SpaceX Wants to Borrow $40 Billion for Nvidia Chips. The 10-Year Treasury Is Already at a 24-Year High.

Ten billion dollars in bank loans. Thirty billion in investment-grade bonds. Forty billion total, according to a Financial Times report that surfaced this week, and every dollar of it earmarked for one purchase: Nvidia chips.

The buyer is SpaceX, or more specifically SpaceXAI, the AI and data center arm Elon Musk's rocket company absorbed earlier this year when it folded in xAI, the lab behind Grok. SpaceXAI runs Colossus, the GPU cluster in Memphis that Musk has called one of the fastest-built supercomputers anywhere. Colossus 2 is the expansion he wants funded. Musk has said SpaceX will use Nvidia hardware exclusively for its data centers, and that Colossus 2 should more than double its chip count by December, on a path toward roughly a million GPUs.

Apollo Global Management is expected to lead the placement, arranging the debt and lining up investors to buy it. That puts Apollo in familiar territory. Nvidia named Apollo one of six financing partners in August, alongside BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, on platforms meant to channel more than $500 billion toward AI infrastructure. The SpaceX deal, if it closes as reported, would be one of the first real tests of that machinery.

[[IMG1]]

What makes the timing notable is where borrowing costs sit right now. The 10-year Treasury yield pushed above 5.3% this week, its highest level since 2002. The 30-year climbed above 5.7%, a similar multi-decade high. The S&P 500 sat within half a percentage point of a record just days earlier, with the 30-year yield already at a 24-year high at the time; since then equities have slipped from those records while yields kept climbing. Both moves trace back to the Fed's September meeting, where officials raised rates by a quarter point, their first hike since 2023, and minutes released this week showed most officials see another increase as likely before year end. Traders were pricing roughly 17% odds of an October move as of Wednesday.

None of that stops a company from borrowing. It does change what the borrowing costs. A $30 billion investment-grade bond sale priced against a 24-year high in benchmark yields carries a steeper coupon than the same sale would have a year ago, before even accounting for whatever spread investors demand from a private, Musk-controlled borrower with no quarterly earnings reports. SpaceX doesn't file the disclosures a public company would, which is part of why a deal like this routes through private credit shops such as Apollo rather than a conventional public bond sale.

Why It Matters

SpaceX's $40 billion ask doesn't arrive in isolation. Amazon, Microsoft, Alphabet, Meta and Oracle together sold $159 billion in corporate bonds through early June 2026, already up 47% from the $121 billion they raised across all of 2025, and JPMorgan CEO Jamie Dimon says AI-linked spending across the hyperscaler ecosystem is climbing from about $700 billion this year to $1 trillion in 2027. JPMorgan recently raised its forecast for 2026 tech-sector bond issuance to $540 billion, the same year the US Treasury needs to find buyers for a deficit north of $2 trillion. SpaceX's $30 billion bond tranche, if it prices as reported, adds directly to that total, and to the competition for the same pool of credit buyers.

The closer parallel is Anthropic. The AI lab lost $42 billion last year, and Broadcom agreed to lend it almost that exact sum to cover compute costs, a deal that drew scrutiny for tying a chipmaker's balance sheet to the fortunes of one of its own customers. SpaceX's arrangement runs a similar playbook from a different angle. Nvidia helps assemble the financing, through partner platforms like Apollo's, that lets a buyer purchase Nvidia's own chips, which supports Nvidia's revenue and the broader case that AI demand still outstrips supply. Whether that demand converts into cash flow strong enough to service the new debt is the question none of these deals answer upfront.

Credit markets have already started pricing in some of that doubt elsewhere. Oracle's cost of default insurance jumped to a record 227 basis points in September, even as Microsoft's, Amazon's and Alphabet's own spreads barely moved, a sign investors are starting to differentiate between AI borrowers with diversified cash flows and those leaning on a single bet. SpaceX, with no public financials and a founder whose other companies carry their own capital needs, sits closer to the Oracle end of that spectrum than the Microsoft end.

What to Watch Next

The Financial Times report is still just a report. Financing of this size takes months to arrange, the deal isn't expected to close until 2027, and that leaves room for the structure, or the total, to change before anything is final. Watch for confirmation from Apollo or SpaceX directly, and watch Colossus 2's build-out through December, when Musk has said the cluster's Nvidia chip count should roughly double.

Also watch the Fed's October meeting. Odds of a hike there sat near 17% this week, but the September minutes left the door open to one more move before year end. A hike, or even hawkish language without one, would push borrowing costs higher right as SpaceX, and likely other AI borrowers behind it, try to lock in financing at scale. Every basis point added to the 10-year yield between now and when this debt actually prices is a basis point added to SpaceX's borrowing cost.

The Pulse24 Take

The interesting thing about this story isn't the $40 billion figure by itself. Big numbers attach to AI infrastructure every week now, and most of them blur together. What's different here is the direction of the money. Nvidia doesn't just sell chips anymore. Through partners like Apollo, it helps arrange the financing that lets customers buy those chips, then benefits again when demand for AI infrastructure shows up in its own stock price. That's a tighter loop than a typical supplier relationship, and tighter loops carry more risk when the music stops.

It's worth separating two different risks that tend to get lumped together in coverage like this. One is whether AI compute demand is real. The capacity constraints described by companies from TSMC to SpaceXAI suggest it mostly is. The other is whether the credit backing that demand is priced correctly, which is a harder question to answer from outside, especially for a private borrower like SpaceX that doesn't publish the quarterly numbers investors would normally use to judge it.

The honest caveat cuts against easy conclusions in either direction. Rising Treasury yields make this debt more expensive, but they don't automatically make it unaffordable, and companies have absorbed higher rates before without the AI buildout slowing down. The real test isn't this deal on its own. It's whether Apollo, Blackstone and the rest of Nvidia's financing partners can keep placing debt this size at reasonable spreads if yields keep climbing from here, or whether SpaceX's $40 billion turns out to be the one that priced right just before that window started closing.