SpaceX's $40B Bet: Chips for Satellites That Don't Exist

A SpaceX Starship launches from Starbase, Texas, in June 2026. The Nvidia chips SpaceX wants to finance would partly power Starmind, an orbital AI constellation that has not launched a single satellite. (Photo: N' Moshe, via Wikimedia Commons, CC BY-SA 4.0)

SpaceX is a rocket company whose most consequential market act this week was not a launch. It was a loan. The Financial Times reported on October 6 that SpaceX is in talks to raise $40 billion in debt — $10 billion in bank loans plus $30 billion in investment-grade bonds, led by Apollo Global Management — to buy Nvidia AI chips. The money would be roughly equal to SpaceX's entire projected revenue for this year. And some of the chips are earmarked for satellites that do not exist yet.

Read that again slowly. A company that lost $541 million in its most recent quarter, already sits on about $39 billion in long-term debt, and burned $18.37 billion in a single quarter wants to borrow $40 billion more — to buy processors that lose roughly 30% of their value every year — for an orbital AI constellation that has launched precisely zero spacecraft. Wall Street's bulls call it "a smart strategic move." The skeptics have a better line: a loan moved the stock, not a launch.

We spent this week reading the FT's report through every relay and rebuttal published since — Reuters, Bloomberg, the bulls, the bears, and the one deeply skeptical piece almost nobody else picked up. Here is the deal, the math, and the strange loop nobody on the bull side wants to talk about.

The deal: $40 billion, Apollo, and a buyer who can't shop elsewhere

The core facts come from the FT's October 6 report, picked up by Reuters the same day. SpaceX plans to raise about $40 billion: roughly $10 billion in bank loans and $30 billion in investment-grade debt. Apollo Global Management is expected to lead the transaction and place the debt with institutional investors. Bond giant Pimco is among a small group of lenders in talks to provide financing. The transaction is expected to close in 2027.

SpaceX, Apollo, and Nvidia did not respond to Reuters' requests for comment. Pimco declined to comment. The stock market, as usual, commented for them: SpaceX shares fell about 1% in extended trading after the FT report broke; Nvidia's rose about half a percent. On Wednesday, SpaceX slid nearly 3% to $167.60, snapping three straight winning sessions.

The one sentence that matters most barely made it into most coverage. Bloomberg reported separately that the talks are in an early stage and could end without a deal being completed. The Reuters relay does not carry that caveat. We checked both versions against each other, and it is worth stating plainly: nothing is signed. This is a $40 billion discussion, not a $40 billion deal.

The strategic context is Musk's full-throated bet on Nvidia. "SpaceX has committed to using Nvidia GPUs exclusively because they are the best," Musk wrote on X on August 4. On the company's earnings call the same day, he went further: "Going forward, we have decided to build exclusively on Nvidia, because we think the Vera Rubin architecture is the best architecture. We're exclusive to Nvidia." The company will use Nvidia hardware across its data centers and space infrastructure. In September, Musk said the Colossus 2 data center then held 110,000 Nvidia GB200 chips and 440,000 GB300 chips — with another 220,000 GB300s expected operational within the week, more in November, and possibly more in December "if we get lucky," per secondary reporting on the post.

The satellites that don't exist yet

Here is where the deal goes from aggressive to genuinely strange. Some of the chips being financed are destined for Starmind, SpaceX's orbital AI program — a planned constellation of up to one million satellites, each carrying Nvidia Rubin GPUs and Vera CPUs, designed to run data-center-class AI workloads from low Earth orbit. SpaceX filed with the FCC in January for authority to operate up to a million of them. The FCC accepted the filing for public comment in February. No authorization has been granted.

Not one Starmind satellite has launched. Prototype testing is not scheduled to begin until early 2027, with the first orbital launch targeted for the fourth quarter of 2027. Nvidia claims its Space-1 Vera Rubin module delivers up to 25 times the AI processing of an H100 GPU — a vendor claim that has not been independently benchmarked in an orbital environment.

Put bluntly: this would be the first investment-grade debt deal in history that asks bond markets to price a bet on satellites that do not yet exist. SpaceX already absorbed xAI in February — the merger that turned a rocket company into Musk's AI infrastructure play — and the AI compute business is real today: about $75 billion in aggregate contracted AI compute revenue, including roughly $1.25 billion a month from Anthropic and $920 million a month from Google. But Starmind is not a business. It is a filing, a factory plan in Bastrop, Texas, and a promise. Bonds maturing in decades are not supposed to be collateralized by promises.

The collateral math nobody on the bull side quotes

Now the math, and it is brutal. The $40 billion is new money — not a refinancing. SpaceX already holds about $39 billion in long-term debt, most of it from a $25 billion bond offering in June and a bridge loan from the xAI merger. A successful raise pushes total debt above $79 billion. Oppenheimer analysts led by Timothy Horan have projected the company could carry more than $400 billion in net debt by 2031 — an estimate relayed through coverage of the firm's research note, not the note itself, but directionally terrifying either way.

