Banks are beginning to move into GPU financing in Asia, a space previously dominated by private credit funds with a higher appetite for risk, significantly widening the pool of capital for the next stage of the AI race.

In recent months, major international banks have played a key role in roughly $3.8 billion of GPU loans for three AI infrastructure providers: GMI Cloud, Zankore and PaleBlueDot AI. According to people familiar with the matter, Citigroup, JPMorgan Chase, Barclays, Deutsche Bank, Santander and Japan's Sumitomo Mitsui Banking Corporation are all currently evaluating GPU-linked loans.

This funding is crucial. PwC estimates that Asia's data center spending could reach $8.2 trillion by 2050, with the vast majority going toward hardware such as GPUs and servers.

International Banks Lead, Asian Banks Follow

Hundreds of data centers are being built across Asia, and their developers are simultaneously raising money to buy chips. GPU financing has already grown rapidly in the United States, while the relatively few such loans in Asia to date have mostly been completed by private credit.

Citigroup served as sole debt advisor on Zankore's $3.1 billion borrowing in Indonesia, while JPMorgan Chase acted as placement agent on PaleBlueDot AI's $255 million credit facility. About six bankers and financial advisors in Asia said they are in talks about, or aware of, more GPU-related financing. Separately, people familiar with the matter said last month that GMI Cloud is in talks with banks and private lenders for a new $300 million loan to buy chips for a data center in Thailand.

Among traditional lenders that have already entered this space, global investment banks are currently the main players, drawing on their experience in handling complex structures, according to people familiar with the matter. One banker said some U.S. banks have brought in headquarters specialists to assess chip values. Citigroup, Barclays and Deutsche Bank declined to comment, while JPMorgan Chase, Santander and Sumitomo Mitsui did not immediately respond.

Local Asian banks are also becoming more active. According to Zankore Chairman Vikram Sinha, Singapore's United Overseas Bank, which co-underwrote Zankore's $3.1 billion loan alongside four other banks, is leading negotiations for its new round of financing. Speaking at a conference on September 22, Sinha said Zankore aims to expand its AI data center capacity tenfold to 1 gigawatt, which will require continuous financing, and working with banks is the "right way." "We are very clear about the scale we want. We want to take the hard road and work with banks and syndicates."

Chip Valuation Remains a Challenge

Eric Tan, a partner in Hogan Lovells Cadwalader's banking and finance practice, said: "As deals get bigger and borrowers demand more competitive pricing, banks will play an increasingly important role." But with technology iterating so quickly, GPU financing also exposes lenders to risks of "rapid depreciation, technological obsolescence and rental volatility."

As banks become a more important source of funding, borrowers will need to more fully demonstrate that project revenue is sufficient to repay loans, and traditional lenders may demand more conservative underwriting standards and higher debt-service reserves.

Private credit has not lowered its bar either. Mike Arougheti, head of Ares Management, one of Asia's largest private credit institutions, said that although GPU financing is the biggest funding gap in the AI boom, Ares still sticks to high standards. "You have to lead with risk appetite, not with the willingness to deploy capital," he said at the Barclays Global Financial Services Conference in September. "At least no one has been able to explain to me clearly what the depreciation curve for this technology looks like." He also said returns on such loans are limited, usually only about 100 to 200 basis points higher than other AI infrastructure loans.

Repayment Hinges on Customers

Most GPU loans in Asia follow the structure pioneered by CoreWeave, one of the earliest and largest users of this type of financing. In many deals, loans are repaid from revenue generated by selling data center compute, with customer contracts and the chips themselves usually serving as collateral. For lenders, the most critical question is therefore whether the customer is reliable and how long the contract term is.

Deals backed by Nvidia are the easiest to get approved. In the GMI Cloud Taiwan project loan completed last month, provided by 14 banks, and in the Zankore loan, Nvidia agreed to buy all unsold compute, providing a backstop if customer contracts fall through. In exchange, the two members of Nvidia's cloud partner program will charge buyers prices above the rates Nvidia committed to and share revenue with Nvidia.

The debate over AI's potential dangers and whether guardrails are needed for its development will also continue to shape risk assessments. Eric Tan said: "It may take some time for the market to find equilibrium. Deal structures may tighten, sovereign support may step in, and the transaction model will evolve accordingly."