NEW YORK — Wall Street cannot shake off concerns that the vast debt taken on to fund artificial intelligence infrastructure could tip individual companies into default now that borrowing costs are high. The collapse in the credit standing of CoreWeave, which rents AI cloud and computing capacity inside data centers to Big Tech firms, has weighed on technology shares day after day. Adding to the strain, a wave of rules restricting data center construction across the United States ahead of the Nov. 3 midterm elections is creating further obstacles for AI infrastructure investment. American opposition to data centers began with economic complaints about the enormous volumes of water and electricity the facilities consume and has spread into religious, philosophical and educational objections that they threaten the value of human existence. Anxiety over AI infrastructure spending has now moved beyond the U.S. market to stock markets in South Korea and elsewhere. With valuations of AI infrastructure companies falling quickly, sentiment toward the sector is unlikely to improve much in the near term.
CoreWeave Aftershocks Persist as Money Shifts From Tech to Blue Chips
On the 20th, the blue-chip Dow Jones Industrial Average, up 1.02%, and the tech-heavy Nasdaq Composite, down 0.02%, diverged throughout the session on the New York market. Unease over AI infrastructure spending, including concerns about distressed CoreWeave corporate bonds, continued to pressure technology shares. Nvidia fell 0.29%, Apple 1.00%, Microsoft 0.31%, Google parent Alphabet 1.13%, SpaceX 0.11%, Facebook parent Meta 1.59%, Broadcom 0.07%, SK hynix 2.16%, Micron 0.98% and AMD 0.62% — all weak even as the Dow rose more than 1% and U.S. Treasury yields declined.
The anxiety was triggered by a report from research firm Gartner on the 14th. Gartner estimated that $3.5 trillion, or about 4,714 trillion won, of AI infrastructure spending would be exposed to depreciation risk by 2030, arguing that revenue from AI infrastructure would fall far short of outlays.
"Given the capital expenditure and financing costs that hyperscalers have announced so far, they would need cumulative revenue of $8.5 trillion by 2030 to cover a minimum 12% return on invested capital, but the actual figure will reach at most $3.6 trillion," Will Somer, a senior analyst at Gartner, said in the report. "Hyperscalers will be able to bring ROIC to 12% only in 2032, and the assets exposed to depreciation risk until then will exceed $4 trillion." Somer added that hyperscalers would spend roughly $1 trillion on AI infrastructure this year and add $6 trillion in investment from next year through 2030, and predicted that more than 80% of neocloud providers would cease operating independently by 2030. Neoclouds are operators that provide cloud services built on graphics processing units and specialized in AI and high-performance computing.
The yield on CoreWeave's bonds maturing in 2032, the largest neocloud operator in the industry, jumped to 14.8% the same day. The price fell to 78% of the face value at which they were sold four months ago, effectively treated as distressed debt. CoreWeave's share price also swung sharply. After plunging 15.46% on the 14th, the stock rebounded 5.94% on the 15th on a single remark from Nvidia Chief Executive Jensen Huang that "financial capability is improving every day," then tumbled 11.91% on the 16th. The stock slipped 0.44% on the 19th and fell more than 6% early on the 20th before barely turning higher, finishing up 1.65%. The probability of default within five years implied by CoreWeave's credit default swap prices has also climbed above 60%. A CDS is a form of insurance bought to guard against the possibility that a particular company defaults. The higher the CDS premium, the greater the market judges the company's default risk to be. CoreWeave operates 51 data centers while carrying $72 billion, or about 96 trillion won, in debt.
