FT recently uncovered an extraordinarily complex financial structure.
Google is set to transfer over $150 billion worth of TPUs into Anthropic, backed by approximately $200 billion in contracts. Yet Anthropic itself has not paid $150 billion to purchase the chips.
The chips are sold by Google to Broadcom, then funneled into a specially created SPV. Apollo, Blackstone, and other institutions extend loans to the SPV, which purchases the hardware and leases the compute power back to Anthropic. Meanwhile, Google provides a guarantee for the data center on the other end.
In June of this year, the first transaction was finalized.
A shell entity called Compute SPV secured $35 billion to acquire approximately 1 GW of capacity—roughly one million TPUs. The $35 billion was structured into three tranches of debt, with about $30 billion in senior debt backed by Broadcom’s residual value support.
The logic is straightforward:
If Anthropic fails to pay rent, the SPV sells the TPUs to repay the debt. If the sale proceeds are insufficient, Broadcom covers the shortfall.
This guarantee is critical. Anthropic has no public credit rating; direct lending to it would carry prohibitively high financing costs. By layering Broadcom’s credit on top, the senior debt’s risk profile becomes nearly equivalent to Broadcom’s own.
And this is only the first transaction.
In April, Google agreed to supply Broadcom with an additional 3.5 GW of TPUs—ultimately for Anthropic’s use. Broadcom’s subsequent financial filings suddenly revealed $128.1 billion in purchase commitments, with $55.2 billion due in fiscal year 2027 and $72.9 billion in fiscal year 2028. According to informed sources cited by FT, nearly all of this volume consists of these Google TPUs.
Today, selling AI compute is increasingly resembling how Boeing and GE once sold aircraft:
Manufacturers sell equipment; financial institutions buy the assets; airlines lease them long-term; the assets serve as collateral; and manufacturers guarantee residual value. This way, tens of billions in hardware remain off both Google’s and Anthropic’s balance sheets.
But there’s one crucial difference between TPUs and airplanes:
A five-year-old Boeing can still find another airline buyer. But what is the residual value of one million TPUs designed specifically for a particular software stack five years from now? No one knows.
Thus, this $200 billion machine ultimately hinges on two bets:
Will Claude’s revenue continue to grow? And can AI chips finally establish a truly credible, financeable “residual value” for the first time?
If both hold true, compute power could become a new class of infrastructure asset.
If they don’t, the risks hidden today within these SPVs will eventually trickle back—layer by layer—onto the balance sheets of Google and Broadcom.
