The bond market is becoming increasingly tough on lower-rated borrowers. CleanSpark is one example: the company, which is building a data center for Meta, had to make significant concessions to investors earlier this month just to secure financing. This deal is one of the clearest signals that financing costs across the data center construction boom are rising broadly, and investors are becoming more selective when choosing new projects. People arranging the deals say the bank loan market is also showing signs of strain, with some lenders—including Societe Generale, Sumitomo Mitsui Banking Corporation, and Mitsubishi UFJ Financial Group—choosing to be more cautious when approving loans for data center projects.

This market dynamic means companies may struggle to borrow the funds needed to build their planned data centers, which in turn threatens the entire AI industry's expansion plans. The AI financing market faces multiple pressures. Major cloud providers such as Amazon, Google, and Microsoft are spending a combined total of roughly $700 billion in capital expenditure this year, and are expected to maintain a similar scale for years to come. These companies have already issued nearly $160 billion in investment-grade bonds this year, flooding the market with massive new supply. Although tech giants have abundant operating cash flow, the financing costs for these hyperscaler bonds have risen more than other high-grade corporate bonds. This year, the spread investors demand on hyperscaler debt has widened by about 0.25 percentage points, while the entire investment-grade bond market has risen by only 0.04 percentage points. At the same time, these large tech companies are shifting a great deal of capital expenditure onto partners that raise funds in the high-yield bond market. Meanwhile, developers building data centers for AI companies such as Anthropic and OpenAI are also heavily reliant on that market. In total, about $55 billion in AI-related bonds have been sold in the high-yield market this year. Affected by a combination of risk aversion among AI-sector investors and macroeconomic pressures, the cost of financing terms on some projects has risen notably in recent weeks. CleanSpark, originally a Bitcoin miner that has expanded into AI data center operations, issued $2.3 billion in bonds at one of the largest discounts seen in nearly a year: investors only had to pay 98.5 cents to subscribe to bonds with a face value of $1. Morgan Stanley data shows that since July, all four data center high-yield bond issuances have priced at a discount. Of the 20 similar deals in the past 12 months, Morgan Stanley led 13; and of the first 10 deals in that 12-month period, only three were issued at a discount. A discount can enhance investors' potential returns beyond interest payments, because at maturity investors get back the full principal. The bond carries a coupon of 7.875%, which is already high compared with similar projects. CleanSpark also agreed to amortize principal in installments, reducing bondholders' refinancing risk. These two types of concession terms are now increasingly common in data center debt. The fact that the CleanSpark deal required so many compromises is especially noteworthy: the end user, Meta, is itself an investment-grade company, which would normally give investors assurance of project repayment. This shows that investors still demand extra compensation to cover project-level risks such as construction. Connor Menear, a fixed-income portfolio manager at Manulife Investment Management, said such discounts "are a newly emerging phenomenon in the current market." He said that earlier this year, investors were much more willing to accept various deal terms. A similar situation occurred in mid-August: Zenith Arc LLC, a developer backed by a fund under Coatue and infrastructure startup Fluidstack, issued bonds at 99.5 cents on the dollar to finance a data center leased to trading giant Jane Street Capital. Bond prices and yields move inversely; even compared with existing bonds of the same rating, investors still demanded higher risk compensation, and Zenith Arc's bond yield climbed further. In some cases, yields have risen to levels that make it difficult for borrowers to bear project costs. For example, if a project needs a target return of 12% and borrowing costs approach that threshold, profit margins would be extremely compressed and the project would no longer be financially viable. AI Risk Overlaps With Macro Market Volatility Meanwhile, Treasury yields have risen sharply, partly because of market concerns that persistent high inflation will force the Federal Reserve to raise rates aggressively. The AI boom itself is also adding pressure: Federal Reserve Chairman Kevin Warsh said last week that part of the rise in Treasury yields is because tech companies are issuing large amounts of debt, crowding out other investors' capital. For data center developers, the massive new supply of bonds across the market is forcing investors to carefully distinguish between different AI projects. Higher Treasury yields mean riskier projects must offer higher returns to attract investors. At a critical stage of expansion, many more speculative projects are finding themselves in trouble. A banker involved in related deals said investors have begun to clearly differentiate between safe projects and high-risk projects—a shift toward "quality selection" across both credit ratings and project levels. In the past, investors were willing to take on projects with long construction cycles and inexperienced developers in exchange for high yields; now they may abandon such projects and instead choose safer bonds for similar returns. Violent swings in the Treasury market are also hurting the IPO window, cutting off another potential source of cash. According to draft IPO disclosures, Anthropic posted a net loss of $42 billion last year while committing more than $500 billion to computing power and infrastructure, highlighting its enormous financing needs. SB Energy and Nscale are building data centers for OpenAI and Anthropic and themselves need huge amounts of capital; both companies filed for IPOs in recent weeks, but listing dates have not yet been set. The market had originally expected Anthropic to file for an IPO as early as this month, but it has still not launched one. Underwriting Syndicates Shrink In the past, when bond market demand was strong, almost all data center-related projects could obtain funding relatively easily; now investor sentiment has reversed, dealing a blow to developers. If borrowers are unwilling or unable to meet the new concession terms demanded by investors, they have to look for other financing channels. One alternative is the bank loan market, but that also faces challenges. Loan negotiations generally take longer, and lenders often demand more protective covenants than bond investors. Some project debt also stipulates that developers can only apply to lenders for the next tranche of funds after reaching specific milestones. Two people familiar with current financing projects said some lenders have become more cautious when selecting deals. This has shrunk the pool of banks willing to participate in syndicated loans, while data centers and other AI infrastructure projects require large syndicated loans worth billions of dollars. One of the people said institutions such as Societe Generale and Sumitomo Mitsui Banking Corporation are reviewing deals more strictly; a second person added that Mitsubishi UFJ, a major lender in large industry deals over the past two years, is also scaling back. These banks had previously participated in many blockbuster deals: Societe Generale led the $7.1 billion debt financing for the first site of OpenAI and Oracle's "Stargate Project"; Mitsubishi UFJ and JPMorgan jointly led $38 billion in financing for two other Oracle projects; Sumitomo Mitsui participated in leading $18 billion in financing for Oracle's New Mexico project. Of course, market participants say no major deals have been canceled so far—that is, no syndicated project currently in progress has been terminated because of poor pricing. In the high-yield bond market, project developers are still willing to bear higher costs to push projects forward. Menear said: "At this stage, the core is still making concessions continuously. For many companies, the primary goal is to bring enough computing power online as quickly as possible, and financing costs take a back seat." But the recent troubles at Oracle's New Mexico project, which secured an $18 billion syndicated loan last year, have sounded an alarm for lenders and prompted banks to carefully review the contracts and covenants behind the loans. Last week, Oracle sent a force majeure notice to the developer, an entity under Blue Owl Capital, after the project encountered unexpected cost overruns. Oracle sought to invoke contractual clauses that allow it to excuse or delay performance obligations when events beyond its control interfere with the project. People familiar with the financing said that from the lenders' perspective, the loan was well structured, and even if the site cannot be powered, Oracle cannot be relieved of its lease payment obligations. But a banker said Oracle's invocation of force majeure may prompt some banks to demand stronger protections and reassess lending risk, further tightening credit supply.