
Google moves billions in Anthropic chip risk off its balance sheet
Quick Answer
Google has partnered with Broadcom and others to finance AI chip sales to Anthropic, offloading risks from its balance sheet.
Quick Take
This $35 billion deal involves leasing TPUs and securing data centers through crypto miners, while Google aims to reduce financial strain amid significant obligations tied to Anthropic's lease payments.
Key Points
- Google's TPU sales to Anthropic involve a $35 billion financing vehicle managed by Morgan Stanley.
- Anthropic, lacking credit, relies on leasing TPUs instead of direct purchases.
- Google's financing lowers borrowing costs for data centers to a median rate of 7.1%.
- The arrangement poses risks with $200 billion in contracts dependent on Anthropic's lease payments.
- Google's cloud revenue is significantly tied to Anthropic, accounting for over 40% of future commitments.
DeepSignal Analysis
What happened
Google has partnered with several financial and tech companies to finance AI chip sales to Anthropic, offloading risks from its balance sheet. This $35 billion deal involves leasing Tensor Processing Units (TPUs) and securing data centers through crypto miners, while Google aims to mitigate financial strain from Anthropic's lease obligations.
Key evidence
- Google's Tensor Processing Units (TPUs) are now sold to external customers and are designed to compete with Nvidia's AI processors.
- Anthropic, lacking a credit rating, cannot purchase the chips directly, leading to a financing structure where Morgan Stanley leases the chips to Anthropic.
- Google could face up to $44 billion in obligations if all lease payments default, but only records $815 million of that liability on its balance sheet.
Why it matters
This financing arrangement reflects a significant shift in AI infrastructure economics, as Google's backing allows for lower borrowing costs compared to competitors relying on Nvidia chips. However, the arrangement also exposes Google to substantial risk, as it has invested in Anthropic while supplying the necessary hardware. The success of this model hinges on Anthropic's ability to meet its lease obligations, which are critical to Google's future revenue from cloud services.
📖 Reader Mode
~3 min readGoogle has teamed up with Broadcom, Apollo, Blackstone, Morgan Stanley, and several crypto mining companies to finance AI chip sales to Anthropic. The Financial Times describes it as one of the largest infrastructure financing programs in history.
The deals center on Google's Tensor Processing Units, which the company has developed with Broadcom since 2016. Originally built for Google's own data centers, TPUs are now sold to outside customers and challenge Nvidia's dominance of the AI processor market. Google sells the chips in "Pods," server racks that connect thousands of TPUs into one computing system.
Nobody wants the chips on their own books
Anthropic needs vast amounts of AI hardware but can't buy the chips itself because it has no credit rating. Banks won't lend the startup that much money, while the other companies involved also want to keep the hardware off their balance sheets, FT sources say. Google is already spending record sums and doesn't want to add more strain to its balance sheet. Broadcom also doesn't want to tie up its capital in the Google chips it resells.
Morgan Stanley has helped set up a financing vehicle that buys the chips and leases them to Anthropic. Outside investors, mainly Apollo and Blackstone, put up the money. The structure first goes into use in June, when a special purpose vehicle called Compute SPV buys about one gigawatt of TPU hardware for $35 billion. That works out to roughly one million TPUs, according to the FT. Broadcom provides a backstop covering about $30 billion of the purchase if Anthropic stops making lease payments.
The structure now serves as a template for more deals. The largest deal to date is an April agreement covering Google's sale of another 3.5 gigawatts of TPU hardware to Broadcom for Anthropic. Broadcom's financial filings list $128 billion in purchase commitments through 2028, and FT sources say nearly all of that amount is tied to Google TPUs.
Crypto miners provide the power and data centers Google needs
Financing the chips solves only half the problem because Google also needs data centers with enough power to run them. The company is turning to crypto miners that have already secured access to large amounts of electricity.
TeraWulf is the first to receive a Google guarantee, which covers a 360 megawatt data center in New York. Morgan Stanley packages the guarantee into a $3.2 billion construction bond, while Google receives an ownership stake in TeraWulf in return.
The same model now extends to other crypto miners, including Cipher Digital and Hut 8. Google has backed ten projects with a combined capacity of 2.4 gigawatts so far, according to FT sources.
Google could face up to $44 billion in obligations if every lease defaults, the FT reports. Yet the company records only $815 million of that liability on its balance sheet, leaving most of the exposure off its books.
Google's financing edge lowers costs but raises its exposure
Google's financial backing is already changing the economics of AI infrastructure, according to the FT. Data center projects backed by Google borrow at a median interest rate of 7.1 percent, compared with 9.3 percent for neocloud operators that rely on Nvidia chips. Jefferies analysts call the gap "a structural cost-of-capital disadvantage" for companies in Nvidia's ecosystem.
The arrangement carries substantial risk because $200 billion in contracts depend on Anthropic continuing to make its lease payments. Google also sits on both sides of the deal as an investor in Anthropic and the supplier of its chips.
An earlier report from The Information says Anthropic has committed to spending about $200 billion on Google Cloud over five years in exchange for five gigawatts of server capacity. That deal accounts for more than 40 percent of Google's committed future cloud revenue. Together with OpenAI, Anthropic accounts for roughly half of the $2 trillion in cloud backlogs at Amazon, Microsoft, Google, and Oracle. Both startups are counting on their revenue growing 20 to 30 times by 2029, and the entire structure could unravel if that growth slows or stalls.
— Originally published at the-decoder.com
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from The Decoder
See more →
An AI model programmed nonstop for 19 days on a single MirrorCode task that cost $2,600 to run
Epoch AI's MirrorCode benchmark reveals Claude Opus 4.7 as the leader with a 56% solve rate, reconstructing a 16,000-line toolkit in 14 hours. Despite this, all models tested struggle with the most complex tasks, highlighting limitations in current AI capabilities. The single task consumed $2,600 over 19 days, raising questions about cost-effectiveness in AI development.

