In a recent research report, Dario Perkins, an economist at research and consulting firm TS Lombard, expressed a bearish outlook on the commoditization of AI computing power, worrying that GPU supply bottlenecks could eventually lead to overcapacity and trigger a price war in the computing power leasing market. Market analysts estimate that approximately $500 billion worth of GPU equipment remains to be installed. As this equipment gradually comes online, GPU leasing prices, currently high due to computing power shortages, could fall significantly.
Perkins points out that data center operators bear fixed costs such as GPU procurement, infrastructure construction, debt, and long-term "take-or-pay" contracts. Even if computing power demand is insufficient, some payment obligations still need to be fulfilled. If leasing prices fall, high fixed costs could force operators to continue providing computing power, further intensifying price competition and cash flow pressure.
The report also warns that cutting-edge model developers face challenges such as declining token prices, continuously rising training costs, and intensified competition in open-weighted models. Due to cross-shareholding and revolving financing arrangements between model companies and data centers, the financial pressures of both parties could be mutually transmitted, creating a risk amplification effect. Perkins believes that if there is a severe oversupply of computing power, the industry may experience a round of financial difficulties and bankruptcy reorganization, while the economic benefits brought by the popularization of AI technology will flow more to end users.