Default Odds Jump for Neoclouds as Warnings Reach SoftBank and Amazon
The CoreWeave turmoil has dragged down shares of other neocloud operators including Nebius, IREN and Oracle. According to Morgan Stanley data cited by the Financial Times on the 20th, CDS premiums for many AI-related companies hit record highs this month. Meta's five-year premium rose to 1.08 percentage points and Nvidia's to 1.01 percentage points, while Alphabet and Amazon also reached all-time highs. Oracle's premium climbed to 3.22 percentage points on the 19th, more than double the 1.44 percentage points recorded early this year. Bloomberg reported on the 20th that credit risk at Japan's SoftBank Group, led by Chairman Masayoshi Son, had also risen to its highest in four years on concerns about AI funding. SoftBank's five-year CDS premium reached 4.32 percentage points on the 20th, the highest since early November 2022. It had crossed 4.00 percentage points only a week earlier, then rose another 0.32 percentage point. That was also the highest level among major Japanese companies.
Against this backdrop, Wall Street and major foreign media outlets have been sounding warnings about AI credit almost daily. The New York Times said on the 18th that the key question was whether the debt problems at CoreWeave, Oracle and SoftBank would remain issues at individual companies or spread into a structural obstacle for data center construction as a whole. CoreWeave's business model involves buying AI chips made by Nvidia and others, signing leases of 15 years or more with data center owners and installing the semiconductors there. It then rents that AI computing space on to Meta, Microsoft, OpenAI and others.
CoreWeave's financial burden arises as it bears the costs of leasing and operating the data centers. If AI technology advances to the point where Big Tech firms no longer need the space by the time the leases of 15 years or more expire, CoreWeave would lose the revenue it needs to repay its enormous debt.
James Mackintosh, a columnist at The Wall Street Journal, also warned on the 18th that a similar outcome could unfold this time, comparing the current AI investment boom to Britain's railway investment bubble of the 1840s and the build-out of fiber-optic networks during the dot-com bubble of the mid-to-late 1990s. Stijn Van Nieuwerburgh, a professor at Columbia Business School, estimated in a Brookings Institution report that total U.S. AI-related investment would reach $9 trillion, or 12,000 trillion won, by 2032. That amounts to 3.6% of annual U.S. gross domestic product over the period.
Mark Habib, global head of digital infrastructure at S&P Global, pointed to a failure to realize expected returns as the biggest threat to hyperscalers' credit at a roundtable in London on the 19th. Citing circular financing structures in which hyperscalers guarantee funding for large data center projects or smaller AI firms, Habib said there was "a high likelihood that much more debt exists that we are not aware of." S&P Global said six major hyperscalers — Amazon, Microsoft, Alphabet, Oracle, Meta and SpaceX — would pour more than $7 trillion, or about 9,000 trillion won, of capital expenditure into data centers and AI infrastructure from 2025 to 2030, and expected free cash flow to remain in net outflow through next year.
Morgan Stanley said in a report on the 19th that capital spending by hyperscalers and neocloud operators would reach $1.58 trillion next year, a 60% increase from $980 billion this year. It also projected that net free cash outflows at AI-focused Big Tech firms would hit a record $540 billion next year. Business Insider, citing a recording of an internal meeting it had obtained, reported that net free cash outflows at Amazon Web Services would continue for the next two years. The Information, a U.S. technology outlet, forecast that the several hundred neoclouds operating now would be winnowed to about seven to 10 within a few years.
U.S. Data Center Rules Emerge as Another Major Variable, With Voters in Both Parties Opposed
Even the timelines Wall Street now offers for hyperscalers and neocloud operators to turn profitable assume there will be no policy restrictions. The problem is that even that assumption looks shaky.
The Financial Times reported on the 18th, citing data from regulatory intelligence firm RegLayer, that at least 47 large data center projects had been delayed, withdrawn or denied approval this year. RegLayer said the figure was "a conservative estimate covering only the largest projects." Bernstein, an investment bank, estimated that only about 35% of planned U.S. data center projects were viable.
