How do you finance $40bn of Nvidia chips without breaking your rating?

On 6 October 2026, the Financial Times reported that SpaceX is seeking to raise about $40bn to pay for Nvidia chips, split roughly between $10bn of bank loans and $30bn of investment-grade debt, with Apollo Global Management expected to lead the deal and place the paper with investors. Reuters, relaying the FT, added that Pimco is among the lenders in talks and that the transaction is expected to close in 2027. None of the parties commented.

The company arriving at that table is four months old as a public issuer. SpaceX listed on Nasdaq in June 2026 at $135 a share in the largest IPOIPOThe first sale of a private company's shares to public investors on a stock exchange, converting private ownership into publicly traded stock.View full definition → on record, received a BBB issuer credit rating from S&P Global Ratings on 18 June with a stable outlook and an expectation that adjusted leverage stays below 2.0x, then priced a $25bn inaugural bond in late June, mostly to refinance the $20bn bridge loan taken out in March to absorb xAI. That debut was well bid, reportedly drawing around $90bn of orders, and it still left a mark: the 2036 tranche came roughly 1.4 percentage points over Treasuries, about 0.4 points wider than the average for comparable BBB credit. The thesis of this piece follows from that gap. In AI infrastructure finance, capital is not the scarce input. The scarce input is a credible answer to one question: how long does the asset earn, relative to how long the debt lives, and who owns the difference? Every structural choice in the SpaceX deal, from the Apollo mandate to the loan-versus-bond split, is an attempt to answer it.

The $10bn loan and $30bn bond split SpaceX is building

The split is doing specific work. Bank loans carry covenants, amortisation and a relationship manager who can be called when a chip order slips; they are the flexible, repriceable, renegotiable layer. The $30bn of investment-grade bonds does the opposite job: it buys tenor and a buyer base. Because SpaceX sits at BBB, the second-lowest investment-grade band, insurance companies and pension funds can hold the paper within mandate limits that would exclude a high-yield issuer. That single notch distinction is the whole reason a $30bn tranche is conceivable rather than fanciful, and it is why thechoix du rating et du récit crcrThe percentage of visitors or prospects who complete a desired action (purchase, sign-up, contact form), calculated as conversions divided by total opportunities.View full definition →édit is a structuring decision rather than an investor-relations chore.

Apollo's role is the tell. Apollo has already financed the exact asset twice: $3.5bn provided in January 2026 to a Valor Equity Partners fund buying data centre infrastructure leased to xAI, followed by a roughly $3.4bn chip-leasing facility reported by Reuters in February. Apollo was also part of the $500bn AI infrastructure financing partnership assembled around Nvidia in August 2026 alongside Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR. A firm that has underwritten GPU residual value three times has a view on it; a bank syndicate selling bonds does not have to hold one.

One disclosure point matters when reading any chip-demand number in this deal. Nvidia is both the vendor and, per reporting around the IPO, a holder of a large SpaceX equity stake. Figures that originate with chip suppliers about end-demand for chips are commercially interested, and the Bank of England's Financial Policy Committee said as much on 30 September 2026 when it flagged the combination of rising leverage, limited transparency and what it called circular arrangements in AI financing.

Why is Apollo leading instead of a bank syndicate?

Because the risk being sold is residual value, not credit quality, and private credit prices residual value for a living. A bank syndicate underwrites a borrower's ability to pay. The $40bn question here is narrower: what is a 2026-vintage Nvidia rack worth in 2030, and who eats the gap if the answer is "less than the loan balance"?

The market has no settled answer. S&P's Andrew Chang has said service lives beyond five years have held so far while the agency still takes a conservative view of the chips' value. Google, Microsoft and Oracle depreciate AI chips over roughly six years. Meta disclosed that extending certain server lives to 5.5 years would cut its 2025 depreciation by about $2.9bn. Michael Burry argues real economic life is nearer two to three years and estimates the gap overstates industry profits by around $176bn across 2026 to 2028. Nvidia has told investors hyperscalers depreciate over four to six years. The spread between those positions, applied to a $30bn chip order, is tens of billions of dollars of collateral value.

Lenders have found two ways to bridge it. One is to attach the loan to a contract rather than to the hardware: CoreWeave closed an $8.5bn GPU-backed facility in March, supported by a Meta contract worth at least $19bn, rated A3 by Moody's, priced near 5.9% fixed and maturing in 2032, led by Mitsubishi UFJ and Morgan Stanley. The other is to move the asset off balance sheet and guarantee the residual explicitly, which is what Meta did with Blue Owl on the $27bn Hyperion campus in Louisiana: Blue Owl funds took 80%, Meta kept 20% and operational control, Pimco anchored the debt, and Meta signed a four-year initial lease plus a residual value guarantee covering the first 16 years of operations. The debt left Meta's books. The economic exposure did not.

