“Build Data Centers First, Ask Questions Later”: Google Extends Massive Credit Guarantees to Anthropic, Risking Heavy Losses if AI Monetization Falters

Google Backs Anthropic’s $15 Billion Financing With Its Own Credit
AI Debt Issuance Forecast to Reach $570 Billion This Year as Off-Balance-Sheet Financing Surges
Protests Spread Across 42 States as Three-to-Seven-Year Grid Connection Delays Intensify Construction Cost Pressures

Google is preparing to guarantee lease payments and electricity costs to support the construction of Anthropic’s artificial intelligence (AI) data center in Texas. The plan is to leverage Google’s investment-grade credit to attract project financing for long-term contracts that Anthropic cannot sustain through its own cash generation alone. However, if AI service revenue fails to keep pace with rising long-term lease payments and power-purchase costs, the data center investment boom could ultimately translate into refinancing pressure for developers and guarantee-related losses for Google.

Google Extends $43.8 Billion in Guarantees, Expanding the TPU Financing Race

According to The Wall Street Journal and other foreign media outlets on Aug. 3, a Morgan Stanley-led banking consortium is considering lending $15 billion to Nexus Data Centers, which is partnering with Anthropic. The financing is expected to consist of $14 billion in bridge loans and revolving credit facilities. Nexus plans to use the funds to develop an AI data center campus in Hubbard, Texas, alongside a 1.6-gigawatt (GW) on-site natural gas power plant. The project is designed to bypass the grid connection process that has delayed data center construction by supplying electricity directly from its own power plant.

Google has agreed to guarantee billions of dollars in data center lease payments and electricity charges in the event that Anthropic fails to meet its obligations. The guarantees cover four data center leases signed by Anthropic and the associated power-purchase agreements. In return for the financial backing, Google is reportedly set to acquire an approximately 20% equity interest in the data center and power-generation projects. Google has provided similar guarantees for data center projects operated by TeraWulf and Hut 8, with Anthropic serving as the end customer. Google parent Alphabet disclosed that its maximum potential exposure from such credit guarantees had risen to $43.8 billion as of the end of June.

The project forms part of Anthropic’s strategy to move beyond renting servers from cloud providers and instead lease data centers directly. Anthropic has signed more than 10 preliminary lease agreements with developers across the United States and aims to secure at least 10GW of data center capacity over the next several years. The company plans to install Google’s tensor processing units (TPUs), AI chips jointly designed with Broadcom, at the facilities. Anthropic will finance the chip purchases through a separate vendor-financing arrangement with Broadcom.

AI Investment Financing Spreads to Pension Funds and Insurers

A financing model in which highly rated technology and semiconductor companies provide payment guarantees and financial support for large-scale infrastructure investment by loss-making AI companies has recently gained traction across the industry. The expansion reflects the difficulty of funding AI capital expenditure solely through internally generated cash. OpenAI has also reportedly discussed providing as much as $250 billion in financial support alongside Nvidia for a data center project in Ohio. This form of AI infrastructure financing uses Big Tech credit and long-term lease agreements to secure construction funding upfront, with principal and interest subsequently repaid from future service revenue. Developers raise bonds and loans against lease agreements, while investment banks and private-credit funds purchase the resulting debt.

However, as pension funds and insurers allocate capital to related products, the operating performance of the AI industry has begun to affect the soundness of a broad range of financial assets. The Bank for International Settlements (BIS) has identified circular financing arrangements—in which AI companies, Big Tech groups, semiconductor suppliers and data center developers enter into reciprocal investment and purchasing agreements—as a source of risk. When the same chips and equipment serve as both collateral and a revenue base across multiple contracts, the true extent of leverage becomes difficult to assess. The BIS has warned in particular that if returns on AI investment fall short of expectations, funding could contract abruptly and the downturn in capital expenditure could become prolonged.

Table 1. Structure and Potential Risks of AI Infrastructure Financing

Data Center Debt Draws Comparisons With the Subprime Crisis

The European Central Bank (ECB) has also classified the opacity of AI investment financing raised through private credit as a financial stability risk. Private loans, which are predominantly funded with long-term capital, differ structurally from the bank lending financed by short-term deposits that characterized the 2008 financial crisis. Nevertheless, the ECB has warned that losses from falling asset prices could still spread to pension funds and insurers. US technology critic Ed Zitron has likewise warned that the AI data center investment boom is creating a risk structure reminiscent of the subprime mortgage crisis. His concern centers on the fact that vast sums are being raised on the basis of unproven future demand rather than the data centers’ demonstrated growth potential, while the associated burden is being dispersed off balance sheet.