AI data-center hardware typically loses around 30% of its replacement value per year. Run that forward: a chip worth $1.00 in 2026 is worth roughly $0.34 by 2029. If the bond tranche stretches to long maturities — SpaceX's June offering ran five to thirty years — then the collateral could be worth a fraction of the outstanding debt long before the notes mature. (An important qualifier: the FT and Reuters did not report maturities for the new deal. The duration argument is conditional on the June precedent, and any honest telling has to say so.)

The credit market has noticed the direction of travel. All five tranches of SpaceX's June $25 billion offering traded below issue price within days; the 2036 bonds priced 1.4 percentage points above Treasuries, meaningfully wider than typical for similarly rated debt. After this week's news, SpaceX credit-default-swap spreads hit their highest level since the contracts began actively trading in June. As Grant Nachman, founder and chief investment officer of Shorecliff Asset Management, put it after the June sale: "The equity market owns the upside, bondholders don't. So you have to get paid for the risk. That's why a company can be worth trillions, yet still need to pay meaningful spread to access debt capital markets."

And the cash flow? In the second quarter, SpaceX reported a $541 million net loss on $7.81 billion in revenue — but capital expenditure was $18.37 billion, of which $15.83 billion went to AI compute infrastructure. For the first half of 2026: $28.5 billion in capex against $3.5 billion in operating cash flow. The company ended the quarter with about $100 billion in cash and securities — which came from the June IPO and bond sale, not from operations. FactSet noted the $40 billion would represent basically all of SpaceX's projected $44.5 billion in revenue this year. AI compute capex went from $0.7 billion to $15.8 billion in four quarters. Whatever else this is, it is not a company buying chips out of cash flow.

'Smart strategic move' vs 'a loan moved the stock'

Wall Street is genuinely split, and both camps made their cases this week.

The bulls: Dan Ives of Yorkville Ives called the reported financing a "smart strategic move" in an October 7 client note, writing: "We view this as a smart strategic move for the SpaceX buildout because it fuels the flywheel between launch, Starlink, and AI." His logic is supply-side — "The constraint on that segment is supply, not demand, so the question has always been how quickly SpaceX can bring compute online and how it pays for it" — and he concluded: "We would rather see this compute financed and deployed than deferred." Ives carries an Outperform rating and a $225 price target. Morgan Stanley's Adam Jonas reiterated Overweight with a $300 target, arguing SpaceX's next chapter is "converting energy into intelligence through physical robots." Goldman Sachs raised its target to $230 with a Buy. And on Mad Money, Jim Cramer — never one to understate — blurted: "SpaceX could end up being Nvidia's largest client at this pace. That's fantastic news for both sides," adding on X that the company could soon generate $3–4 billion a month renting out computing capacity.

The bears: Inference.org's October 7 piece had the week's punchiest line — "A loan moved the stock, not a launch" — and the week's sharpest frame: "SPCX holders take the leverage, Nvidia keeps a buyer that cannot shop elsewhere." Wolfe Research's Chris Caso estimated SpaceX's 10-gigawatt commitment could translate into roughly $350 billion in Nvidia hardware sales, at about $35 billion per gigawatt. Susquehanna's Charles Minervino raised his price target to $173 but kept a Neutral rating, "citing uncertainty over future earnings."

Notice the shape of the disagreement. The bulls are pricing the demand story: compute is scarce, SpaceX has $75 billion in contracted revenue, finance it and deploy it. The bears are pricing the capital-structure story: the collateral melts at 30% a year, the debt compounds, and the one customer who cannot leave is the one whose chips you are buying. Both can be true at once. That is usually the sign of a cycle's late innings.

The strangest loop in corporate finance

And now the loop, which is where this story stops being a financing and starts being a hallucination. Nvidia holds roughly 122.8 million SpaceX Class A shares, worth about $17 billion as of August — a stake born when Nvidia invested $10 billion in xAI's January Series E, which converted to SpaceX equity after the February merger. In August, Nvidia partnered with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR on financing platforms intended to mobilize more than $500 billion for AI infrastructure — meaning Nvidia partially backstops loans that are themselves backed by the value of its own chips.

Trace the circle: Nvidia sells SpaceX chips. Nvidia holds $17 billion of SpaceX equity. Nvidia partially guarantees the financing that lets SpaceX buy more Nvidia chips. The chips serve as collateral for the loans. Nvidia's Jensen Huang has rejected the circular-financing characterization, per industry coverage of his remarks. But the geometry is the geometry. If this financing closes, Apollo will lead a deal in which a chipmaker's chips finance a chip purchase by a customer the chipmaker part-owns, backstopped partly by the chipmaker. Nobody involved is acting irrationally. That is precisely what makes it alarming.

