Apollo is underwriting SpaceX's $40B Nvidia-chip order — and the chip-financing frontier just expanded past the hyperscalers

SpaceX is in talks with banks and investors to raise roughly $40 billion of debt — about $10 billion in bank loans and $30 billion in investment-grade paper — to fund a single order of Nvidia chips, with Apollo Global Management leading the financing and Pimco among a small group of lenders in talks, according to a Financial Times report on Oct 6, 2026, confirmed by Reuters and Bloomberg. Closing is targeted for 2027; the talks are early-stage and could end without a deal. SpaceX shares fell about 1% to $169.79 in postmarket trading; Nvidia ticked up about 0.5%.

The structure matters more than the headline number. The $10 billion in bank loans takes the leveraged portion; the $30 billion in investment-grade paper — a much larger sum than most private-company issuers carry — signals Apollo expects SpaceX to clear the rating thresholds. That expectation rests on the S-1 balance sheet filed in May 2026: consolidated FY2025 revenue of $18.67 billion (+33% YoY) anchored by Starlink at $11.4 billion, with a $4.9 billion net loss driven by Starship R&D ($3.0 billion) and xAI compute ($1.9 billion). For comparison, a $40 billion raise would equal roughly 2.1× FY2025 revenue — a balance-sheet event on the scale of the Broadcom deal to lend Anthropic up to $42 billion disclosed on Oct 1, 2026, but without an AI-lab valuation narrative attached.

The chip order is the input to a very specific build-out. Musk said on Sept 25, 2026 that SpaceX's Colossus 2 supercomputer currently runs 110,000 Nvidia GB200 chips and 440,000 GB300s — a 550,000-GPU base. Three additional 220,000-chip batches are queued: one next week, another in November, and a final 220,000 \"if we get lucky\" in late December. The target is roughly 1.21 million to 1.44 million GPUs in operation by year-end, alongside a 1.2-gigawatt dedicated power plant being built to bring the systems online.

At roughly $40,000–$55,000 per GPU slot in a rack-scale system — GB200 NVL72 racks run about $2–3 million and GB300 NVL72 racks about $3–4 million for 72 GPUs each — a single 220,000-chip tranche works out to roughly $8–12 billion. A full ~1 million-GPU order lands at $40–55 billion, the same order of magnitude as the FT-reported $40 billion raise. In other words, the deal sizes to roughly one year's worth of Nvidia chip purchases for a single AI cluster.

Apollo didn't invent this market on Oct 6. On Aug 10, 2026, Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to set up independent AI compute financing platforms intended to mobilize \"over $500 billion\" of third-party capital over time. Apollo President Jim Zelter called it a combination of Nvidia's technology ecosystem with Apollo's \"flexible, long-term capital base.\" The framing — Nvidia compute as an \"investable asset\" with usage-linked, long-duration revenue — is what makes the underlying chips credible collateral for non-recourse or asset-backed structures.

The SpaceX deal is the framework's first concrete mega-ticket. Two months ago the partners described a platform; today Apollo is using it to underwrite $30 billion of investment-grade paper for a single issuer's chip order. The mechanism is simple and important: by treating the chips themselves — deployed in a contracted compute business with Alphabet as anchor tenant — as the loan's economic asset, Apollo can lend against the cash flows those chips generate rather than against SpaceX's launch franchise alone. That is also why the credit story still works for an issuer reporting a $4.9 billion net loss in FY2025: the loan is collateralized by an Nvidia compute fleet with multi-year contracted offtake, not by SpaceX's GAAP earnings.

Every dollar SpaceX borrows flows to Nvidia, then to TSMC for wafer fabrication and to SK Hynix, Samsung, and Micron for the HBM stacks bolted onto every Blackwell GPU. That pipeline is already running hot. TSMC reported August 2026 revenue of NT$514.81 billion — up 53.3% YoY and the strongest monthly print on record — bringing eight-month 2026 revenue to NT$3.39 trillion, +39.3% YoY, per the Sept 10, 2026 release. September's number drops Oct 8; Q3 results land Oct 15. Consensus is calling for $45.8 billion of revenue and a 66.5% gross margin, per Hudson Labs.

Memory is the more under-appreciated chokepoint. Every GB200 and GB300 ships with multiple HBM3E stacks; Blackwell Ultra (GB300) bumps that 1.5×, per Nvidia's GB300 NVL72 product page. SK Hynix holds mid-50% of Nvidia's HBM4 allocation (UBS puts it at ~70% for the Rubin platform), Samsung mid-20%, and Micron ~20%, per industry tracking. Micron reported record FY2026 revenue of $133.19 billion — 3.6× FY2025's $37.38 billion — with Q4 FY2026 revenue of $54.23 billion and Q1 FY2027 guidance of $61.5 billion (±$1.5 billion), per the Sept 30, 2026 release. A new $40 billion chip order doesn't just feed Nvidia's top line — it locks in another quarter of HBM pull-through for the Korean and US memory makers already sold out through 2026.

Until this week, the largest debt-financed Nvidia purchases had three shapes. Hyperscalers (Microsoft, Alphabet, Meta, Amazon) self-funded or used captive vehicles. AI labs — Anthropic, xAI, OpenAI — were lent to by chip makers themselves: Oracle's $40 billion Nvidia buy for an OpenAI US data center in May 2025, Broadcom's $42 billion Anthropic loan on Oct 1, 2026. The third bucket — operator-backed compute businesses — barely existed as a debt-financed category.

SpaceX is the first big test of bucket three. The borrower is not an AI lab and not a hyperscaler; it's a launch-and-broadband operator whose Starlink cash flow is being re-pledged against compute. Morgan Stanley and JPMorgan project big tech will need to borrow roughly $1.5 trillion by 2028 to keep building AI infrastructure, with $800 billion of that coming from private credit, per Morgan Stanley research. That $1.5 trillion is the macro context: this single deal is 2.7% of the entire projected 2028 debt-financed AI capex bill, and it's going to a non-hyperscaler. Apollo, by underwriting it, has now demonstrated that the $500 billion August framework works for buyers outside the obvious set.

Three things move at once. Nvidia gets another non-hyperscaler top-tier customer, paid for with debt whose collateral story is the chips themselves rather than the borrower's GAAP earnings. Apollo proves the August $500 billion framework can underwrite a single-name mega-ticket in 60 days, validating the firm's pivot into AI infrastructure credit. And TSMC, SK Hynix, and Micron absorb another quarter or two of pull-through at a moment when every line is sold out.

The harder-to-see shift is structural. The AI capex stack used to bifurcate cleanly into self-funded hyperscalers and vendor-financed AI labs. With SpaceX in the middle, Apollo has carved out a third lane: private credit for non-AI-lab compute operators with hard assets and contracted offtake. If the deal closes in 2027, expect that lane to fill fast — first with Tesla (Musk-controlled, AI-adjacent, collateral-rich), then with sovereign AI projects, then with any operator whose CFO can show Apollo a multi-year compute contract and a chip vendor willing to take an order. The chip-financing frontier has officially moved past the hyperscalers; the question now is which unusual buyer is next in line, and how much of Morgan Stanley's projected $1.5 trillion 2028 financing gap Apollo plans to fill.