re getting the C.M.B.S. market comfortable with this type of asset,” said Dan McNamara, the founder and chief investment officer of Polpo Capital, a hedge fund that focuses on C.M.B.S. He added that while most traders in the market were well versed in assets like office space or industrial buildings, with data centers, “it’s not traditional ‘bricks and sticks’ commercial real estate.”
To complicate matters further, the share of single-asset-single-borrower securities (S.A.S.B.) — for example, the assets inside the bond being sold are all from the same company or a single data center — is rising, with 13 percent of all S.A.S.B. deals coming from data centers, according to Goldman Sachs.
“It’s one company, and these assets are quite similar. If there’s a problem with A.I. data centers, like if their current chips are obsolete in five years, you could have big losses in these deals,” McNamara said. “That’s the knock on S.A.S.B.: When things go bad, they go really bad.”
https://www.nytimes.com/2025/11/08/business/dealbook/debt-has-entered-the-ai-boom.html
Laura Chambers, CEO of Mozilla—maker of the Firefox browser—sees the situation as a classic, straightforward bubble. Funding is abundant, it is easier than ever to make low-grade products, and most AI companies are running at a loss, she says. “Yes. It’s really easy to build a whole bunch of stuff, and so people are building a whole bunch of stuff, but not all of that will have traction. So the amount of stuff coming out versus the amount of stuff that’s going to [be sustainable] is probably higher than it’s ever been. I mean, I can build an app in four hours now. That would have taken me six months to do before. So there’s a lot of junk being built very, very quickly, and only a part of that will come through. So that’s one piece of the bubble,” she said.
“I think the most interesting piece is monetization, though. All the AI companies, all these AI browsers, are running at a massive loss. At some point that isn’t sustainable, and so they’re going to have to figure out how to monetize.”
Babak Hodjat, chief AI officer at Cognizant, said he believed diminishing returns were setting in to large language models. The DeepSeek launch from earlier this year—in which a Chinese company released an LLM comparable to ChatGPT for a fraction of the cost—was a good example of this. Building AI was once a huge, expensive, and difficult undertaking. But today, many AI use-cases (such as custom-built, task-specific AI agents) don’t need huge models underpinning them, he said.
“The bulk of the money that you see—and people talk about a bubble—is going into commercial companies that are actually building large language models. I think that technology is starting to be commoditized. You don’t really need to use that big of a large language model, but those guys are taking money because they need a lot of compute capacity. They need a lot of data. And their valuation is based on, you know, bigger is better. Which is not necessarily the case,” he told Fortune.
https://www.aol.com/finance/tech-execs-admit-ai-bubble-141038315.html