We didn't build this industry so that the decade's most important trust experiment would happen off-balance-sheet, in the dark, priced by two banks and a private credit fund.
The story broke quietly, the way the ones that matter usually do. Bloomberg reported that SpaceX is working with Goldman Sachs and Apollo Global Management on a financing vehicle of roughly $40 billion to fund AI infrastructure — a special purpose entity whose stated purpose, in the reporting's own framing, is to move the liability off the balance sheet and protect the parent's investment-grade rating. The effort has run into friction, per unnamed sources, as credit market volatility complicates negotiations over borrowing costs.
Sit with that for a moment. The headline says SpaceX raises $40 billion for AI. The structure says something else entirely: SpaceX is building a legal container engineered to make a $40 billion obligation invisible to the people who will ultimately own it. And in an industry that spent fifteen years arguing about settlement finality, verifiability, and who bears the loss when the music stops, almost nobody has asked the only question that matters — who is holding the residual when this paper matures, and with what collateral standing behind it.
Let me be honest about the evidence before I build anything on top of it. What we have is a single-source report, unnamed sources, no disclosed debt-to-equity split, no tenor, no coupon, no recourse structure, no named collateral, no statement from SpaceX, no rating agency commentary, and no lender on record. The use of proceeds is described only as large-scale AI infrastructure, which could mean an xAI training cluster, Starlink edge compute, or something more speculative like orbital data centers. That is not enough to conclude anything. It is enough to build a tracking framework, and it is more than enough to identify a market structure forming in real time.
The mechanics matter, so let's be precise about them. A special purpose vehicle is a bankruptcy-remote entity that holds assets and debt separately from its sponsor. When a sponsor is investment-grade and the underlying asset is not — and a GPU cluster leased to a pre-profit AI lab is emphatically not — the SPV is how you borrow against the risky thing without dragging the parent's rating down with it. That is the legitimate use. The illegitimate use, the one that got us into trouble in 2007, is when the isolation is more legal than economic.
Here is where the reported rationale starts to strain. The three major agencies apply look-through treatment to SPVs that carry implicit support, reputational linkage, or common control with the sponsor. If SpaceX guarantees the debt, commits to repurchase, builds the facility, or if the tenant is a Musk-controlled affiliate, the analysts will fold that obligation back into adjusted debt — in full or pro rata. The clean isolation the structure promises exists on the org chart, not in the rating model.
And the sequencing tells us this is not an isolated transaction. Through 2025 we watched the same shape appear repeatedly: the Meta and Blue Owl Hyperion data center vehicle in the high-$20-billion range, xAI's reported GPU purchase-and-leaseback arrangement with NVIDIA participating on the equity side, and a growing pipeline of asset-backed securities collateralized by compute hardware. Each one was larger than the last. If SpaceX's $40 billion prices, it becomes the benchmark that every subsequent AI infrastructure vehicle is measured against — the reference spread for an entire asset class. If it stalls, it becomes the reference point in the other direction.
That is why this belongs on a crypto desk and not only on a credit desk. The instruments being assembled here — securitized cash flows, tranched exposure, collateral with contested residual value — are the same primitives our industry is now marketing under the language of real-world assets and tokenized credit. Same buyers, same underlying risk, different wrapper. We should understand the thing before we tokenize it.
Start with the leverage arithmetic, because nobody in the coverage published it. SpaceX's revenue is widely estimated in the $13–16 billion range for 2024–2025, with Starlink carrying the majority. EBITDA, generously, lands somewhere around $7–9 billion. Add $40 billion of new debt and you are stacking five to six turns of incremental leverage onto a base that an investment-grade profile typically wants to keep between 2.5x and 3.5x net. The SPV is not financial engineering for its own sake — it is the only way this balance sheet survives the addition. That single calculation explains the entire structure, and it explains why the structure exists at all rather than a straightforward bond deal.
Now run the carrying cost, because this is where the reported obstacle becomes legible. Assume seventy percent debt, so $28 billion, priced at 7.5 to 9 percent for what is structurally sub-investment-grade risk wearing an investment-grade sponsor's clothes. That is $2.1 to $2.5 billion of interest a year, before a single dollar of principal. Assume $25 billion of that capital goes to compute hardware with a four-to-six-year economic life — which implies $4 to $6 billion of annual depreciation. The vehicle needs to generate something in the range of $6 to $8.5 billion a year just to stand still, against a sponsor whose entire EBITDA is $7 to $9 billion. When the reporting says negotiations over borrowing costs have become complicated, that is the complication. A 200-to-400 basis point gap between what lenders want and what the sponsor will pay is not market noise. It is the credit market asking whether a single AI tenant's compute revenue can service a charge larger than its parent's operating cash flow, and declining to answer yes for free.
The physics are worse than the finance, and this is the part I keep coming back to. Convert the money into matter. A GB300 NVL72 rack runs roughly $3 to $4 million for 72 GPUs including networking, which is $40,000 to $55,000 per accelerator all-in. AI data center shell, power distribution, and cooling run $8 to $12 billion per gigawatt excluding IT equipment. Split $40 billion at sixty percent silicon and forty percent facilities, and you get somewhere between 450,000 and 550,000 top-tier GPUs drawing 1.3 to 2.0 gigawatts of IT load, with total supply capacity closer to 1.8 to 2.5 gigawatts. Two gigawatts is a large nuclear unit. It is a mid-sized city.
