On October 1, Reuters got hold of Anthropic's IPO prospectus and found a number buried in it that explains a lot about how the AI industry is actually funding itself right now: Broadcom has agreed to lend Anthropic up to $42 billion to finance its own infrastructure buildout. The catch, which Broadcom disclosed in its own filing, is who that money is ultimately paid to — Broadcom. Anthropic borrows from the chipmaker, then spends most of it leasing compute hardware from the same chipmaker.
This isn't a one-off curiosity. It's the latest and arguably starkest entry in a pattern that's been building across the entire AI stack for two years, and if you're betting your product, your runway, or your trading capital on the assumption that AI demand is as real as the revenue numbers suggest, it's worth understanding exactly how the money actually moves.
The Deal, In Plain Terms
The $42 billion facility could cover roughly one-third of Anthropic's $125.2 billion five-year TPU lease commitment — a number also disclosed in the same IPO filing. Broadcom is already installing one gigawatt of Ironwood chips for Anthropic this year. This loan sits on top of an earlier $35 billion tranche Broadcom arranged in June through a financing platform built with Apollo Global Management and Blackstone, where Broadcom insisted at the time it was providing no direct financing — just "modest residual value guarantees." The $42 billion commitment is a materially deeper entanglement than that framing suggested just four months ago.
There's a real risk clause buried in the filing too: Anthropic put cash into a restricted account for Broadcom's benefit back in April, and may have to top it up. If Anthropic misses a payment or trips another covenant, a large chunk of its lease obligations could become due immediately — right as its access to the $42 billion facility that was supposed to cover those obligations gets restricted. Broadcom is, in effect, on both sides of the transaction: the lender and the landlord, with first claim if the tenant stumbles.
This Is Not An Isolated Deal — It's The Pattern
Anthropic isn't unique here. Nvidia has structured similar arrangements with OpenAI (up to $100 billion in direct investment tied to 10 gigawatts of data center capacity, plus reported discussions around a $250 billion guarantee and $350 billion in chip-purchase financing for a single Ohio campus). Oracle has committed roughly $300 billion in purchases tied to OpenAI. Microsoft has committed around $250 billion of its own. OpenAI, in turn, committed $90 billion to AMD and took a direct equity stake in Nvidia. Every major AI lab is now financially entangled with the same handful of hyperscalers and chipmakers that sell it compute.
Michael Burry — the investor who shorted subprime mortgages before 2008 — has been the loudest voice connecting these dots, posting a Bloomberg-sourced diagram tracing roughly $46 billion in direct equity stakes and $879 billion in multi-year purchase commitments moving between Microsoft, Oracle, Amazon, Google, Meta, OpenAI, Anthropic, xAI, CoreWeave, Nvidia, and AMD. His argument is simple: when a hyperscaler funds a lab, and the lab spends that money buying compute from the same hyperscaler, the hyperscaler books it as revenue. Run that loop often enough and top-line growth stops telling you anything about real customer demand.
The Bank for International Settlements backs up part of the concern with harder numbers: it estimates the five largest hyperscalers are carrying roughly $1.65 trillion in off-balance-sheet debt through special-purpose vehicles — more than the $1.35 trillion they report directly. That debt doesn't show up in standard leverage ratios. It shows up in credit default swaps, and Nvidia's five-year CDS spread has roughly doubled over the past two months.
The Counterargument, Fairly Stated
Jensen Huang has called the circular-financing framing "ridiculous," and it's worth taking the pushback seriously rather than dismissing it. Nvidia generated close to $48 billion in free cash flow in a recent quarter — that's real cash, not an accounting artifact. CoreWeave's CEO Michael Intrator has pointed out that Nvidia's roughly $300 million stake in his company is a rounding error against CoreWeave's $25 billion-plus in independently raised capital. Wedbush's Dan Ives frames the current moment as "1996, not 1999" — early-cycle enthusiasm, not bubble-stage mania, given that enterprise AI adoption is still in the single digits as a share of total addressable spend.
Both things can be true at once: the underlying technology can be generating genuine value, and the financing structure funding its buildout can still be fragile. Vendor financing isn't inherently fraudulent — it's a normal feature of capital-intensive industries building ahead of demand (telecom did this in the late '90s, and some of that fiber is still in the ground getting used today). The problem is concentration and circularity, not financing itself: when the same handful of counterparties are lender, customer, supplier, and equity holder to each other simultaneously, a stumble at any single node propagates fast.
Why This Matters If You're Building, Not Just Watching
If you're a founder or engineer whose product sits on top of Anthropic's API, OpenAI's models, or any infrastructure provider two or three hops downstream of this financing web, you have real exposure you probably haven't priced in:
- Pricing and availability risk. Companies carrying this much leveraged, interlinked debt have strong incentives to protect margins if credit conditions tighten — through price increases, rate limits, or deprioritizing less profitable workloads.
- Vendor concentration is now a balance-sheet question, not just a technical one. Multi-vendor fallbacks (which most teams treat as a nice-to-have resilience pattern) are starting to look like basic financial risk management given how intertwined the major labs now are with each other's funding.
- "Unprofitable but valued at $1.8 trillion combined" (Anthropic plus OpenAI) is not a stable long-term equilibrium regardless of how good the underlying models are. If credit markets reprice AI-linked debt even modestly, the squeeze lands on API pricing and feature availability before it lands on anyone's balance sheet headline.
None of this means the models stop working next quarter. It means the companies providing them are leveraged in ways that make their pricing and roadmap decisions less purely product-driven than they look from the outside — and that's worth factoring into any dependency you're building that assumes today's API terms are stable for the next three years.
Top comments (0)