Wall Street banks are increasingly moving massive AI infrastructure debt off their balance sheets, turning to the bond market to redistribute risk and reshape how AI data centers are financed.
According to a report on August 5, a banking syndicate led by Morgan Stanley plans to refinance $15 billion in debt tied to a Texas data center backed by Google and leased to Anthropic. The debt package, expected to be split into multiple bonds or partially refinanced through the leveraged loan market, will be pushed to bond investors shortly after the loan is drawn. Insiders note the scale and structure of this deal reflect a deeper transformation in AI infrastructure funding.
For banks, selling this debt not only reduces their overall AI exposure but also frees up balance sheet capacity to support more lending. Meanwhile, the bond market, with its depth and liquidity, has become the preferred channel for raising long-term capital for hyperscale AI projects, offering faster access and lower costs compared to traditional bank loans.
Banking Strain: AI Infrastructure Debt Piles Up, Traditional Markets Under Pressure
Wall Street's caution over holding AI infrastructure debt is not isolated. Over recent months, several large banks have spent considerable time seeking buyers for more than $50 billion in construction debt linked to multiple data center projects leased to Oracle. Some lenders have also used risk transfer transactions to manage their exposure.
The traditional bank infrastructure financing market, once focused on projects like natural gas pipelines and airports, is now overwhelmed by the massive capital demands of the AI building boom. As banks take on new projects, they are accelerating the offloading of existing debt to maintain lending capacity.
According to reports, the data center campus covers roughly 2,000 acres in Hubbard, Texas, and is being developed by Nexus Data Centers. It is still under construction and will house Google's custom TPU chips, though chip financing is separate from the $15 billion debt package.
Notably, Google's backing only kicks in after the data center is completed. During construction, investors must bear risks such as delays and cost overruns. As a result, the bonds are expected to receive speculative-grade ratings.
Self-Built Power: Avoiding Delays but Adding Complexity
To bypass Texas' increasingly strained power grid, the data center campus will include its own natural gas power plant, ensuring independent operation and avoiding external supply delays and cost pressures.
However, combining data center and power asset financing significantly increases deal complexity, as lenders must evaluate and price two distinct types of risk simultaneously. Similar precedents exist: Meta's "Project Walleye" data center also used on-site power, offering higher yields to compensate investors for the added risk.
The influx of data center projects into Texas has raised concerns about tighter local electricity and water supplies, as well as rising utility costs, further complicating financing for such projects.
Nevertheless, the bond market is replacing bank loans as the core financing platform for hyperscale AI infrastructure projects. Its depth enables projects to secure capital more cheaply and quickly, which is especially attractive for multi-billion-dollar AI ventures.
The refinancing of this $15 billion debt package is the latest example of this trend. As AI infrastructure investment continues to expand, the division of labor between banks and bond markets is becoming clearer: banks provide initial construction loans, while the bond market takes on long-term capital needs, jointly supporting this unprecedented wave of AI infrastructure development.