Here is the mismatch that no financing structure can paper over: capital can be deployed in eighteen months, but grid interconnection queues in the United States run three to seven years. Interest accrues from day one. Revenue waits for the substation. And the asset on the other side of the loan depreciates faster than the loan amortizes — GPU economic life is commonly pegged at three to six years with residual value trending toward scrap, while project debt of this type typically runs five to seven years. A power plant finances over thirty years against a thirty-year asset. A toll road finances over thirty years against a thirty-year asset. Compute finances over seven years against a four-year asset. That inversion, not chip supply, is the structural defect at the heart of AI infrastructure lending, and it is precisely why rating agencies treat hardware collateral with visible skepticism. NVIDIA's allocation priority for a customer of this scale is not the bottleneck. The bottleneck is the interconnect agreement, the transformer lead time, and the fact that next-generation silicon will make today's collateral look obsolete before the loan is half-amortized.
Then there is the knot nobody wants to pull. SpaceX and xAI share a controlling shareholder, and SpaceX has reportedly invested around $2 billion into xAI. If the SPV buys GPUs and leases them to xAI, the credit being lent is xAI's credit, which is not investment-grade, and the only way that lease prices acceptably is with SpaceX standing behind it — which collapses the isolation the SPV was built to create. The structure is logically circular. That circularity, not credit market volatility, may be the actual reason the negotiations are stuck, and it is a governance question dressed as a financing question. There is a second motive the coverage skipped entirely: persistent market chatter about a SpaceX IPO in the 2026 window. If that is real, whether $40 billion sits on the balance sheet determines the leverage ratio and the valuation multiple printed on the front of the prospectus. Protecting the rating and protecting the multiple are different objectives that happen to require the same accounting treatment.
This is where my own work keeps pulling me back to verification. In 2024 I ran a pilot with five developers and two sociologists, wiring autonomous agents into decentralized compute through Golem and testing whether oracle networks could suppress hallucination in local news aggregation across the Philippines. We processed ten thousand data points and cut misinformation by roughly forty percent. The lesson I carried out of that project was not that decentralized compute wins. It was that an unverifiable claim about a machine is indistinguishable from a lie, and credit markets run on claims about machines. Can anyone independently confirm that a given GPU exists, is powered, is utilized at the contracted rate, and will still be economically relevant in year five? In DeFi we at least see collateral on-chain in real time. Here, the buyer of a GPU-backed note sees a press release and a trustee report filed quarterly.
That gap is the actual opportunity, and it is being mis-sold. The pitch circulating right now is that tokenized private credit will democratize access to AI infrastructure yields. I have watched this film. It is the same shape as the omnichain narrative that VCs manufactured in 2022, where users were solemnly informed they should care how many chains a contract was deployed on. They never did. Users care whether the thing works and whether the money is safe. Wrapping a four-year depreciating asset leased to a single non-investment-grade tenant inside a smart contract does not reduce the risk by one basis point. It adds a distribution layer, a governance attack surface, and a new class of holder who cannot read a DSCR covenant.
So let me make the contrarian case plainly: the reported obstacle is the best news in this entire story. A $40 billion vehicle that priced without friction would mean the credit market had stopped discriminating. The 200-to-400 basis point gap is the market doing its job — insisting that AI cash flows be underwritten on evidence rather than on faith. The failure mode we should fear is not a delayed closing. It is a closing that happens because everyone assumes someone else did the diligence.
We didn't learn that lesson properly in 2021. I spent a weekend in a Manila dormitory teaching forty classmates to use hardware wallets and read contract source, after watching people around me get liquidated by JPEG mania. I audited the five hottest projects that week and flagged one as a rug pull two days before launch, which saved maybe fifteen thousand dollars of combined student savings. The lesson was not put it on chain. The lesson was that someone has to actually read the thing. Later, in the 2022 winter, two hundred of us in a resilience DAO audited lending protocols through Code4rena contests, landing fifteen findings against Aave and Uniswap for eight thousand dollars in bounties. Transparency there was a process — adversarial, continuous, unglamorous — not a disclosure document.
We didn't get a second chance at 2008 inside crypto. We may be getting one in compute. When the tranches are assembled and the senior notes land with insurers, pension allocators, and CLO managers who have no framework for judging silicon residual values, the risk will have migrated from an industry that understands technology cycles to institutions that do not. That is the exact topology the Bank for International Settlements and the Bank of England have been flagging for two years under the bland heading of AI-related leverage and concentration. And there is a familiar circularity underneath it: chipmaker invests in customer, customer buys chips, chipmaker books revenue, valuation rises, chipmaker invests again. Any equity participation by a silicon vendor in this vehicle would quietly convert a market-clearing price into a related-party transfer, and the benchmark value evaporates.
We didn't need a token to verify a GPU last year. We needed a meter, an auditor, and a counterparty with skin in the game. The same is true now, at forty billion dollars of scale. What we should be building is not another yield wrapper but the verification rail — attestation of physical asset existence, power delivery, utilization, and contractual cash flow, published continuously and checkable by anyone. That is the trust architecture this moment actually requires, and it is the one part of the stack our industry is genuinely good at.
Watch three numbers over the next two quarters, not the headlines. The spread at which the vehicle finally clears, because it becomes the pricing anchor for every AI infrastructure SPV that follows. The identity of the tenant, because a third-party lease is a commercial real estate story and an affiliate lease is a governance story. And whether any agency publishes a look-through opinion, because that single document decides whether the isolation is real or rhetorical. If the deal reprices wider or slips into 2026, the second-order effect lands on highly levered compute landlords, GPU-backed ABS holders, and every direct lending fund that marked its AI book at par.
The $40 billion question is not whether SpaceX can borrow. It is whether we will have built the verification layer for machine collateral before the shadow banking system of artificial intelligence finishes forming — or whether we will spend another cycle teaching people, after the fact, how to read a document nobody handed them.
