Google moves $44 billion in Anthropic chip risk off balance sheet

Google has assembled one of the largest infrastructure financing programs in history by partnering with Broadcom, Apollo, Blackstone, Morgan Stanley, and several crypto mining companies to finance the sale of its Tensor Processing Units (TPUs) to AI startup Anthropic. The core problem is straightforward: Anthropic requires vast amounts of hardware but has no credit rating and cannot borrow the necessary capital from banks. Google itself, already spending record sums, does not want to strain its balance sheet further. Broadcom, which resells Google's TPUs, also wants to avoid tying up its own capital. The solution centers on special financing vehicles. Morgan Stanley structured a deal in which a special purpose vehicle called Compute SPV purchases the chips and leases them to Anthropic. In June, Compute SPV bought about one gigawatt of TPU hardware (roughly one million chips) for $35 billion, with Apollo and Blackstone providing the capital. Broadcom backstopped the transaction by covering about $30 billion of the purchase if Anthropic stops paying. This model has become a template: Broadcom's financial filings show $128 billion in purchase commitments for Google TPUs through 2028, nearly all from similar arrangements.

Google also needed to solve the second half of the problem—securing data centers with sufficient power to run the chips. The company turned to crypto mining operators that had already secured access to large amounts of electricity. TeraWulf received the first Google guarantee, covering a 360 megawatt data center in New York. Morgan Stanley packaged the guarantee into a $3.2 billion construction bond, and Google took an ownership stake in TeraWulf in return. The same model has now extended to other crypto miners, including Cipher Digital and Hut 8. Google has backed ten projects with a combined capacity of 2.4 gigawatts to date. The financial engineering delivers significant benefits: data center projects backed by Google borrow at a median interest rate of 7.1 percent, compared with 9.3 percent for competitors using Nvidia chips—a structural cost advantage that Jefferies analysts say puts Nvidia's ecosystem at a disadvantage.

Yet the arrangement carries substantial risk. Google could face up to $44 billion in obligations if every lease defaults, but the company records only $815 million of that liability on its balance sheet, leaving most exposure hidden. The entire structure depends on $200 billion in contracts contingent on Anthropic's ability to pay rent. An earlier report reveals that Anthropic has committed to spending about $200 billion on Google Cloud over five years for five gigawatts of capacity—a deal that accounts for more than 40 percent of Google's committed future cloud revenue. Anthropic and OpenAI together account for roughly half of the $2 trillion in cloud backlogs across Amazon, Microsoft, Google, and Oracle. Both startups are projecting revenue growth of 20 to 30 times by 2029. If that growth slows or stalls, the entire structure could unravel, leaving Google holding leverage across multiple sides of a single bet.