Meta Data Center Deal Tests Limits of AI Debt Markets

Bond markets are getting tough for lower-rated borrowers. One example: CleanSpark, which is developing a data center for Meta Platforms, had to offer investors big concessions to land financing earlier this month.

That deal is among the clearest signs yet that financing the data center boom is getting expensive across the board, and investors are getting picky about which new projects they’ll back. The bank loan market too is showing signs of strain, with some lenders like Société Générale, Sumitomo Mitsui Banking Corp. and Mitsubishi becoming more selective in lending to data center projects, people arranging the deals say. Those dynamics mean companies could struggle to borrow for planned data centers, jeopardizing growth plans for the broader AI industry.

Pressures on the AI financing market are coming from a number of different directions. Big data center operators like Amazon, Google and Microsoft will together spend some $700 billion on capital expenditures this year, and are expected to continue at around that level the next few years. They’ve issued nearly $160 billion in investment-grade debt this year, flooding the market with new supply.

That hyperscaler debt is getting more expensive relative to other highly rated corporate bonds, though the tech companies generate enormous operating cash flows. The extra yield investors are demanding to hold hyperscaler debt has risen by about 0.25 percentage points this year, compared with just 0.04 percentage points for the broader investment-grade market.

At the same time, several of these big tech companies are offloading spending onto other ventures raising money in the high-yield bond market. Meanwhile, developers building data centers for AI firms like Anthropic and OpenAI are leaning on the market too. All told, the high-yield market has seen around $55 billion of AI-related bonds sold this year.

Terms for some of those projects have gotten noticeably costlier in recent weeks, thanks to a collision of AI-specific investor wariness and broader macro pressures. The project from CleanSpark, a bitcoin miner expanding into AI data centers, had to offer one of the biggest discounts of the last year to sell $2.3 billion in bonds, making investors pay just 98.5 cents to buy a dollar in principal of bonds.

All four high-yield data center bond deals going back to July have been sold at some type of discount, according to data from Morgan Stanley, which has led 13 of the 20 deals over the last 12 months. Of the first 10 during that period, just three included a discount.

Discounts boost investors’ return potential beyond the interest payments because they get back the full principal. The coupon of 7.875% was already high compared to similar projects. CleanSpark also agreed to repay the principal over time, reducing the refinancing risk for bond buyers.

Both moves are concessions that have become more common for data center debt. That the CleanSpark deal needed them was especially notable because the end user, Meta Platforms, is an investment-grade company, which typically has given investors comfort that a project will pay off. That suggests investors demanded extra compensation for construction and other project risks.

Such discounts are “a newer development in the current market,” said Connor Minnaar, a fixed-income portfolio manager at Manulife Investment Management. “Early in the year, there was a lot more appetite” from investors to accept whatever structure the deals had, he said.

A similar dynamic played out in mid-August, when developer Zenith Arc LLC, backed by a Coatue Management venture and infrastructure startup Fluidstack, sold debt at 99.5 cents on the dollar to finance a data center to be leased by trading giant Jane Street. Yields, which move inversely to prices, spiked further for the Zenith Arc debt once it started trading, as investors sought extra compensation even compared to older but comparably rated project bonds.

In some cases, yields may be rising to levels where borrowers can’t afford to pay for the project costs. If, for example, they need returns of 12%, borrowing costs approaching that threshold squeeze profit margins so tightly that the project may no longer make sense financially.

AI Risks, Macro Volatility Collide

Meanwhile, government bond yields are spiking, fueled in part by fears that persistently high inflation will prompt aggressive Federal Reserve interest rate hikes. The AI boom itself is adding to the pressure—Fed Chair Kevin Warsh said last week that government yields have climbed in part because heavy tech debt issuance is crowding out investors.

For data center developers specifically, the rush of new debt across markets is forcing investors to take a closer look at what differentiates one AI project from another. Rising government yields mean riskier projects have to offer even higher compensation to entice investors. That’s put more speculative projects in a tough position at a critical moment for the build-out.

Investors have started drawing sharper distinctions between safer and riskier projects, according to one banker working on some of the deals, who called the shift “a move up in quality” across credit ratings and projects. Investors who once took more risk on projects with long construction timelines or less experienced developers in exchange for yield may pass them up to get similar returns on safer debt, the person said.

Meanwhile, volatility in the Treasury markets is causing the initial public offering market to wobble, jeopardizing another potential source of fresh cash. Anthropic’s $42 billion net loss last year and more than half a trillion dollars in compute and infrastructure spending commitments, highlighted in reported draft IPO paperwork, underscore its enormous financing needs.

SB Energy and Nscale, which are developing data center capacity for OpenAI and Anthropic and will need substantial financing themselves, made their IPO paperwork public in recent weeks, but the timing of their debuts remains unclear. Anthropic had been expected to reveal its IPO filing as soon as earlier this month but hasn’t pulled the trigger.

Dwindling Syndicates

Investor pushback is an abrupt change for developers accustomed to previously strong bond market demand that helped most anything related to data centers land financing with relative ease. Borrowers that are unwilling or unable to offer the concessions investors are now demanding may be forced to look elsewhere for financing.

One alternative is the bank loan market, but that comes with its own challenges. Deals are generally slower to arrange and lenders often demand more protections than bond investors. And some project debt requires developers to return to lenders periodically to get the next slug of financing once certain milestones are met.

Some lenders are becoming more selective about the deals they’ll participate in, according to two people with knowledge of current financings. That narrows the pool of banks available to syndicate large multibillion-dollar loans needed for data center projects and other AI infrastructure.

Société Générale and Sumitomo Mitsui are among the firms that have become more selective, one of the people said, while the second person said that Mitsubishi, which has been a larger lender to the industry over the last couple of years, is also stepping back.

Those banks had previously funded high-profile deals, with Société Générale leading a $7.1 billion debt deal for the first site in OpenAI and Oracle’s Project Stargate. Alongside JPMorgan Chase, Mitsubishi led a $38 billion financing for two more Oracle projects, while Sumitomo helped lead the $18 billion financing for Oracle’s New Mexico project.

To be sure, market participants said no major deals have been pulled, meaning no active syndication processes have halted due to weak pricing. And project developers are still stomaching higher costs in the high-yield bond market to keep projects moving.

“At this point, it is still a story of concessions,” Minnaar said. “It’s a race to sort of get as much capacity online as possible and less about the cost of financing for a lot of these companies.”

But recent snags with Oracle’s New Mexico project, financed last year with an $18 billion loan sold into the bank syndicated loan market, was a wake-up call for lenders that has prompted them to take a closer look at contracts and covenants tied to their loans.

Last week Oracle sent a force majeure notice to the developer, a company owned by Blue Owl Capital, after the project ran into power delays. Oracle is seeking to invoke a contract clause that can waive or delay its obligations if events outside its control disrupt the project.

People familiar with the financing said the loan was “well structured” from the lenders’ perspective and wouldn’t let Oracle off the hook on paying the lease, even if the site doesn’t get power. But Oracle’s force majeure call could prompt some banks to seek tougher protections or reassess lending risks, one banker said, further cutting into potential lending supply.