AI demand surges, but the financing "credit gate" is tightening! Under the heavy pressure of the 5% U.S. Treasury yield, compute power losers are gradually emerging

According to Zhitong Finance APP, companies in the stock market that are benefiting from the unprecedented wave of AI computing power infrastructure construction are simultaneously facing expanding demand and significantly tightened financing conditions: AI agent applications are bringing a continuously growing demand for AI computing resources, while soaring long-term U.S. Treasury yields above 5% have sharply raised the capital threshold for converting excess AI orders into actual cash flow, due to the higher cost of funds.

In the week ending September 25, CoreWeave's stock price rose nearly 8%, while Oracle fell around 7% and is down roughly 30% year-to-date, indicating that investors are beginning to scrutinize the financing, delivery, and cash flow capabilities of different companies more carefully. JPMorgan predicted in June that by 2030, AI infrastructure construction would require about $4.1 trillion in debt financing—a cumulative multi-year estimate, not a single-year issuance, nor entirely equivalent to corporate bonds. When such massive construction demand continuously enters the capital markets, interest rates, credit conditions, and funding availability directly impact the pace of expansion.

Muse and Astra, which represent the hottest AI agent applications, have become new and sustainable engines of robust growth for computing power demand. A single user command can trigger planning, retrieval, browser operations, code execution, and result validation; multiple rounds of model calls further increase the need for context processing and state preservation. Mechanically, GPUs handle model computation, CPUs execute tools and task orchestration, HBM, server DRAM, and enterprise-grade SSDs collectively support data access and cache management, and data center optical interconnects enable high-speed data transmission. A detailed engineering analysis by Nvidia highlights that CPU execution speed affects GPU utilization, while tiered caching through GPU memory, CPU memory, local NVMe, and remote storage reduces redundant computation. Therefore, the investment value brought by agent penetration must ultimately manifest in each unit of capital being able to deliver more high-complexity, billable tasks, meet larger AI computing resource demands, and generate stronger operating cash flows for enterprises.

The changes on the funding side are undoubtedly global. On September 25, the yields on U.S. 10-year and 30-year Treasuries briefly touched approximately 5.23% and 5.53%, respectively—the highest levels since 2007 and 2004. Energy supply shocks, sticky inflation, and resilient economic demand have prompted the market to revise up its expectation for future policy rate paths; the Federal Reserve raised rates by 25 basis points in September, and a Charles Schwab analysis on September 25 shows that the futures market began pricing in nearly three additional rate hikes by June 2027. As the "anchor for global asset pricing," a higher 10-year Treasury yield raises the pricing benchmark for new long-term financing, and a higher required rate of return suppresses the valuation of future corporate cash flows.

The “Credit Gate” of the AI Frenzy: Financing Expansion and Tighter Qualifications May Occur Simultaneously

The current AI computing power expansion boom appears to be undergoing a “credit gate” test—demand determines how much companies want to raise capital, while lenders decide which orders and projects are eligible for funding.

An executive from Mitsubishi UFJ Lease’s U.S. subsidiary believes that only part of the candidate list of new cloud enterprises will truly win market support; this is an industry practitioner’s judgment, but it reveals a critical shift: higher rates can compensate for some risk, but if projects lack reliable power, delivery assurance, or executable customer contracts—or if an AI project relies excessively on a single large client—they may not secure financing by simply raising the bid. As a result, overall sector borrowing may still grow, but capital will increasingly concentrate with companies that have stronger financial cushions, high-quality contracts, and robust delivery capability.

Interest rate shocks also differentiate by debt structure. Newly issued or refinanced fixed-rate debt faces repricing; outstanding floating-rate loans usually adjust in line with short-term benchmarks such as SOFR, so the rise in 10-year Treasury yields doesn’t directly translate into all loans. CoreWeave’s SEC filings show that, based on its outstanding floating-rate debt as of June 30, 2026, a 100-basis-point increase in rates would result in approximately $30 million in extra interest expenses over three months, or about $61 million over six months. The company also disclosed it uses interest rate swaps to mitigate some of this risk. These numbers reflect the sensitivity of specific outstanding debts, while future incremental borrowing, changes in financing costs, and credit spreads will further impact the actual cost of funds.

SoftBank demonstrates the continued strength of financing demand. On September 24, it finalized bond issuances totaling approximately $11.1 billion, consisting of $10 billion in dollar bonds and 1 billion euros in euro bonds; the longest-dollar bond matures in 7.5 years with a coupon of 9.75%, issued at par, and is expected to settle on September 29. A portion of the funds raised will go toward follow-up investments in OpenAI. This suggests that when a large company believes missing the AI window would incur a higher cost, it may still accept expensive capital, although a higher interest burden means that future investment returns need to clear a higher hurdle rate.

Another variable that robust demand cannot easily offset is the time gap between construction expenditure and operating cash inflows. Equipment purchases, engineering construction, and financing costs are all incurred upfront, while income depends on power connection, project acceptance, and customer usage. “Force majeure” controversy over Oracle’s Project Jupiter in New Mexico involves arrangements for contract payments in the event of potential delays; previous media reports citing insiders indicated related arrangements might extend the lower rent phase for the construction period and delay the start of higher operational rents, although Blue Owl Capital said its financial commitment remains unchanged. This shows that the financing value of long-term AI infrastructure orders still depends on delivery conditions and risk-sharing mechanisms, and cannot be counted as cash-in-hand. For investors, the “credit gate” ultimately screens for AI cloud computing and computing power companies that can convert demand into timely debt service cash flows while maintaining returns to shareholders after interest payments.

