Four hundred billion dollars. The largest single AI infrastructure financing ever attempted. The number is not the story.
The story is who is in the room. Goldman Sachs is running the syndication. Apollo Global Management is seated across the table. When a company carrying an investment-grade rating has to place the largest alternative credit manager in the world beside its lead bank, the syndicated market has already answered the question. It said no. Or it said not at this price.
The reporting behind this deal is thin — a single wire item, unnamed sources, no term sheet, no tenor, no coupon, no recourse structure, no identified collateral. I do not read that as a funding announcement. I read it as a signal about how the credit market has started to price artificial intelligence itself.
That distinction matters more than the $40 billion.
The structure is a special purpose vehicle. Its stated purpose is to move the liability off SpaceX's balance sheet and protect the parent's investment-grade rating.
Read that backwards and you get the actual disclosure. If SpaceX could carry $40 billion directly, it would. It cannot. Reported revenue sits in the $13–16 billion range, with Starlink doing most of the work. Optimistic EBITDA lands at $7–9 billion. Forty billion of new debt is five to six turns of incremental leverage. Investment-grade net leverage tolerance is roughly 2.5 to 3.5 turns. The SPV is not a tax structure. It is a rating-protection device, and the arithmetic is the reason it exists.
I learned this discipline the hard way. In 2017 I put 50 ETH against twelve ICO whitepapers and rejected eleven — not because the ideas were bad, but because the structures could not survive a stress case. One survived. It returned 40x. The lesson was never about picking winners. It was about refusing to price a claim whose collateral I could not see.
I cannot see this collateral either. So I am going to price the structure instead.
Start with the lender group. Goldman plus Apollo is not a coincidence. It is a dual-track backstop. Public syndication for the cheap money, private credit for the money that actually shows up. Private credit costs 300 to 500 basis points over investment-grade spreads, arrives with tighter covenants, and executes in weeks rather than quarters. You only pay that premium when the cheaper channel is shut.
The line about credit market volatility complicating cost negotiations is the most informative sentence in the reporting. It locates the obstacle in price, not access. My read: a 200 to 400 basis point gap between what the sponsors will pay and what lenders will accept for a first-of-kind AI infrastructure credit.
Now run the carrying cost. Assume 70% debt — $28 billion — priced at 7.5% to 9% for structural, sub-investment-grade economics. That is $2.1 to $2.5 billion of annual interest before a single accelerator is switched on. Assume $25 billion of the total goes to compute hardware depreciating over four to six years. Add $4 to $6 billion of annual depreciation. Total cash carrying cost lands at $6 to $8.5 billion per year.
Against an undisclosed revenue contract. Against a tenant that has not been named.
The use of proceeds is described only as large-scale AI infrastructure. That phrase covers three very different assets: an xAI training cluster, a Starlink edge compute layer, or orbital data centers. The third is the only one that escapes the interconnection queue entirely — solar power, radiative cooling, no grid — and it is also the least commercially validated. The ambiguity is not sloppiness. It is the difference between a related-party lease and a third-party commercial lease, and those two assets price 200 basis points apart.
Which brings us to the tenant problem. SpaceX and xAI share a controlling person. SpaceX reportedly put roughly $2 billion into xAI. If the SPV buys accelerators and leases them to xAI, the underlying credit is xAI's — not investment grade — and the paper only prices with SpaceX's implicit backing. That backing is precisely what destroys the isolation the SPV was built to create. The structure closes on itself. I suspect that loop, not market volatility, is where the deal is actually stuck.
Rating agencies are not naive about this. Any vehicle with implicit support, reputational linkage, or a common controller gets look-through treatment. If SpaceX builds the facility and xAI leases the capacity, the debt returns to the adjusted balance sheet in proportion or in full. The claim that off-balance-sheet placement protects the rating is a journalist's simplification. Technically it does not hold. Structures fail at the seam where the asset's life ends and the debt's life continues.
Then there is the physical layer, which nobody in this debate prices. Forty billion converts to roughly 450,000 to 550,000 GB200/GB300-class GPUs at $45,000 to $55,000 all-in with interconnect, and 1.3 to 2.0 GW of IT load, 1.8 to 2.5 GW of total supply capacity. US interconnection queues run three to seven years. The capital can be deployed in eighteen months. The power cannot arrive in eighteen months. Interest accrues from day one. Revenue waits for the grid.
And the tenor is wrong. GPU economic life is three to six years, and residual value collapses with each silicon refresh. Project debt runs five to seven years. The asset dies before the loan does. A power plant amortizes over thirty years. A toll road over thirty. An accelerator does not amortize at all — it depreciates to scrap on a schedule set by a chip designer. A loan is a claim on a future cash flow. Collateral is a claim on a thing that must still be worth something when the future arrives.
The entire public argument is whether AI demand is real. That is the wrong question, and it is why the smart money keeps mistiming this cycle.
Demand can be entirely real and the structure can still default. Residual value in year four is not a function of demand. It is a function of the next chip. You can have infinite inference demand and still hold collateral worth fifteen cents on the dollar.
Here is the part the Web3 crowd keeps missing. This is the trade on-chain credit spent 2021 through 2023 promising and never delivering. Bring real cash flows onto a ledger, let the market price them. What actually happened is that underwriting migrated back to humans with legal recourse and a trustee. The GPU-backed ABS and the SPV are the same idea executed off-chain. The chain added settlement. It never added price discovery. On-chain credit did not fail because of regulation. It failed because nobody could mark the collateral. Same collateral. Same problem.
And the risk transfer is not a side effect. It is the product. Direct lending funds moved AI and data center exposure from the periphery to the core of their books across 2024 and 2025. A failed print here forces every one of them to re-mark the entire book. That is the contagion channel — not SpaceX. This paper ends up with insurers, pensions, and CLO tranches: holders who cannot mark a three-year-old accelerator. Tokenized wrappers will eventually hand that residual-value bet to retail and call it democratization.
The number to watch is not $40 billion. It is the basis points.
The first benchmark print on this paper sets the floor for every AI infrastructure SPV, every neocloud ABS, every tokenized compute note that follows. Priced at investment grade plus 350, the asset class exists and is underwritable. Priced at plus 600, the asset class exists and the equity is worth less than the deck claims. Unpriced, the repricing begins somewhere else and arrives here last.
The architecture of trust is built, not inherited. Right now the market is deciding what to charge for the mortar.