According to an analysis by Nikkei Asia, the off-balance-sheet obligations of Alphabet, Amazon, Meta, Microsoft and Oracle are estimated at $1.65 trillion. Adding the $1.35 trillion in debt officially reported by the five companies brings their total obligations to $3 trillion. Off-balance-sheet obligations stem primarily from leases for data centers still under construction and purchase agreements for graphics processing units (GPUs) that have yet to be delivered. Before the facilities are completed, much of the exposure remains confined to footnotes in financial statements. Once operations begin, however, it crystallizes into expenses that must be paid over extended periods.

Zitron has focused in particular on the use of special-purpose vehicles (SPVs). Data center developers and private-credit funds raise financing through separate legal entities, while Big Tech companies lease the completed facilities over long periods. Although the debt formally remains with the SPV, lease guarantees or loss-protection agreements can ultimately shift the economic burden back to Big Tech companies and investors. Zitron described the arrangement by arguing that “the risk has not disappeared; it has merely been divided into pieces and hidden from view.” He noted that mortgage-related risks also appeared to have been dispersed across multiple financial products during the 2008 financial crisis, when in reality they spread throughout banks, insurers and pension funds.

AI Infrastructure Investment Triggers Community Resistance

These are not the only variables confronting the industry. In areas where data center construction is concentrated, local opposition has increasingly become a political issue. According to Reuters, civic groups staged simultaneous protests against data center construction at 142 locations across 42 US states last month. Communities with differing political leanings cited the same concerns, including higher electricity rates, water shortages, noise and environmental damage.

Data Center Watch, an industry watchdog, estimated that projects worth more than $75 billion were delayed or halted by community opposition and permitting problems in the first quarter alone. The total value of affected projects reached $130 billion. The number of anti-data-center groups across the United States also increased from 396 at the end of last year to 833 in March, while state legislatures introduced approximately 300 bills addressing construction restrictions and cost-sharing arrangements.

Residents are concerned that data centers’ continuous, large-scale electricity consumption could shift the cost of grid upgrades and additional power generation onto households. The US Energy Information Administration (EIA) estimated that data center servers accounted for approximately 7% of commercial-sector electricity consumption in 2025. Given that server power demand remains constant around the clock, its overlap with air-conditioning demand during heat waves could exert sustained pressure on reserve margins and wholesale electricity prices.

Three-to-Seven-Year Connection Queues Put Power Supply in Control of Investment Timelines

The physical limitations of power infrastructure are also constraining data center expansion. According to BMI, a research unit of consultancy Fitch Solutions, data center construction typically takes 18 to 24 months in the United States, but grid connections require three to seven years in some regions. This means some projects could complete their buildings and install all their servers, only to wait years for electricity. Google’s head of energy and sustainability has also said that several utilities quoted grid connection timelines of four to five years, some indicated waits of 10 years, and one utility said it would take 12 years merely to review a connection schedule.

Grid operators’ interconnection queues are already saturated. PJM Interconnection, the largest grid operator in the United States, suspended new applications and overhauled its review process after requests surged. Lead times for extra-high-voltage transformers and gas turbines have also stretched to several years, making it difficult even for power-plant developers to secure the necessary equipment. Big Tech companies are becoming increasingly concerned because the longer the interval between financing completion and the start of electricity supply, the greater the accumulation of construction-period interest and the risk of contractual default.

With grid bottlenecks extending beyond the ability of individual operators to address them, federal regulators have begun overhauling interconnection rules. In June, the US Federal Energy Regulatory Commission (FERC) ordered six regional grid operators, including PJM, to justify or revise their connection rules for large electricity consumers within 60 days. The principal objectives include shortening interconnection reviews, preventing transmission costs from being shifted to other users and establishing connection standards for data centers equipped with their own generation facilities. The measure is intended to revise both connection speeds and cost-allocation principles amid recognition that the existing grid-management system cannot absorb rapidly growing data center demand.

Because the US government regards its AI competition with China as both an economic and national security priority, it is likely to combine faster permitting with stricter cost-allocation requirements. Permitting procedures are therefore expected to be accelerated while a larger share of grid-expansion and on-site generation costs is assigned to project operators. However, if AI service revenue fails to keep pace with the growth of long-term lease payments and power-purchase costs, developers and private-credit investors will be the first to face refinancing pressure. Guarantors such as Google could also be forced to absorb losses by assuming lease rights or making payments under their guarantees.