Bond markets are now tightening credit for lower-rated borrowers, and one of the clearest examples is CleanSpark, a company building data centers for Meta that had to make major concessions to investors earlier this month just to secure financing.

That deal stands as one of the most telling signals that, across the board, the cost of funding the data center construction boom is rising and investors are becoming pickier about which new projects they back. People involved in arranging the transactions say the bank loan market is showing similar strain, with some lenders, including Societe Generale, Sumitomo Mitsui Banking Corporation, and Mitsubishi UFJ Financial Group, applying stricter screening standards when making loans to data center projects.

This market dynamic means companies' planned data center projects may struggle to borrow, which in turn could drag down the entire AI industry's expansion plans. The AI financing market is facing multiple pressures at once. Major cloud service providers such as Amazon, Google, and Microsoft are spending a combined roughly $700 billion in capital expenditure this year and are expected to maintain similar levels for years to come.

These companies have already issued nearly $160 billion in investment-grade bonds this year, flooding the market with a massive supply of new debt. Although these tech giants generate enormous operating cash flow, their bond premiums are still rising relative to other high-grade corporate debt. This year, investors are demanding an extra yield of about 0.25 percentage points to hold these hyperscaler bonds, while the entire investment-grade bond market has seen an increase of only 0.04 percentage points.

At the same time, several large tech companies are shifting construction spending onto partners that raise money in the high-yield bond market. Developers building data centers for AI companies such as Anthropic and OpenAI also rely on that market for financing. Combined, the high-yield bond market has issued about $55 billion in AI-related bonds this year. Hit by a combination of investor risk aversion toward the AI sector and broader macroeconomic pressures, the cost of financing terms on some projects has risen noticeably in recent weeks.

CleanSpark, once a bitcoin miner that has now expanded into AI data center operations, offered one of the largest discounts of the past year when it issued $2.3 billion in bonds: investors only had to pay 98.5 cents to receive a bond with a face value of $1. Data from Morgan Stanley shows that all four data center high-yield bond deals completed since July were issued at a discount. Of the 20 similar deals in the past 12 months, Morgan Stanley led 13, while in the first 10 deals during that period, only three were issued at a discount.

A discount gives investors higher potential returns beyond the coupon, because they can recover the full principal at maturity. The bond's 7.875% coupon rate is already on the high side compared with similar projects. CleanSpark also agreed to amortize principal in installments, reducing refinancing risk for bondholders. Such concessionary terms are now increasingly common in data center debt.

The CleanSpark deal is particularly noteworthy: end user Meta is itself an investment-grade company, which would normally give investors confidence that the project would be repaid. This shows that investors are demanding extra compensation to cover construction and other project risks. Connor Minar, a fixed-income portfolio manager at Manulife Investment Management, said such discounts "are a relatively new phenomenon in the current market." He said: "Earlier this year, investors were much more accepting," willing to participate regardless of the deal's structure.

A similar situation emerged in mid-August: developer Zenith Arc LLC, backed by a fund under Coatue and infrastructure startup Fluidstack, issued bonds at 99.5 cents on the dollar to build a data center that would be leased to trading giant Jane Street Capital. Bond prices and yields move inversely, and as investors demanded greater risk compensation than on existing bonds of the same rating, the yield on Zenith Arc's debt surged further.

Yields on some projects have risen to levels that make it difficult for borrowers to afford project costs. For example, if a project needs a 12% internal rate of return 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 high inflation will persist, forcing the Federal Reserve to raise rates aggressively.

The AI boom itself is also adding pressure. Federal Reserve Chairman Kevin Warsh said last week that higher Treasury yields are partly due to large-scale debt issuance by tech companies, which is crowding out market capital. For data center developers, the flood of new debt in the market is forcing investors to carefully distinguish among AI projects. Rising Treasury yields mean riskier projects must offer higher returns to attract investors.

At a critical stage of construction expansion, a group of more speculative projects is now struggling. A banker involved in related deals said investors have begun to more clearly distinguish between safe projects and high-risk ones, a shift he called "quality selection" at the credit rating and project level. 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 invest in safer bonds for similar returns.

At the same time, volatility in the U.S. Treasury market is weighing on the IPO window, another major potential source of funding for companies. A draft IPO filing disclosed in reports shows that Anthropic posted a net loss of $42 billion last year while committing more than $500 billion to computing power and infrastructure, highlighting its enormous funding needs. SB Energy and Nscale, which build data centers for OpenAI and Anthropic, also need huge amounts of financing themselves. Both companies have publicly filed IPO documents in recent weeks, but the timing of their listings remains uncertain.

The market had originally expected Anthropic to file for an IPO as early as this month, but it has not yet done so. Underwriting Syndicates Shrink Developers had grown used to vigorous demand in the bond market, with almost all data center-related projects able to secure funding relatively easily, but investor attitudes have now reversed. If borrowers are unwilling or unable to meet the concessions investors currently demand, they must find other financing channels.

The bank loan market is an alternative, but it has its own difficulties. Loan negotiations usually take longer, and lenders often demand stricter protective terms than bond investors. Some project debt also stipulates that developers can only request the next tranche of funds from lenders in stages after completing specific milestones. Two people familiar with current financing business said some lenders are being more cautious when screening deals.

The pool of banks able to form syndicates and lead multibillion-dollar loans for large data center and AI infrastructure projects has therefore shrunk. One of the people said Societe Generale and Sumitomo Mitsui Banking Corporation have tightened project access standards; another person said Mitsubishi UFJ, one of the industry's main lenders over the past two years, has also begun reducing its exposure to the business. These banks had previously participated in several blockbuster deals: Societe Generale led $7.1 billion in debt financing for the first site of OpenAI and Oracle's "Stargate Project"; Mitsubishi UFJ and JPMorgan Chase co-led $38 billion in financing for two other Oracle projects; and Sumitomo Mitsui helped lead $18 billion in financing for Oracle's New Mexico project.

Of course, market participants say no major deals have been canceled so far, and no ongoing syndication process has been halted because of poor pricing. Project developers are still absorbing higher costs in the high-yield bond market to keep projects moving forward. Minar said: "At this stage, the core feature is still all kinds of concessions. For many companies, the priority is to get computing power in place as quickly as possible, while financing costs take a back seat." But the recent troubles at Oracle's New Mexico project have sounded an alarm for lenders.

The project completed $18 billion in financing last year through the syndicated loan market. The incident prompted banks to re-examine loan-related contract terms and restrictive covenants. Last week, Oracle sent a force majeure notice to the project developer, an entity under Blue Owl Capital, after the project suffered billions of dollars in cost overruns. Oracle sought to invoke a contract clause that could exempt or delay its performance obligations if events beyond its control disrupt the project.

People familiar with the financing said that from the lenders' perspective, the loan is "structurally sound"; even if the site cannot be powered up, Oracle still has to pay rent. But a banker said Oracle's invocation of force majeure may prompt some banks to add stronger protective clauses, reassess lending risk, and further squeeze credit supply. Key Takeaways: AI data center construction relies heavily on debt financing, and rising U.S. Treasury yields combined with declining investor risk appetite are leading to bond discounts and stricter financing terms.

The force majeure dispute over Oracle's project is further pushing banks to strengthen risk controls on infrastructure projects; as funding tightens, some AI data center projects with thin returns may be shelved.