Table of Contents
Three AI financing deals, one shared lender risk
In sum – what we know:
- One recurring name – Apollo Global Management is reportedly in discussions on all three financings, alongside Goldman Sachs, Blackstone and banks.
- Debt outlives the edge – Oracle stretches server useful life to six years while frontier hardware loses its competitiveness in two to three.
- Custom silicon as collateral – The Broadcom/OpenAI 10-gigawatt program, with Jalapeño still on engineering samples, broadens what lenders are being asked to underwrite.
Three separate AI-hardware financing efforts surfaced at the same time this week, each enormous on its own, and collectively painting a picture of an industry scrambling to fund compute at a scale. Broadcom is reportedly arranging more than $50 billion tied to OpenAI’s custom-accelerator program, according to the Wall Street Journal. SpaceX is reportedly seeking roughly $40 billion to acquire Nvidia hardware, per the Financial Times and Reuters. And Oracle is reportedly in discussions over a separately financed entity that would buy chips and lease them back, also reported by the Wall Street Journal.
None of these are closed deals. They’re talks, at varying stages of maturity and involving different structures. But one name keeps appearing across all three. Apollo Global Management is reportedly among the firms in discussions on the Broadcom/OpenAI financing, reportedly leading the SpaceX effort, and reportedly in talks with Oracle alongside Goldman Sachs.
The common thread isn’t private credit specifically — SpaceX’s reported package, for instance, includes roughly $10 billion in bank loans and $30 billion of investment-grade debt. The thread is that AI buyers want compute capacity now without fully absorbing the balance-sheet and ratings consequences of owning it. And lenders, asset managers, and structured-finance vehicles are being asked to step in and carry the hardware while the operators pay for access.
How the structures separate owning compute from using it
The pattern across these reported deals is remarkably consistent, even though the details differ. A lender group or financing vehicle funds the purchase, takes title to the equipment, and the operating company pays a recurring fee for access. The capex decision gets converted into scheduled payments. The borrower’s leverage profile stays cleaner while it simultaneously funds land, power, buildings, and networking.
For the lender, the appeal is a claim on a defined asset pool, ideally backed by a creditworthy contracted lessee. The protections come from covenants, reserve accounts, step-in rights, and residual-value provisions. In theory, both sides benefit. In practice, the risk hasn’t disappeared. It’s moved. If the operating company still owes lease payments, guarantees residual value, or otherwise supports the vehicle, it carries meaningful financial exposure regardless of where the hardware formally sits.
It’s worth keeping the three structures distinct, because they aren’t the same thing dressed up three ways. Oracle’s reported arrangement is a leaseback — an entity buys chips and leases them back to Oracle. SpaceX’s is reported as a direct debt-heavy package combining bank loans and investment-grade debt to acquire Nvidia hardware. And Broadcom’s role is supplier-side arranging, where the company enabling the hardware is also reportedly helping organize the capital that lets the customer take delivery. Compressing them into a single label misses the structural differences and the different risk profiles each carries.
Debt terms against product cycles
AI hardware is subject to several overlapping clocks, and they don’t tick at the same rate. Physical life is how long the equipment continues to function. Accounting life is the period over which the owner depreciates it. Technological life is how long it remains competitive against the next architecture. And economic life is how long it earns enough utilization and pricing to service debt. These can diverge sharply, and that divergence is where the risk lives.
There’s a common framing that GPUs depreciate in two to three years, but primary filings don’t support that as a blanket statement. Oracle extended the estimated useful life for servers and networking equipment from five years to six years, effective at the beginning of fiscal 2025. Microsoft discloses a two-to-six-year range. CoreWeave uses six years for technology equipment, including GPU servers. The two-to-three-year figure is defensible as a product-cycle or competitiveness concept — frontier hardware does lose its edge that fast — but it shouldn’t be confused with a depreciation schedule.
That gap between accounting life and competitive life is precisely the problem. A lender underwrites against a six-year asset. The hardware’s high-margin period might be considerably shorter. If the collateral is less attractive when a lease ends or a borrower defaults, what’s actually there may not match what underwriting assumed. This concern is sharpest for custom accelerators, where a secondary market beyond the buyer that specified them may be thin to nonexistent. Nvidia GPUs at least have broad ecosystem support and resale channels. A bespoke inference chip designed for one company’s workloads is a different proposition entirely. The debt may outlive the hardware’s moment at the frontier, and the structure only works if the equipment stays rentable or replaceable long enough to bridge that gap.
