Google's $200B AI Chip Financing Deal for Anthropic Takes Aim at Nvidia

Google's $200 Billion AI Chip Engine for Anthropic

Google has assembled one of the largest infrastructure financing programs in corporate history: a network of contracts valued at roughly $200 billion that is designed to put more than $150 billion of its AI chips into the hands of Anthropic. The structure brings together Google, Broadcom, the private credit arms of Apollo and Blackstone, Morgan Stanley and a group of crypto miners, according to people familiar with the arrangements.

The chips at the center of the program are tensor processing units, the AI processors Google has co-developed with Broadcom since 2016. Long used almost exclusively inside Google's own data centers, TPUs are now being sold to outside customers in racks that can link thousands of chips into a single system. The sales are a direct challenge to Nvidia, whose processors have dominated the AI chip market.

The elaborate financing was necessary because Anthropic, the AI company buying the capacity, does not yet have a credit rating, making conventional corporate borrowing difficult. So the risk has been split across the parties: Google, already an Anthropic investor, guarantees data center leases; Broadcom commits to buying the chips and helps finance them; and Apollo and Blackstone provide private credit that buys the hardware and leases it back to Anthropic. Morgan Stanley helped adapt the sale-and-leaseback model long used to finance aircraft.

The scale is unprecedented. In June, a special-purpose vehicle called Compute SPV paid $35 billion for 1 gigawatt of AI hardware, roughly 1 million TPUs, funded by Apollo and Blackstone. In April, Google agreed to sell Broadcom 3.5 GW of TPUs for distribution to Anthropic; Broadcom's filings include $128 billion in purchase commitments, $55.2 billion slated for fiscal 2027 and $72.9 billion for fiscal 2028. To house the hardware, Google has backed crypto miners including TeraWulf, Hut 8 and Cipher Digital, helping to arrange construction debt — five projects in Texas and Louisiana totaling 1.4 GW have raised $15 billion, and Google has supported ten projects with 2.4 GW of power for TPUs. The flip side is concentration: nearly all of the roughly $200 billion in contracts ultimately depends on Anthropic's ability to keep paying.

How the TPU Sale-and-Leaseback Model Challenges Nvidia

Why Google Is Taking Its Home-Built Chips to Market

Google has effectively turned its internal TPU program into a merchant chip business in all but name. The reported deals mean TPUs are no longer just a cost center serving Google's own model training and cloud business; they are a revenue product competing directly with Nvidia's data center GPUs. The economics are favorable on paper: Google builds the chips with Broadcom, sells them at scale, guarantees the hardware is placed, and uses outside capital so the inventory never sits on its balance sheet.

That said, the strategy carries an execution risk that Nvidia does not have. Google is now responsible for hardware sold to a third party whose future revenue depends on Anthropic's commercial success. A senior banker involved is quoted saying AI chips are the most valuable product ever made — the financing structure shows how much of that value is actually being built on credit.

An Aviation Model, Refitted for Chips

The core innovation is borrowed from aircraft financing. Instead of airlines buying planes outright, a special-purpose vehicle buys them and leases them back. Here, Compute SPV — capitalized with about $35 billion of Apollo and Blackstone debt — bought roughly 1 million TPUs in June and leases them to Anthropic. Broadcom has agreed to cover any shortfall if Anthropic stops paying, effectively guaranteeing about $30 billion of the $35 billion facility, according to the report.

This matters beyond Anthropic. If the model works, it can be repeated for other AI labs and cloud operators, turning chips into a standardized, asset-backed financing class. It also underscores why Broadcom, despite being a chip supplier, insisted it did not want to be in the financing business — the SPV structure lets Broadcom and Google monetize AI demand without holding tens of billions of dollars of depreciating hardware.

The $44 Billion Question on Google's Balance Sheet

Google's reported exposure from guaranteed leases is up to $44 billion if every data center lease in the program fails — yet the company's balance sheet shows only $815 million for these obligations. The gap is large enough that investors should ask how the guarantees are structured, what collateral backs them, and what triggers could move the contingent liability onto the balance sheet.

The disclosure gap is not just an accounting footnote. Jefferies analysts cited in the report estimate Google-supported data center projects carry leverage of about 7.1 percent, versus 9.3 percent for cloud operators built around Nvidia chips. That 2.2 percentage-point difference is, in their words, a structural cost-of-capital advantage for Google's ecosystem — and a reason Nvidia-based operators may struggle to compete on price even if their hardware remains the industry standard.

One Customer Holds the Whole Chain Together

The most striking feature of the structure is its concentration. Almost every layer — Google's lease guarantees, Broadcom's purchase commitments, Apollo and Blackstone's credit, the crypto miners' data centers — is ultimately backed by one unrated company's ability to pay: Anthropic. If Anthropic's revenue growth stalls or its funding environment deteriorates, the cascade could hit every participant.

Jefferies analyst Jonathan Petersen frames it as a macro risk: an entire world of AI infrastructure has been built under a handful of players, and if their investment appetite fades, all of it slows down. For now, Google's team says its focus is power — finding enough electricity and sites to host the 4.5 GW of hardware it has agreed to sell. The next phase of the story will be about Anthropic's cash flows, not just data center construction.

What Investors and AI Companies Should Watch Next

What to Watch

For investors, customers and counterparties tied to this AI infrastructure chain, several concrete variables will determine whether the model holds together.

• For investors in Google and Broadcom: track Broadcom's fiscal 2027 and 2028 purchase commitments ($55.2 billion and $72.9 billion) and any changes to Compute SPV's $35 billion, 1 GW financing terms in quarterly filings — they are the clearest signal of whether the program is scaling as planned.
• For investors in AI infrastructure: compare the reported 7.1 percent leverage on Google-backed projects with the 9.3 percent on Nvidia-based operators; a widening gap suggests Google's ecosystem can undercut rivals on price even before chip performance is considered.
• For Anthropic's customers and suppliers: watch Anthropic's funding and revenue metrics, since the roughly $200 billion contract network and Google's up-to-$44 billion lease exposure depend on its ability to keep paying.
• For cloud buyers: treat TPU supply as an increasingly credible second source next to Nvidia; Google's financing model means capacity can be built more cheaply, which may translate into lower AI compute prices.
• For crypto miners with power assets: the TeraWulf, Hut 8 and Cipher Digital deals show a template — Google-backed debt for AI data centers — but any new project depends on securing a creditworthy tenant or guarantor like Google, not just on AI market hype.