Anthropic borrowed roughly $50 billion to build out its computing infrastructure before its revenue came anywhere close to justifying the bet. Most of that debt closed in early 2026, months before the company’s annualized revenue jumped from about $9 billion to more than $47 billion by May. The timing answers a question the AI industry keeps asking nervously: can capital markets fund buildouts this large, or does financing become the wall compute scaling eventually hits.
Epoch AI, the research group behind the newsletter Gradient Updates, laid out the mechanics in a post published Thursday by analyst Campbell Hutcheson. The mechanism is the interesting part. Anthropic did not post collateral or draw down its own cash. Lenders instead advanced money against the company’s long-term lease commitments, its promise to pay for TPU systems and data center capacity over five to ten years. Two of its suppliers, Google and Broadcom, agreed to take on some of the shortfall should Anthropic or its data centre partner Fluidstack ever stop paying. That guarantee turned a young company’s growth story into something institutional lenders were willing to price.
On the compute side, funds managed by Apollo, along with Blackstone and a group of banks, supplied $34.5 billion through a special purpose vehicle named AI XPV Platform. It purchases the TPU racks and leases them to Anthropic for five years. Broadcom backstops $30 billion of that debt, capping its own exposure at $29 billion if Anthropic defaults. The remaining $4.5 billion carries no Broadcom protection and pays 8.5 percent interest, against 5.75 percent for the backstopped senior tranche. That 2.75 percentage point gap is roughly what Broadcom’s guarantee is worth to lenders.
The data center side follows the same logic at smaller scale. Five project companies raised about $15.2 billion to build 1.43 gigawatts of capacity that Fluidstack will lease to Anthropic, with Google standing behind the rent. At the Lake Mariner site, developer TeraWulf is completing the buildings while Google agreed to cover missed rent or step into the lease if Fluidstack cannot pay, in exchange for rights to acquire TeraWulf shares. Lake Mariner has no unbackstopped tranche for comparison, so the discount Google’s guarantee produced cannot be measured precisely.
This is a single case study, and Anthropic is not a typical borrower. Two of the best-capitalized companies in tech have direct commercial reasons to keep its buildout running, and both have long operating histories that make their guarantees credible to lenders. A lab without a supplier willing to backstop its debt, or without Anthropic’s revenue trajectory, would likely face higher rates or a thinner pool of willing lenders. For this financing model to spread across the industry, more vendors need balance sheets large enough, and commercial stakes deep enough, to make the same trade Google and Broadcom made.
The pattern is already becoming an infrastructure product. Broadcom, Apollo, and Blackstone describe the Anthropic deal as the opening transaction in a platform meant to finance more than 20 gigawatts of frontier lab deployments, including OpenAI’s, through 2028. AI Insiders covered the same dynamic from the vendor’s side this week: an argument that Nvidia’s financing arm now behaves less like a captive chip lender and more like a hyperscaler in its own right. Both cases point to the same shift: suppliers using their balance sheets to keep buildouts financed ahead of the labs’ own revenue.
For operators watching the AI infrastructure market, the takeaway is not that capital is unlimited. It is that whoever controls a scarce input, TPU supply, power-ready land, or a lease counterparty labs need, can turn that position into cheaper capital for the borrower. Any lab without a vendor willing to make that trade should expect a materially higher cost of capital than Anthropic paid.
Epoch AI published this analysis on its Substack newsletter, Gradient Updates, on August 13, 2026.
