Paul Yeung/Bloomberg
Major global banks are moving deeper into a financing frontier few would have touched a few years ago: loans backed directly by GPUs, the chips powering Asia’s rapidly expanding AI data center buildout, a market PricewaterhouseCoopers estimates could reach $8.2 trillion in spending by 2050.
Citigroup, JPMorgan Chase, Barclays, Deutsche Bank, Banco Santander and Sumitomo Mitsui Banking Corporation are all now evaluating GPU-linked loans, with lenders having already structured roughly $3.8 billion in chip-related financing for AI infrastructure providers GMI Cloud, Zankore and PaleBlueDot AI.
The individual deals illustrate the scale already in motion. Citigroup is acting as sole debt advisor on a $3.1 billion loan for Zankore in Indonesia, co-underwritten alongside Singapore’s United Overseas Bank and four other lenders. JPMorgan Chase placed a $255 million credit facility for PaleBlueDot AI, while GMI Cloud is seeking $300 million to fund chips for a Thailand data center. Zankore chairman Vikram Sinha said the financing is needed to expand the company’s capacity “tenfold to 1 gigawatt,” a target that underscores how fast Asia’s AI infrastructure buildout is scaling.
The lending carries risks that are unusual even by the standards of infrastructure finance. Eric Tan, a partner at Hogan Lovells Cadwalader, said lenders are grappling with “rapid depreciation, technological obsolescence, and rental volatility” in GPU-backed deals, a cluster of risks that doesn’t map cleanly onto traditional asset-backed lending models. Mike Arougheti, head of Ares Management, put the core problem bluntly: “At least nobody can clearly explain to me what the depreciation curve of this technology looks like.” He argued lenders need to “prioritize risk appetite, not willingness to invest,” a distinction that matters when the underlying collateral, cutting-edge AI chips, could lose much of its value within a few product cycles.
Most of these loans are structured to be repaid through the sale of data center computing power, with customer contracts and the chips themselves serving as collateral. The easiest deals to get approved, notably, are the ones where Nvidia itself guarantees to purchase any unsold capacity, a backstop that effectively transfers some of the depreciation and demand risk back onto the chipmaker whose hardware is being financed in the first place.
