“Dissecting the Reality and Illusion of US$500 Billion in Computing Power Financing” — Dr Zhou Yuanfan, Chairman of the Board of Anrong (Hong Kong) Credit Ratings, in an Exclusive Interview with Econo

On August 14, Economic Observer conducted an exclusive interview with Dr Zhou Yuanfan, Chairman of the Board of Anrong (Hong Kong) Credit Ratings Co., Ltd. On the same day, the interview was published in Economic Observer and on the Economic Observer website under the title “Dissecting the Reality and Illusion of US$500 Billion in Computing Power Financing.” The article provided a comprehensive analysis of the underlying credit logic and substantive risks of computing power financing, focusing on key market issues including: “Can computing power become an independent credit asset? How can the mismatch between financing terms and the economic life of GPUs be addressed? Can structured arrangements such as SPVs and leasing truly achieve risk transfer?”

Dr Zhou believes the essence of the current round of computing power financing is redefining GPU computing power as an infrastructure asset capable of generating stable cash flows. From a credit rating perspective, however, its credit foundation is a multi-layered composite structure. The credit quality of the ultimate lessee is the most critical cornerstone, while GPU equipment and infrastructure serve only as risk buffers and secondary repayment sources. This logic has been repeatedly validated by mainstream overseas computing power financing cases. Regarding the widely discussed mismatch between financing terms and GPU economic life, Dr Zhou stated that this is the most significant structural challenge currently facing credit ratings for computing power financing. Generally, GPUs have a prime profit-generating period of only two to three years and a full-cycle economic life of approximately six to eight years. During the first three years, they support high-value training workloads, before shifting to lower-margin inference operations, resulting in a stepwise decline in rental rates. If financing terms are extended to seven to ten years, this essentially relies on rolling cash flows from multiple generations of equipment to cover debt from a single financing period, creating substantial maturity mismatch risk.

Dr Zhou further explained that when financing terms significantly exceed the economic life of GPUs, credit ratings follow the principles of “conservative assumptions” and “dynamic buffers,” conducting stress tests across three dimensions—residual value, rental income, and refinancing—rather than applying the linear assessment logic of book depreciation. Regarding market practices that seek to transfer risk through arrangements such as SPVs, leasing, minimum usage commitments, or residual value guarantees, Dr Zhou stated bluntly that such structured arrangements are more often accounting-driven off-balance-sheet designs than genuine transfers of credit risk. Rating agencies strictly adhere to the principle that “substance prevails over form,” identifying all implicit credit obligations through a look-through approach and incorporating them into a unified assessment of the entity’s debt-servicing capacity.

Dr Zhou explained that, in legal form, independent SPVs can achieve bankruptcy isolation of assets; operating leases may be accounted for off-balance-sheet under certain standards; and minimum usage commitments and residual value guarantees are often packaged as “commercial cooperation terms” rather than fixed liabilities. However, in economic substance, such arrangements often leave the core risks with the originator or core lessee: minimum usage commitments are essentially unconditional payment obligations, no different from conventional lease liabilities; residual value guarantees are commitments by the originator to absorb asset depreciation and constitute typical contingent credit liabilities.

Regarding the fundamental question of whether computing power itself can become an independent credit asset, Dr Zhou offered a clear judgment: under current technological conditions and market infrastructure, computing power does not yet meet the necessary conditions to become a credit asset on its own. The market’s discussion of “investable computing power” refers to asset allocation at the broad asset-class level, whereas the core financing support on the credit side remains long-term usage contracts, operating cash flows, and the corporate credit of computing power lessees. GPUs and computing power facilities can only serve as collateral buffers and secondary repayment sources.

At the conclusion of the interview, Dr Zhou stated that AI infrastructure development is a core pillar supporting the growth of the digital economy, and computing power financing provides an important channel for industrial capital participation. However, amid market enthusiasm, it is even more necessary to return to the essence of credit and rationally assess the risk-return characteristics of computing power assets. Financial institutions participating in computing power financing should look through to the underlying credit entities, fully assess technology iteration risks and maturity mismatch pressures, remain vigilant against hidden off-balance-sheet risks concealed by structured designs, and prudently plan their business deployment. In the future, as the computing power trading market system continues to improve and valuation and pricing mechanisms gradually mature, the credit attributes of computing power assets are expected to strengthen further, promoting the standardisation and maturity of the computing power financing market and better empowering the high-quality development of the AI industry.