The next wave of artificial intelligence debt financing is taking off in Asia.

Companies building hundreds of data centers across the region are also seeking funds to buy the advanced computer chips that will power them. While borrowing to acquire graphics processing units has grown rapidly in the US, the few such loans secured in Asia have mostly involved private credit funds that were willing to take more risk.

Now banks are becoming more comfortable with GPU financing, significantly widening the pool of capital available for the next phase of the AI race. That money will be crucial given PricewaterhouseCoopers LLP estimates Asia’s spending on data centers could reach $8.2 trillion by 2050, with the vast majority going to hardware including GPUs and servers.

In recent months, banks played key roles in GPU loans totaling roughly $3.8 billion to GMI Cloud and two other AI infrastructure providers, Zankore and PaleBlueDot AI. Citigroup Inc. was the sole debt adviser for Zankore’s $3.1 billion borrowing in Indonesia, while JPMorgan Chase & Co. was the placement agent for the credit facility raised by PaleBlueDot AI.

Half a dozen bankers and financial advisers across Asia said they were either in talks for or aware of more financing deals linked to GPUs, though lending is still nascent because of concerns over how to value the chips longer term and assess risks tied to geopolitical tensions and AI itself. They asked not to be identified discussing private information.

“As deal sizes grow and borrowers push for more competitive pricing, banks will be increasingly important,” said Eric Tan, a banking and finance partner at Hogan Lovells Cadwalader. But GPU financing also exposes lenders to “rapid depreciation, technology obsolescence and volatile rental rates” due to how quickly the technology changes, he said.

Among traditional lenders that have waded into Asian GPU financing, global investment banks have so far been the main players because of their expertise in the complex structures, according to people familiar with ongoing talks. Some US lenders have drawn on specialists in their head offices to assess chip valuations, one banker said.

Citigroup, JPMorgan Chase, Barclays Plc, Deutsche Bank AG, Banco Santander SA and Japan’s Sumitomo Mitsui Banking Corp. are currently evaluating GPU-linked loans, according to people familiar with the matter. GMI Cloud is in talks with banks and private lenders for a new $300 million loan to buy chips for its Thai data center, separate people familiar said last month.

Representatives for Citigroup, Barclays and Deutsche Bank declined to comment. JPMorgan, Santander and SMBC did not immediately reply to requests for comment.

Still, Asian banks are starting to get more active in the space. Singapore’s United Overseas Bank Ltd., which jointly underwrote Zankore’s $3.1 billion loan with four other banks, is now leading talks as the data center operator tries to raise a fresh $6 billion from banks, according to Zankore Chairman Vikram Sinha.

Working with banks is “the right way” because Zankore will need to keep raising money as it expands its AI data center capacity by 10 times to 1 gigawatt, Sinha said at a conference on Sept. 22. “We were very clear [about] the scale we are looking at. We wanted to go the hard way — we wanted to work with banks and a bank syndicate.”As banks become a bigger source of funding, companies seeking to borrow for chip purchases will have to do more to prove their projects will generate enough revenue to make good on their loans. Traditional lenders will likely demand more conservative underwriting and higher debt service reserve requirements.

Mike Arougheti, who runs Ares Management, one of Asia’s largest private credit lenders, has said that his firm is also keeping the bar high even though GPU financing has the biggest funding gap in the AI boom.

“You have to be leading with your risk appetite and not your appetite to deploy,” Arougheti said at the Barclays Global Financial Services conference in September. “No one could really articulate, at least to me, what the depreciation curve looks like for that technology,” he said, adding that returns are limited and usually only around 100 to 200 basis points higher than other kinds of AI infrastructure lending.

GPU loans in Asia have largely followed a structure pioneered by CoreWeave Inc., one of the earliest and biggest users of this type of financing. In many deals, the loans are repaid with revenue from selling a data center’s computing power. Customer contracts and the chips themselves often serve as security for the loans.

For a lender, the most important question then becomes: how reliable are those customers and how long are the contracts?

Deals backed by Nvidia Corp. have been an easy yes. In both the GMI Cloud and Zankore loans completed last month, the US technology giant agreed to purchase any unsold computing capacity, providing a backstop if customer contracts fall through. In return, the companies, which are part of Nvidia’s cloud partner program, will charge buyers higher prices than Nvidia’s promised rate and share their revenue with Nvidia.

Major Chinese tech companies have emerged as another key source of demand for computing power in Asia. They’re a relatively safe bet since they’re also able to draw in Chinese banks, according to a loan banker at a European lender.

Tencent Holdings Ltd. would be the end user of chips GMI Cloud wants to buy for its Thai data center, people familiar said last month. Xiaohongshu, a popular Chinese social media platform, was set to buy computing capacity from chips purchased by PaleBlueDot AI for its project in Japan, separate people familiar said in December.

But Chinese clients also come with their own baggage. The US already restricts chip exports to mainland China and is considering new legislation that aims to curb Chinese firms’ remote access to Nvidia chips via cloud computing, which would restrict them from renting computing capacity located in other countries.

“Large Chinese hyperscalers generally have strong credit quality,” said Yijing Ng, an analyst at S&P Global Ratings. However, banks will have to conduct more due diligence to ensure they’re complying with regulations and the ultimate users of the computing power aren’t violating any sanctions or restrictions, she said.

At the same time, the existential debate about AI’s dangers and the need to impose guardrails on its growth will continue to influence risk assessments. “The markets are likely to take a while to find an equilibrium,” said Tan from Hogan Lovells Cadwalader. “Deal structures may tighten up, sovereign support may come into play and the model for deals will evolve as a result.”