Financing Is the Product: Broadcom Lines Up $50 Billion So OpenAI Can Buy Its Own Chips

The Wall Street Journal reported this week that Broadcom is arranging more than $50 billion in private-credit financing to fund OpenAI’s purchase of the custom AI accelerators Broadcom is building for it. Apollo Global Management and Blackstone are among the lenders in talks, the facility is targeted to close before year-end, and the talks are early enough that the final size could still move. None of the companies involved has commented publicly.

On its face it is a routine-sounding financing headline with a very large number. Look closer and it is one of the stranger transactions the chip industry has produced: the supplier is raising the money its customer will spend on the supplier’s own product — a bespoke processor line, codenamed Jalapeño and Serrano under OpenAI’s “Nexus” program, that exists for exactly one buyer.

What is actually being financed

The backdrop is the October 2025 strategic collaboration between OpenAI and Broadcom: 10 gigawatts of custom AI accelerators to be deployed by the end of 2029, with first silicon rolling out from 2026. For scale, 10 GW of accelerator power is roughly the electricity demand of several million U.S. homes — an entire nation-state of compute, committed on paper more than four years ahead.

The chip program itself has been moving fast. OpenAI unveiled Jalapeño, its first custom inference accelerator co-designed with Broadcom and fabricated on TSMC’s N3P process, in June 2026, and presented detailed architecture and benchmarks at Hot Chips in August. Early numbers showed 1.5x–1.9x more inference work per watt than the best shipping merchant accelerators. A second-generation part, Serrano, is in development. What the program lacked — until now — was the balance sheet to buy the fleet at 10-gigawatt scale. The project reportedly hit an $18 billion financing snag back in May, per The Information. This week’s $50 billion-plus package is the fix.

The third one since June

What makes this more than a one-off is the pattern. Broadcom has now put together financing for both of its biggest AI-lab customers in under five months:

• June 2026: a completed $35 billion partnership with Apollo and Blackstone to expand Anthropic’s compute capacity.
• October 2026: banks began syndicating a $60 billion package for Anthropic — $42 billion of senior debt backed by Broadcom guarantees plus $18 billion of junior debt without them, per the Financial Times.
• Now: more than $50 billion in private credit for OpenAI’s Nexus program.

Add them up and Broadcom-linked customer financing exceeds $145 billion in a single year — against Broadcom’s own trailing-twelve-month revenue of roughly $89 billion. The OpenAI package alone equals about 56% of what Broadcom earns in a year. Meanwhile, in a detail that tells its own story, Broadcom has been reducing its own debt (from $68.9 billion in late 2024 to $59.4 billion) while its customers’ leverage grows.

$90 billion of AI chip debt in one week

The Broadcom-OpenAI facility did not arrive alone. In the same week, per the WSJ and subsequent reporting:

• Oracle is negotiating with Apollo and Goldman Sachs on an off-balance-sheet vehicle that would buy hardware for a 1-gigawatt data center and lease it back to Oracle — compute on the income statement, assets and debt parked elsewhere.
• SpaceX is seeking roughly $40 billion for Nvidia processors, structured as $10 billion in bank loans led by Apollo and $30 billion in investment-grade bonds, with closing expected in 2027.

That is roughly $90 billion of chip-bound debt sought in a single week — and Apollo appears in all three deals. AI-related U.S. leveraged finance has grown from about $20 billion in early 2025 to $88 billion in 2026, a 4.4x increase in under two years. When borrowing moves from public bond markets toward bespoke private credit, that is generally the market saying the amount, the speed, or the structure is something public investors will not comfortably absorb.

The macro backdrop is not helping. The 10-year Treasury yield touched a 24-year high of 5.326% this week, and Reuters reported that the wave of tech borrowing is adding strain to sovereign bond markets. “The AI build-out started on cash,” deVere Group’s Nigel Green put it. “It’s increasingly running on credit, and credit changes the risk profile entirely.”

Why financing is becoming the product

Nvidia pioneered this playbook at even larger scale — its multi-hundred-billion-dollar financing program with Apollo, Blackstone and others exists so customers can buy Nvidia systems. Broadcom is now running the same logic one level down, in custom silicon where the lock-in is deeper: a chip designed for one customer, financed by its own supplier, is a mutual commitment neither side walks away from lightly.

For OpenAI, the strategic arithmetic is straightforward. Inference cost dominates the P&L of any frontier lab, and Jalapeño’s performance-per-watt gains only matter if OpenAI can actually deploy the fleet at gigawatt scale. Off-balance-sheet structures and private credit let a pre-IPO company commit to 10 GW of silicon without torching its own balance sheet — though guarantees can pull risk back onto Broadcom’s books, and the WSJ report does not specify who borrows in this deal or whether Broadcom guarantees any tranche. That is the number experienced watchers will look for when terms surface: a guaranteed Broadcom tranche means the chipmaker is underwriting its customer’s credit; an unguaranteed one means lenders alone carry it.

There is also a competitive read. Oracle shopping for off-balance-sheet compute and SpaceX issuing $30 billion of bonds to buy Nvidia chips are both evidence that hyperscaler-style capex is now the industry-wide norm, not an exception. When three unrelated companies go looking for $90 billion in the same week, the financing market — not the fab — may become the real gating factor on AI capacity growth.

What to watch

1. Whether the facility closes before year-end as targeted, and at what final size — the May $18 billion snag shows these deals can stall.
2. Guarantee structure. Broadcom-backed senior debt (as in the Anthropic package) versus fully lender-borne risk changes who actually holds the exposure.
3. Apollo’s concentration. One private-credit house pricing three of the largest AI financings simultaneously is a dependency worth monitoring if any facility slips.
4. The ripple into token prices. Claude Haiku 5.5 and GPT-6 Luna both landed at $0.10 per million input tokens this week. Cheap inference priced on debt-financed hardware is not evidence that inference is cheap — it is evidence that the bill has been deferred.

The AI buildout’s first phase ran on venture cash and hyperscaler balance sheets. Its second phase is running on structured finance. Broadcom’s $50 billion bet is that in this industry, the company that finances the chip sells the chip — and OpenAI’s bet is that the future’s compute is worth mortgaging.