SpaceX Seeks $40 Billion in Debt to Buy Nvidia Chips: AI Is Now Financed Like Infrastructure
The Financial Times reported on October 6 that SpaceX is negotiating $40 billion in debt, led by Apollo, to buy Nvidia chips. This isn't a closed purchase: it's a snapshot of how AI gets financed in 2026.
What Was Reported on October 6
The British daily, citing people familiar with the matter, put Apollo Global Management at the head of the deal. Reuters, Bloomberg, The Straits Times, Benzinga and TechCentral all carried the report between October 6 and 7. None of them describes an official announcement: all of them describe negotiations.
The Structure: $10 Billion in Loans and $30 Billion in Investment-Grade Debt
The package would split into two very different tranches:
- About $10 billion in bank loans.
- About $30 billion in investment-grade debt.
That the larger tranche is investment grade matters: buyers of that debt treat it as low default risk, and if the market accepts those terms to finance chips, appetite for the sector remains high.
Read also
What the Report Doesn't Say
The 2027 closing is an estimate from the sources, not a commitment. Nor does it specify which chips would be bought or their price, so it's best not to fill that gap with estimates. Neither SpaceX nor Nvidia has confirmed anything, and analysis speculating about an intermediary vehicle holding the chips as collateral is one outlet's interpretation, not a fact published by the parties.
Why It's an AI Story
Three years ago, the chips conversation belonged to labs and a handful of hyperscalers. Today the buyer is a company that also launches rockets.
From Private Lab to Listed Company
On February 2, 2026, SpaceX absorbed xAI in an all-stock merger, and June 12 brought its IPO, the largest in history, with a valuation the coverage placed around $1.77 trillion (a different figure from the merger's ~$1.25 trillion; the two shouldn't be mixed). The practical effect: buying compute stopped being a lab's expense and became the capex of a listed company, which taps debt markets at a scale a private lab can't reach.
Colossus 2: Doubling the Chip Count by Year-End
On September 25, Musk said Colossus 2, the Memphis data center, could more than double its Nvidia chip count before the year is out. That's his statement, not an audited figure, so the prudent move is to talk about hundreds of thousands of accelerators. Demand is high enough that Nvidia struggles to control where its servers go: in October, a California executive was accused of smuggling $300 million in Nvidia-chip servers bound for China.
Money That Has to Be Repaid
The interesting part isn't the number but the payment: buying hundreds of thousands of accelerators costs more than the quarterly profit of almost any company.
Investment-Grade Debt and Collateral
Investment-grade debt is what rating agencies consider low default risk, which makes it easier and cheaper to place; the bank tranche is the more conventional piece. Together they turn compute into infrastructure with a maturity date, not a research expense.
Not an Isolated Case
According to the Financial Times on October 2, a record package of roughly $60 billion in debt was being assembled to fund Broadcom chip leases for Anthropic. The same week, Nvidia announced a $150 billion buyback and $180 billion in contracts with Anthropic. It has also been reported that Nvidia was exploring products with insurers to cover lenders on chip-backed loans: if the collateral is the hardware, the lender wants to know what it will be worth in two years, and chips depreciate fast.
The Calendar Question
The underlying doubt is whether returns arrive before maturity: the debt has a payment date and AI revenue is less certain. On October 6, Bank of America strategists warned of a possible tech bubble and advised hedging with derivatives. That's the view of a group of analysts, not a fact.
The Other Route: Build Instead of Buy
Terafab is the Tesla and SpaceX semiconductor plant project in Grimes County, Texas, presented on March 21, 2026, with a stated goal of more than one terawatt of AI compute per year. Its numbers should always come with a date: the first stage has been put at more than $16.8 billion, and the full complex could approach $119 billion.
Intel, TSMC and the Division of Roles
Intel signed on April 7 as a lead manufacturing partner and Musk has mentioned its 14A process; that doesn't make Intel the owner, it only contributes technology. On October 3, Musk confirmed early talks with TSMC, and on October 7 he clarified that his companies will build and run the complex and that TSMC might sublease part of the capacity, while Intel CEO Lip-Bu Tan confirmed to Bloomberg that Intel remains on board. In plain terms: nobody is leaving, but the division of roles is still open.
What This Means for You
If you buy or rent compute, GPU access will stay expensive, scarce and increasingly financialized: a provider's availability now depends on its balance sheet too, as the episode of memory scarcity until 2028 reminds us. If you decide tech in a company, look at the provider's financial health when signing compute contracts. And if you build hardware, chip-backed debt changes useful life and depreciation: the refresh cycle stops being purely a technical decision. Big tech firms are already designing their own inference chips to avoid depending on the same supplier.
What Comes Next
If the closing is confirmed in 2027, the questions will be how much of that $40 billion is actually drawn, whether Terafab ends up with Intel, TSMC or both, and whether the market keeps buying AI debt at the same price.
Conclusion
The story isn't that SpaceX is buying chips: it's that it's borrowing to do so. That says more about the state of the sector than any accelerator count. If you want to follow how AI's foundations move, keep reading the blog: every week we break down chips, data centers and financing with verified data.
