Nvidia's $21B SpaceX Stake Lays Bare AI's Circular Financing

Two Nvidia filings landed within days of each other, and read together they describe less a pair of transactions than a financing model: the chipmaker now takes equity and credit positions in the same companies that buy its accelerators. One disclosed a holding of nearly 123 million SpaceX shares worth close to $21 billion at the end of June; the other committed $1.5 billion to SB Energy, the data center and power developer building for OpenAI.

The SpaceX position

That stake originated in Nvidia's January investment in xAI, completed just before Elon Musk merged the AI lab into his rocket company, and a post-IPO slide in SpaceX shares has since marked it down to roughly $17 billion.

The commercial return matters more than the mark. Musk used SpaceX's first public earnings call to commit the company exclusively to Nvidia, praising the Vera Rubin architecture, while projecting a jump in computing capacity from 2 gigawatts at the end of this year to nearer 10 gigawatts than 5 by the end of 2027, expansion that routes almost entirely to a supplier Nvidia part-owns.

The Ohio project

The SB Energy deal is structured more aggressively still: alongside the equity, Nvidia extends up to $105 billion in credit toward OpenAI's Ports-Pike facility near Cincinnati and locks in sole-supplier status for compute at a site engineered to start at 4.25 gigawatts and scale to 8.

The cap table underscores how closed the loop has become, since SB Energy already counts both OpenAI, the eventual tenant, and SoftBank, which sold $5.8 billion of Nvidia stock last November to fund other AI bets, among its shareholders.

Power is the binding constraint

Serving that compute requires a 9.2 gigawatt natural gas plant on Department of Energy land that once enriched uranium for the American nuclear arsenal and for Navy submarines, at an expected cost of $33 billion, a figure swollen by the 66% rise in gas plant construction costs over two years that BloombergNEF has tracked.

The knock-on effect is what utilities are watching, because plants like this will eventually compete for gas with export markets in a squeeze that could triple prices in some regions. The cost of AI capacity, in other words, does not stay inside the data center fence.

The near-term tell will be whether the Ports-Pike timeline holds. Gas turbine lead times, interconnection queues, and local permitting have slipped repeatedly across comparable projects, and a site obliged to build its own generation has considerably more ways to run late than one plugging into an existing grid.

A pattern, not two transactions

These deals sit inside more than $100 billion Nvidia has committed to AI companies over two years, with CoreWeave, Thinking Machines, Safe Superintelligence, and Cursor among the recipients, and alongside a newly outlined consortium of over $500 billion assembled with Apollo Global, Blackstone, BlackRock, Brookfield Asset Management, Goldman Sachs, and KKR, through which Nvidia will partially guarantee customer loans collateralized by the value of its own chips.

The risk nobody has priced

Vendor financing is not a new instrument. Telecom equipment makers extended it liberally to carriers in the late 1990s, and it worked well right up until the demand it presumed failed to arrive. What distinguishes this version is its scale relative to the rest of the market, and the fact that the collateral is the vendor's own product.

The exposure compounds rather than diversifies. A supplier that books revenue from a customer, holds equity in that customer, and guarantees the loans that customer used to buy its hardware has stacked three claims on a single assumption, which is that AI compute demand keeps rising.

If it does, the circularity reads as foresight, and Nvidia captures value at every layer of a buildout it helped underwrite. If demand merely plateaus, the same structure transmits the shock three ways at once: slower orders, written-down stakes, and guarantees backed by hardware whose resale value would be sliding at exactly the wrong moment.

For everyone else in the market, the consequence is competitive rather than financial. Capital that arrives bundled with a supply guarantee is cheaper than capital alone, so rival accelerator vendors increasingly find themselves bidding against a balance sheet rather than against a product. That dynamic tends to entrench an incumbent well past the point where its technical lead alone would have done so.

A disclosure question is worth watching too. Equity marks move with quarterly fair-value accounting while loan guarantees typically sit off the balance sheet until they are called, which means reported financials can lag the underlying risk by some margin. Investors may get their clearest read not from Nvidia's revenue line but from how candidly it quantifies these commitments in future filings. Nvidia did not respond to a request for comment.