What the bond market has already priced into SpaceX debt

Spread, not access. SpaceX raised $25bn in June with a heavily oversubscribed book, and within weeks the paper sold off. According to MarketAxess data cited in the FT's reporting, its 2056 bonds have traded around 85 cents on the dollar at a yield roughly 2.27 percentage points over Treasuries, closer to junk pricing than to the BBB median. Investors approached about the earlier chip financing described receiving a two-page term sheet. Thin disclosure from Musk-controlled entities has made some credit investors cautious, and that caution now has a visible price.

The scale context is no longer marginal. The Bank of England's July 2026 Financial Stability Report noted that the five AI hyperscalers accounted for about 3% of outstanding US investment-grade debt at the end of 2025, but more than 15% of year-to-date issuance by early May 2026. August 2026 set a US investment-grade monthly supply record of $145.2bn, a third consecutive record month, driven by data centre borrowing. FactSet put incremental annual debt at 32% of hyperscaler capital spendingcapital spendingCapital Expenditure (CapEx) is money spent to acquire, upgrade, or extend long-lived assets like equipment, property, or software that deliver value over multiple years.View full definition → on a trailing basis by mid-2026, up from 9% in fiscal 2024. Moody's warned in July 2026 that heavy capital spending relative to revenue would push free cash flowfree cash flowFree Cash Flow is the cash a company generates from operations after funding the capital expenditures needed to maintain and grow its asset base.View full definition → down, in some cases negative, and damage leverage ratios where the spending is debt financed, citing around $1.2 trillion of 2026 data centre commitments. The FPC, citing JPMorgan, put potential debt-financed AI capexcapexCapital Expenditure (CapEx) is money spent to acquire, upgrade, or extend long-lived assets like equipment, property, or software that deliver value over multiple years.View full definition → at roughly $4.1 trillion between 2026 and 2030, with Morgan Stanley estimating about $700bn of data centre capex funded through private credit from 2026 to 2028.

The life, lock and residual test for financing capex

Three questions, asked in order, before any instrument is chosen. Write them on the whiteboard.

Life: what is the honest economic life of the asset, as opposed to the accounting life in your fixed asset register? If those two numbers differ by more than a year, the smaller one governs the financing.

Lock: how much of the asset's cash flow is contracted, by whom, and for how long? CoreWeave's $8.5bn facility worked because a named counterparty had signed for at least $19bn. A general procurement raise has no such lock.

Residual: who owns the asset at the end, and is that ownership documented? Meta's Hyperion guarantee is the honest version. If nobody has signed for the residual, you own it whether or not it appears on your balance sheet.

The rule: fund an asset with debt no longer than the contract that pays for it, plus only the residual someone else has guaranteed in writing. Everything beyond that line is a refinancing bet, and it should be sized, disclosed internally and priced as such.

A hypothetical to show the arithmetic, using only published reference points. Take $30bn of chips with a four-year economic life: straight-line depreciation is $7.5bn a year. Fund it with ten-year bullet bonds at 6%, within the 4.55% to 6.45% range Meta paid on its May 2026 senior unsecured notes. Interest runs $1.8bn a year, $18bn over the life of the bond, and the full $30bn principal falls due in year ten, when the hardware bought in year zero is several generations old. To retire that principal from the asset itself, the compute has to throw off roughly $7.5bn a year of cash above operating costs and interest during the first four years. If contracted revenue covers half of that, $15bn is a bet on the 2036 refinancing market. That is the number a board should be shown, not the coupon.

Where your context differs from a $2 trillion issuer

Most CFOs will never place $30bn. The transferable parts are smaller and sharper. SpaceX can tap a BBB bond market that absorbed a $25bn debut; a mid-cap issuer's equivalent move is a club facility where therelation bancaire et les covenants négociés determine whether a capex slip becomes a waiver conversation or a default. SpaceX reported about $28bn of non-cancellable AI-related commitments as of June 2026, much of it falling due in 2027, against roughly $15.8bn of compute capex in the second quarter alone. Your version of that number sits in the purchase obligations footnote, and almost nobody reads it until a lender does.

What to do on monday

Build an asset-life ladder for every capex line above your materiality threshold: four columns showing economic life, accounting life, contracted cash coverage and funding tenor. Any line where funding tenor exceeds contracted coverage by more than 24 months goes to the audit or finance committee with a number attached.

Pull your non-cancellable purchase obligations due within 18 months and put the total in the next board pack, alongside the cash and facilities available to meet them.

Call your rating analyst before they call you, and open with depreciation assumptions. Agencies are already divided on chip lives; being the issuer who raised it first is cheaper than being the issuer who was asked.

Ask your DCM desk for the spread your last issue paid versus the peer median at the same rating. If there is a premium, find out how much of it is disclosure quality rather than credit, because that part is fixable inside a quarter.

MapMapUsing software to automate repetitive marketing tasks and campaigns, enabling personalisation at scale across channels like email, web, and social.View full definition → the circularity in your own capital stack: any supplier that is also an equity holder, guarantor or lender to you. Name them in one page.

SpaceX's $40bn is not yet signed, and the parties have said nothing. What is already settled is the price of the questions it raises: a 2.27 point spread on the 2056s, against a BBB rating earned four months ago. Tenor discipline and disclosure quality show up in basis points long before they show up in a downgrade.