The Financial Times said its own research found that more than 1,000 communities across the United States had attempted to impose moratoriums or other restrictions on AI infrastructure this year. Oracle, for instance, had planned to bring a 902-megawatt data center in Wisconsin online from 2028 but ran into a delay in approval for grid construction. An Oracle data center project in New Mexico, started through private equity firm Blue Owl, is also struggling to obtain power supply permits. Some of the $18 billion in debt tied to that project recently traded at a distressed 80 cents on the dollar. Both projects are part of Stargate, the vast data center venture led by SoftBank and OpenAI.
Opposition to data centers is now a major issue in the U.S. midterm elections. Many residents, regardless of which party they support, are fiercely opposing data centers in their own areas, citing surging electricity bills, depletion of water resources and noise. A perception has taken firm hold nationwide that data centers threaten residents' safety, drive up living costs and enrich large corporations while creating almost no jobs.
New York was the first state to move against AI data center construction. Gov. Kathy Hochul signed an executive order on July 14 — the first of its kind — suspending construction of large data centers for one year. Its core provision temporarily halts environmental permits for up to a year for large data centers that use 50 megawatts of power or more. Chicago Mayor Brandon Johnson signed an executive order on Aug. 11 calling for a temporary moratorium for the same reason. That order prompted the Chicago City Council, where opposition members hold an overwhelming majority, to begin reviewing a moratorium. Pennsylvania decided on Aug. 18 to require all data center projects applying for permits to comply with the Governor's Responsible Infrastructure Development requirements, known as GRID, a set of rules covering relief from energy costs, community engagement, workforce and economic development, transparency and environmental protection.
New Jersey Gov. Mikie Sherrill signed a bill on Aug. 27 requiring data centers in the state to report water and electricity use statistics to the state Board of Public Utilities every six months. Maine Gov. Janet Mills also used her veto earlier this year to put data center construction on hold. In Texas, emerging as a new hub for the facilities, Gov. Greg Abbott ordered energy audits of new data center projects. Despite President Donald Trump's advocacy for data centers, Abbott, a member of the same party, told ABC in July that companies building data centers "failed to win community support and dug their own grave."
In a Gallup poll conducted in March and released in May, 71% of Americans said they opposed data center construction in their area, citing environmental concerns above all. Those who said they "strongly oppose" accounted for 48%. In a national survey by Emerson College in the United States, opposition to data centers jumped to 63% in July from 42% last December, while support fell to 27% from 33% over the same period. The survey found that negative views rose with education level, with opposition reaching 79% among holders of master's degrees. An August survey by the Annenberg Public Policy Center at the University of Pennsylvania also found opposition to new data centers in respondents' own areas rising to 61% in July from 49% in March. In that survey, opposition was highest among adults under 30, a group expected to be favorable toward new technology, at 70%.
Anthropic, Preparing to List, Posts 107 Trillion Won Net Loss in First Half as Global IPOs Reel
Beyond worries about damage to local infrastructure, a considerable number of people object on grounds of religious belief or job losses. On July 25, when news emerged that the U.S. Environmental Protection Agency was moving to scale back public comment procedures for AI data center construction, about 200 civic groups and more than a dozen state governments voiced opposition. Jang Sung-kwan, head of U.S. strategy at D&A Advisory, a U.S. political consulting firm, told a briefing for Korean correspondents hosted by the Korean Consulate General in New York on the 14th of last month that "ordinary American citizens did not even form positive or negative opinions last year because they did not know what an AI data center was, but since May this year even centrist governors from both parties have been pushing policies to block construction." He added, "Ordinary citizens lump AI together with issues such as unmanned surveillance cameras, and because President Trump keeps making remarks defending data centers, Republican candidates in swing states are in an awkward position."