AI Computing Power Demand Frenzy vs. Financing Pressure: Soaring Bond Yields Raise the Risks for AI Infrastructure Firms Hungry for Debt Financing

As U.S. Treasury yields rose this week to their highest level since 2007, companies dependent on debt financing will face increased borrowing costs. This means that the already historic scale of AI infrastructure buildout is about to become even more expensive.

In a deep-dive report in June, JPMorgan estimated that, by 2030, $4.1 trillion in AI-related debt could be issued. Data center companies and others caught up in the AI craze are racing to expand capacity to meet “nearly insatiable” demand, according to many industry experts.

When borrowers re-enter the financing market, they are now met with a 10-year U.S. Treasury yield approaching 5.17%, up roughly a full percentage point since the start of the year. This means debt issuers must offer more attractive returns to lure investors.

The market has not yet panicked—at least not yet. Heavily indebted new cloud company CoreWeave’s stock remains robust, up nearly 8% this week; whereas Oracle, with high financial leverage and dependent on long-term debt markets to support its AI expansion, has fared worse, down 7% this week and about 30% year-to-date.

Meanwhile, Japanese investment giant SoftBank, one of the main capital providers to AI projects, raised $11.1 billion this week by issuing junk bonds, with the 7-year tranche yielding as much as 9.75%.

“They’re basically price-insensitive for this funding, which means they are price takers. In my view, a lot of these companies must be less sensitive to price. They need to raise as much capital as possible to compete,” said Mark Malek, chief investment officer at Siebert Financial, in an interview.

At the heart of the AI boom are leading model developers OpenAI and Anthropic, with private market valuations both topping the $1 trillion “super” threshold. To provide these advanced models—and the supporting infrastructure for the models and services of many others—tech giants Amazon, Google, Meta, and Microsoft have this year pledged hundreds of billions of dollars in capital expenditures, expected to increase to $1 trillion in AI capex by 2027.

While a significant portion of these investments are funded by debt, these tech giants all possess investment-grade credit ratings, allowing them to access capital at lower cost. Still, some market participants believe other companies may face greater difficulties ahead.

According to reports citing insiders, a seasoned private credit investor told the media that future financing of new cloud projects will become much tougher as these companies have less of a cushion to absorb costs. This investor requested anonymity to speak frankly.

Riley Thompson, Vice President at Mitsubishi HC Capital America, said in an interview that even as borrowers accept higher rates, lenders are being more selective about which projects they are willing to fund.

“If you list 50 new cloud companies, the market may only really be interested in 20 of them,” said Thompson.

CoreWeave, which went public last year, warned of the risk of rising interest rates in its SEC filings. In its latest quarterly report, the company said that based on its outstanding floating-rate debt as of June, every 100-basis-point (i.e., one percentage point) rise in rates could increase its interest costs by $30 million.

This week may have already seen an early warning sign. On Friday, according to sources cited by the media, Oracle filed a “force majeure” notice for its New Mexico data center project to avoid higher costs, sending its stock price down. If the so-called Project Jupiter campus fails to go live as planned by 2028, Oracle hopes to delay associated payments. Oracle said the project “is still moving forward according to our established timetable.”

Rising interest rates are not the only current challenge. Before yields spiked this week, the CEOs of Anthropic and OpenAI had already begun calling for a slowdown in AI development—following public concerns from industry researchers about the risk of advanced models escaping human control.

Meanwhile, as the U.S. heads into midterm elections in November, nationwide opposition to AI data centers has become a major topic. NBC News Decision Desk, in a recent SurveyMonkey-powered poll, found that 69% of respondents opposed building such facilities in their region, mostly due to unprecedented spikes in residential electricity and water bills and carbon emissions resulting from the AI infrastructure boom. On Monday, Texas Republican Governor Greg Abbott, who faces a hotly contested re-election bid, ordered a temporary halt to all environmental permit approvals for data centers, after pausing grid access approvals last month.

Nonetheless, AI service demand continues to surge. One recent example is Meta’s personal assistant app Muse, which has rapidly gained popularity since its launch in early September. In its first two weeks online, Muse was downloaded over 2.5 million times globally and topped the Apple App Store chart, surpassing ChatGPT. According to Evercore’s Mark Mahaney, who spoke to CNBC this week, Muse could reach 100 million users in six to twelve months.

Andrew Judice, head of global corporates, project, and infrastructure finance at rating agency KBRA, said that while higher rates may affect future deals, he has yet to see a major impact on borrowing demand.

“In a normal environment, people might take a step back and pause a little. But I don’t see that happening here. I think you’ll continue to see significant levels of issuance,” Judice said.

Chaim Zaltzman, vice chair of emerging companies and growth at Latham & Watkins, said there’s no doubt that someone has to bear these rising costs. As someone who’s worked in core AI infrastructure financing for years, Zaltzman said: “But when you have demand that’s this robust, it’s much easier to digest these costs.”

For Bernie Margulis, CEO of American Compute, which advises on risk management for GPU financing, the calculation is even simpler. He said borrowers remain eager for financing, even at higher costs—especially those with contracts with OpenAI or Anthropic, both of which have signed deals locking in computing power supply for years to come.

“If you have a big multi-year contract with Anthropic, is an extra 50 basis points really going to make you stop?” Margulis said.