Financial institutions are becoming increasingly reluctant to hold AI debt liabilities, including the banks financing a Google-backed Texas data center leased to Anthropic.

According to people familiar with the matter, the project’s lenders, including Morgan Stanley, plan to offload their debt commitments for the 2,000-acre Texas data center campus by selling bonds as soon as the loans are drawn. The aim is to reduce exposure to AI-related risk and free up capital for the lenders.

The move reflects a broader shift in how Wall Street is funding AI infrastructure. Banks remain eager to participate in the sector’s growth, but fewer are willing to keep large AI-related loans on their own balance sheets, especially as projects combine construction risk, power needs and uncertain timelines.

Data center projects require billions of dollars in upfront investment, involve long construction timelines, and depend on sustained demand from AI companies whose long-term revenue prospects remain uncertain. Selling the debt into bond markets allows banks to recycle capital while limiting exposure to a single high-risk sector.

Inside Google’s AI financing structure

The bank-centered infrastructure market is already under pressure from the scale of AI funding needs. Construction, energy, and other infrastructure costs can run into the billions of dollars. In Google’s case, the Texas campus project is supported by an extensive network of deals and a financing program of more than $150 billion spanning hardware manufacturing and data center construction.

The facility itself is financed through the main debt package, while Google’s custom TPU chips are being financed separately.

The structure relies on a syndicate that includes Broadcom, Apollo, Blackstone, and Morgan Stanley to support the core debt package. If Anthropic fails and the chips lose value, Broadcom is expected to cover the gap. Apollo and Blackstone are also providing private credit through a special purpose vehicle that leases the hardware to Anthropic, with Morgan Stanley serving as financial adviser and lender on the deal.

A Google executive described the arrangement by saying, “This is each of us putting our balance sheet to work. We’re doing it on the data center side, [Broadcom’s] doing it on the chip side.”

Even with the various financing arrangements, Google still expects to hold about a 20% equity stake in the Hubbard, Texas, campus, according to sources.

The company also plans to use an on-site natural gas plant to avoid grid delays and lower energy costs. At the same time, the large number of upcoming data center projects in Texas is already raising concerns about grid capacity, water shortages, and higher utility rates, highlighting how AI buildouts can ripple beyond the companies directly involved.

Combining data center assets with power assets also makes financing more complex, since lenders must assess two different risk profiles at once. In the case of Project Walleye, a separate Meta data hub with its own dedicated energy supply, lenders demanded and received higher yields to compensate for the dual-asset risk.

Banks may sell the debt in bond deals

For the Texas campus, people familiar with the transaction said the $15 billion debt will be split into multiple bond sales to reflect a delayed-draw structure, allowing developer Nexus Data Centers to access funds in stages as construction reaches specific milestones.

Some of the financing could also be repackaged through leveraged loans. Analysts say the bond may receive a speculative-grade rating because Google’s backstop is not fully in place until the data center is completed. Investors would also face the risk of construction delays and budget overruns.

Even so, investors are expected to remain interested because the project is backed by companies such as Google and Anthropic. Infrastructure-linked assets have drawn strong demand and are often seen as more attractive than investment-grade debt. Pricing, however, will still reflect construction risk and uncertainty over the project timeline before Google’s guarantees take effect.

The bond market has recently become a preferred source of long-term funding for AI projects because it can be faster and less expensive than bank loans, giving lenders a way to fund big builds without carrying the full exposure for years.

Over the past few months, banks have also aggressively sought to unload $50 billion in Oracle infrastructure loans, turning to risk-transfer markets to reduce their exposure.

Analysts say the result of the Texas financing could shape how future AI infrastructure projects are funded. If investors absorb the bonds without demanding much higher yields, other technology companies may adopt similar structures for multi-billion-dollar data center developments.

If demand is weak or credit spreads widen, borrowing could become more expensive and could change how banks participate in the AI infrastructure boom.