As sentiment toward AI investment sours, investors are also viewing companies that recently listed on global markets or are preparing initial public offerings far less favorably. According to a securities filing obtained by The Wall Street Journal on the 20th, Anthropic posted a net loss of $80.7 billion, or about 107 trillion won, in the first half of this year, more than double the $40.2 billion recorded for all of last year. Revenue rose sharply to $16.3 billion, or about 22 trillion won, in the same period this year from $1.1 billion in the first half of last year, but losses ballooned as well. Anthropic spent $9 billion, or about 12 trillion won, on securing computing capacity and infrastructure in the first half alone, more than the $7.3 billion it spent in all of last year. Its computing contracts with the world's three largest cloud providers — Google, Amazon and Microsoft — amount to $450 billion, or about 597 trillion won. Adding a contract to lease Google chips through Broadcom, worth more than $160 billion, and deals with smaller cloud operators pushes future infrastructure costs above $600 billion. Including additional negotiations now under way, the company plans to invest up to $750 billion, or about 995 trillion won, on securing computing capacity alone.
According to Reuters, at a pre-IPO presentation held at Anthropic's San Francisco headquarters on the 14th, some investors were disappointed by a slowdown in the company's growth and discussed a valuation of about $1.5 trillion, or roughly 2,000 trillion won, far below the $2 trillion, or about 2,700 trillion won, initially expected. Anthropic, which submitted a draft registration statement to the U.S. Securities and Exchange Commission on June 1, had originally planned to list in New York this month but has pushed the timing back to after the Nov. 3 midterm elections.
Shares of Elice, a South Korean AI infrastructure company, closed at 61,500 won on the 20th, its first day of trading on the KOSDAQ, 32.04% below its offering price of 90,500 won. The debut landed as a considerable shock on Wall Street. In an article headlined on the collapse of an AI company's listing stoking fears that the AI investment boom is ending, Bloomberg said that Elice, which had been little known outside the Korean market, "has found itself at the center of a fierce debate over the valuations of AI companies." Bloomberg described the case as emblematic of weakening investor confidence in neocloud companies. Reuters said Elice's first-day plunge "amplified investor concerns about AI spending" and saw it as fueling weakness in AI-related shares in New York that day. CNBC said that while the size of Elice's IPO was relatively small, investors were treating it as a signal of sentiment toward AI infrastructure investment.
The Financial Times argued on the 18th that the recent weakness in AI infrastructure shares stemmed from a string of abandoned IPOs by large Australian data center companies. Firmus, a data center operator backed by Nvidia and Blackstone, had sought a valuation of 44 billion Australian dollars, or about 41 trillion won, which would have been the second-largest in Australian history, before abruptly withdrawing its IPO plan on the 10th. AirTrunk, another Australian data center operator, also canceled a plan to list a data center real estate investment trust in Singapore, citing deteriorating market conditions.
Microsoft canceled a lease with Nscale, a British AI data center operator pursuing a New York listing. Bloomberg also reported that Nscale had argued for an IPO valuation of $30 billion, or about 40 trillion won, while some investors were pushing back with proposals of $16 billion to $19 billion, or roughly 22 trillion to 26 trillion won. Nscale publicly filed listing documents with the New York Stock Exchange on the 18th of last month but has been unable to hold a formal roadshow even after the mandatory 15-day waiting period, as it works through SEC comments. The Financial Times reported on the 18th that Anthropic, which along with Microsoft is one of Nscale's two biggest customers, could also walk away from a $45 billion contract if the company fails to meet its funding timeline. As of August, Nscale was running 25,000 advanced AI chips, far short of the 500,000 it plans to operate.
With views of AI infrastructure investment darkening both inside and outside the industry, technology shares are unlikely to rise sharply for some time. AI-related companies could still engineer a short-term rebound by unveiling striking results during earnings season. The difference is that high interest rates and American resistance to data centers are likely to last longer than that. Whether the current episode ends as a financial problem at a handful of companies or spreads into broader distress across the AI infrastructure industry remains to be seen.
Note: "Trump Stocker" is a column delivering on-the-ground reporting and analysis of U.S. markets, companies, policy, politics and diplomacy that may be useful for investors during President Donald Trump's term. Subscribe to receive news from the United States.