There is precedent for the plumbing, if not the poetry. Apollo previously arranged a $3.5 billion loan in November 2025 and a $3.4 billion loan in February 2026 to vehicles that bought Nvidia chips and leased them to xAI for its Colossus data center in Memphis. This $40 billion discussion is different: new-money procurement financing, at investment grade, for chips — some of which will fly on satellites that have not been built yet.

What we checked

A note on how we read this story, because the coverage this week rewards skepticism. We compared the Reuters relay, Bloomberg's caveat, and the Tech Times' deeply skeptical piece against each other. The FT original sits behind a paywall; everything below it is relay. Two things the relays disagree on: only Bloomberg carries the "early-stage, could end without a deal" line — the most important sentence in this story, and the one missing from most coverage. And the "30-year maturities" framing is conditional, not reported: the June offering had 5–30 year tranches, and the new deal's maturities have not been reported anywhere. Presenting them as confirmed terms would be fabrication; several outlets came close. The Oppenheimer >$400B-by-2031 projection reaches us through two relays, not the firm's note. We attribute it as reported rather than treating it as primary fact. The skepticism in this article is sourced, not decorative.

Our take: the AI boom's strangest trade

Our judgment: this is the AI boom's strangest trade, and it deserves to be watched as a sentiment indicator, not just a financing. Every participant is acting rationally within the frame the boom built. Musk needs compute and has decided, exclusively and publicly, that Nvidia is the only supplier. Nvidia needs a buyer that cannot shop elsewhere and has one. Apollo needs yield and has found a borrower whose equity investors will absorb the pain. Bondholders need to be paid for risk that is, at this point, almost philosophical: what is the 20-year value of a chip that will be obsolete in five?

The bull case is that AI compute demand is so real, so contracted, so hungry that the financing structure is a detail. $1.25 billion a month from Anthropic is not a promise; it is a contract. The bear case is that the structure is the story: $79 billion in debt and climbing, collateral that depreciates faster than almost any asset class ever financed at investment grade, and a flagship use case — Starmind — that is currently a regulatory filing. Morgan Stanley estimates AI infrastructure will need $1.5 trillion in external financing by 2028; JPMorgan strategists count more than $300 billion in AI-related debt raised since last November. Somebody is going to own the senior tranches of that pyramid. This week, Apollo volunteered.

Whether this is genius or the top of the cycle will be decided by one variable: whether compute demand survives the debt it took to build it. If the demand is real, Ives is right and the flywheel spins. If it is not — if the AI buildout is being financed by the same optimism it is meant to justify — then a rocket company's $40 billion chip loan will be the trade this cycle is remembered by. Either way, it is the story of the boom, told in a single financing.

Frequently asked questions

Is the SpaceX $40 billion deal done? No. The FT reported on October 6 that talks are underway, led by Apollo; Bloomberg added that discussions are at an early stage and could end without a deal. Nothing has been signed or announced by SpaceX.

What is Starmind? SpaceX's planned orbital AI program — up to one million satellites carrying Nvidia Rubin GPUs and Vera CPUs for data-center-class AI workloads in low Earth orbit. SpaceX filed with the FCC in January 2026; no authorization has been granted and no satellite has launched. Prototypes are expected in early 2027.

Why would bondholders lend against depreciating chips? Because AI compute is contracted and scarce: SpaceX reports about $75 billion in aggregate AI compute contracts. Lenders are pricing the revenue stream, not the resale value of the chips — though the 30%-a-year depreciation is exactly why the credit market is demanding wider spreads.

What is the 'circular financing' criticism? Nvidia sells SpaceX the chips, holds about $17 billion in SpaceX equity (from its xAI investment), and partially backstops the financing consortium — so the chipmaker is on multiple sides of a deal to buy its own chips. Nvidia's CEO has rejected that characterization.

Sources: Financial Times (Oct 6, 2026) via Reuters; Bloomberg; Tech Times; Motley Fool; Inference.org; TipRanks; StockTwits; Finbold; Market Periodical; the FT's October 6 report as relayed.

Related: SpaceX Snapped Up the Airwaves T-Mobile Was Forced to Sell — and Telecom Stocks Crashed — SpaceX's other expansion play, into wireless spectrum. SpaceX Starship Flight 14: First Orbital Launch Sept. 22 — the rocket side of the company that is now borrowing like an AI lab. SoftBank's $64.6B OpenAI Bet: 13% Stake Now Complete — the other mega-financing of the AI-debt wave.

By Daniel Rao
Daniel covers AI policy, platform security, and the geopolitics of technology for TechPulse USA.