Where the profit sits, and who holds the downside
The suppliers are booking the boom now. Samsung’s October 8 earnings guidance estimated third-quarter 2026 sales of 195 trillion won and operating profit of 107.4 trillion won — nearly ninefold year over year, off a base of 12.17 trillion won, and a record. TSMC reported third-quarter revenue of approximately NT$1.49 trillion, up about 50% year over year. Marvell, at its October 6 investor day, said it expects approximately $20 billion in fiscal 2028 total revenue and more than $12 billion in fiscal 2029 custom-silicon revenue. The demand is real and it’s showing up in earnings, not just forecasts.
One counterweight to the headline numbers. Micron’s Taoyuan union authorized a strike by a vote of 1,994 to 14, seeking a one-time bonus equal to 83 months’ salary and a permanent profit-sharing mechanism tied to 15% of operating profit. The boom’s distribution is contested even at the companies manufacturing the chips that make it possible.
The transfer of risk is straightforward. Semiconductor companies recognize revenue when equipment ships or contracts close. Utilization, pricing, and residual risk belong to the compute buyer and the capital behind it. That doesn’t make supplier-side profits artificial. But a financing boom can pull demand forward, shifting part of the economic uncertainty from the chip seller to borrowers, special-purpose vehicles, and credit investors. Marvell’s more than $12 billion custom-silicon target for fiscal 2029 shows that financing demand is expanding beyond Nvidia’s GPU ecosystem into specialized accelerators — which is precisely why the Broadcom and OpenAI talks matter beyond their headline figure. If the market is beginning to finance not only GPU fleets but larger custom-accelerator programs, the range of collateral that lenders are being asked to underwrite is broadening in ways that haven’t been tested yet.
Broadcom on both sides of the transaction
OpenAI and Broadcom announced a strategic collaboration in October 2025 to deploy 10 gigawatts of OpenAI-designed AI accelerators, with deployments beginning in the second half of 2026 and completing by the end of 2029. In June 2026, the two companies publicly unveiled Jalapeño, a first-generation “Intelligence Processor”; OpenAI said engineering samples were running workloads in the lab.
The financing layer is separate though. The Wall Street Journal reported that Broadcom is working to arrange more than $50 billion for OpenAI’s custom chip program, reportedly called Nexus internally, with Apollo and Blackstone among the firms in discussions and a closing targeted before year-end. The report identified a successor chip called Serrano, though that name comes from the Journal’s reporting, not from OpenAI or Broadcom’s public materials.
What this creates is a potential supplier-arranger loop. Broadcom is the company enabling the hardware. Broadcom is also reportedly the company helping organize the capital that lets the customer take delivery. That dual role doesn’t automatically mean something is wrong, but it does raise questions about where one interest ends and the other begins. Nothing published establishes that Broadcom is lending to, underwriting, or guaranteeing OpenAI, or that it’s selling to a special-purpose vehicle. The scope of what’s been reported is arranging, not direct financial exposure. Still, a supplier with a commercial interest in the sale also facilitating the financing of that sale is a structure worth watching closely as details emerge.
What has to hold for this to work
The condition that matters isn’t whether AI demand grows. The condition is whether utilization and rental pricing hold long enough to turn fast-moving hardware into debt-serviceable infrastructure. That’s a narrower and harder bet. Demand can be enormous and financing can still fail if the hardware doesn’t earn enough per unit over its financed life.
Take-or-pay terms, high utilization, fast amortization, and creditworthy counterparties can all make these structures work. They aren’t reckless by definition. But the open question is who is actually providing those protections, and how durable they are against a product cycle that moves in two-to-three-year increments while debt can stretch to six years or more. Custom accelerators sharpen the question further. If Jalapeño’s successor offers a substantial performance-per-watt improvement — as each generation tends to — the first-generation hardware’s earning power declines even if it still functions perfectly well.
