Anthropic filed for an IPO, and buried inside the prospectus is a financing arrangement more interesting than the valuation chatter around it. According to Reuters' report on the filing, Broadcom has agreed to lend Anthropic up to $42 billion to support the AI lab's infrastructure lease agreements. That is roughly one third of Anthropic's $125.2 billion five-year commitment to lease tensor processing unit capacity, disclosed in the same filing.
The number is big enough to be a headline. The structure behind it is why an engineer and a data scientist should both care: it is not a loan in the way the word is normally used, and the same architecture is quietly becoming the standard way frontier labs buy compute.
How the $42 billion is actually built
Read past the shorthand and the deal looks less like a bank wiring cash and more like a co-signer on a very large apartment lease. The layers, assembled from the filing coverage by Pulse 2.0 and Crypto Briefing:
| Layer | Detail | Source of figure |
|---|---|---|
| Anthropic's lease commitment | $125.2B over five years for TPU capacity | IPO filing, via Reuters |
| Broadcom's exposure | Up to $42B, backstopping senior debt tranches | IPO filing, via Reuters |
| Senior tranche estimate | $30B to $31B estimated exposure, some disclosures cap near $29B | Crypto Briefing |
| Junior financing | About $18B package led by Blackstone | Moneycontrol (Bloomberg) |
| Core private credit | Roughly $35B closed June 2026, led by Apollo and Blackstone | Crypto Briefing |
| Capacity covered | Initial financing exceeds 1 GW; platform targets over 20 GW by 2028 | Crypto Briefing |
| Follow-on talks | $60B to $70B in additional senior debt, same SPV structure | Bloomberg via Moneycontrol |
The money does not sit on Anthropic's balance sheet as hardware debt. Special purpose vehicles buy Google-designed TPUs, lease them to Anthropic, and Apollo and Blackstone provide the private credit. Broadcom's contribution is the backstop: it supports the senior tranches, the safest slices that get paid first, and it only pays if leases default or reselling the hardware fails to cover what is owed. No payouts had occurred as of the later 2026 filings.
One more detail from the prospectus: Anthropic deposited cash into a restricted account for Broadcom's benefit in April 2026 under the convertible debt arrangement, and the notes can convert into equity. Anthropic says it does not expect to sell any of the notes before the IPO completes.
Why This Matters: the balance sheet is the product
From a software engineer's view, this is an interface decision as much as a financial one. Anthropic gets compute delivered through an abstraction layer: the lab calls a lease endpoint, the SPV handles ownership, default risk falls through a waterfall where Broadcom absorbs the floor. What the abstraction hides is exactly what the filing warns about. Anthropic's own prospectus flags that Broadcom's dual role as hardware supplier and financing partner creates "potential conflicts of interest" affecting its access to compute, because Broadcom's decisions on hardware availability and pricing could influence how much capacity Anthropic can secure.
From a data scientist's view, the interesting quantity is what this structure implies about expected costs. Labeling the arithmetic below as an illustration with stated assumptions, using only figures from the coverage above:
- Assumption 1: $125.2B buys capacity ramping to the 3.5 GW of TPU access Anthropic secured with Google and Broadcom in April 2026, beginning 2027.
- Assumption 2: straight-line the spend over the five-year lease, so roughly $25B per year.
- Result: at the revenue run rate Anthropic reported crossing $30B by end of 2026 (up from about $9B at end of 2025, per Crypto Briefing), lease payments alone would consume somewhere around 80% of run-rate revenue.
That ratio is the real story the $42B figure hides. The structure exists because no operating cash flow at any plausible 2026 growth rate covers a commitment this size, so the capital has to arrive as vendor-backed debt, private credit, and a contingent guarantee rather than as profit. It is the same economics behind USD.AI's record GPU-backed loan, where $128.9M was underwritten against the resale value of 2,304 specific chips, just scaled four orders of magnitude up and with the chipmaker itself standing behind the paper.
The circularity is the part skeptics will focus on. Broadcom co-designs the TPUs, so backstopping their purchase keeps demand for its own products flowing. Broadcom projects roughly $115B of AI semiconductor revenue in fiscal 2027 and $230B in fiscal 2028, and Anthropic is expected to become its largest compute customer in 2027. The supplier financing the customer who buys the supplier's product is the pattern Nvidia normalized with its customer investments, now formalized into a $42B guarantee. Each party's revenue line depends on the others staying solvent, which is rational while utilization holds and brittle if it doesn't.
What could go wrong, by the terms
The guarantee activates precisely in the scenario the prospectus describes: chips depreciate faster than the leases run, or a newer TPU generation makes the pledged hardware worth less than the remaining obligation. Recovery proceeds from reselling older TPUs then fall short, and Broadcom's contingent liability stops being contingent. Three specifics from the filing worth noting:
- Certain payment or performance defaults could accelerate a substantial portion of Anthropic's $125.2B in lease obligations, making them immediately payable.
- Anthropic may be required to contribute additional cash to the restricted account under certain circumstances, so its downside is not capped at the current deposit.
- The $60B to $70B follow-on talks use the same SPV, lease, and backstop structure, and they were not limited to Anthropic. Other AI labs were in the same conversations, which means this template, not this deal, is the industry signal.
For Apollo and Blackstone, the calculus is friendlier: with a corporate backstop under the senior tranches, they hold AI growth exposure with a cushion beneath them. That is private credit writing the middle of the AI capital stack, a role banks abandoned for this risk profile.
Outlook: read the filing, not the valuation
The IPO is reported at a potential valuation of about $2 trillion, but the prospectus detail that will age better is the disclosure itself. A lab that keeps hardware debt off its books through SPVs is optimizing for one event: pricing day. The structure buys speed and flexibility heading into that event, and it transfers depreciation risk to the vendor with the most information about how fast its own silicon ages. Broadcom took that trade knowingly, because the alternative was a competitor financing the demand.
The signal to watch is the follow-on raise. If the $60B to $70B tranche closes on the same terms with the same backstop, the market has priced TPU depreciation as manageable and this becomes the default financing shape for every frontier lab. If it reprices, or if a second lab's prospectus reveals a haircut on senior exposure, you will know the resale-value assumption underneath all of it is cracking, well before any lab reports slowing revenue.