AI bubble: how Google is financed by Wall Street Anthropics chips

Google has entered into a complex financial arrangement with chip designer Broadcom, investment firms Apollo and Blackstone, and several banks to facilitate the sale of its Tensor Processing Units (TPUs) without these transactions appearing directly on its balance sheet. A significant portion of these TPUs are being sold to Anthropic, a leading artificial intelligence company. The contracts involved in this structure amount to approximately $200 billion in total value, according to the Financial Times. The arrangement involves Google selling its TPUs to Broadcom, which then sells them to a specially created entity called Compute SPV. This entity borrows money from Apollo and Blackstone to finance the purchase of the chips and subsequently leases them to Anthropic. If Anthropic fails to make its payments and Compute SPV cannot meet its loan obligations, Broadcom has guaranteed to cover up to $30 billion of the $35 billion currently circulating in this round of financing. This structure, as revealed by the Financial Times, could serve as a template for future deals involving another 3.5 gigawatts of TPU hardware already agreed upon by Google, Anthropic, and Broadcom. The model resembles the financing strategies used by Boeing and General Electric in the sale of aircraft and engines. In addition to securing funding for TPUs, Google aims to ensure sufficient power supply and data center capacity for its operations. Former cryptocurrency miners have played a role in this effort by providing excess capacity. These partnerships are common, with companies such as CoreWeave transitioning from mining to operating AI data centers. TeraWulf was the first to build a 360-megawatt data center expansion in Upstate New York under a Google guarantee. Google assured the lease payments for Anthropic’s site. Morgan Stanley structured this deal into construction bonds worth $3.2 billion, effectively taking on the financial responsibility if needed. In return, Google received so-called penny warrants allowing it to purchase shares in TeraWulf for symbolic cents. Google expanded this model to other projects, including collaborations with Cipher Digital and Hut 8 in Texas and Louisiana. According to the Financial Times, there are ten such projects totaling 2.4 gigawatts of power capacity secured by Google as a backstop. Google risks up to $44 billion in potential losses if all rental agreements fail, yet its balance sheet only reflects $815 million allocated for this purpose. This intricate web of guarantees fits within the broader context of the current investment race in artificial intelligence. Recently, it was reported that Nvidia is negotiating a $250 billion guarantee for OpenAI. Collectively, major technology companies have already invested over a trillion dollars in expanding their AI capabilities, despite growing doubts about the long-term sustainability of this boom.