(Yicai) May 22 -- The explosion in computing power demand driven by artificial intelligence agents is reshaping the industry's competitive landscape, but token pricing offered by Chinese providers is "far below that of US peers," with a much smaller performance than price gap, according to an analyst at UBS Securities.

The practical application of AI agents has surged this year, with the computational power required to complete an agentic task significantly more than that of traditional conversational AI, Xiong Wei, China internet analyst at UBS Securities, said at the Asia Investment Conference media briefing yesterday. "The token consumption for completing an agentic task can be more than 1,000 times higher than that in AI chat use."

Global token consumption was up about tenfold for the week through May 15 from a year earlier, according to third-party data from a model application programming interface aggregation platform.

Token use is on an exponential growth trajectory, Xiong stressed. Demand would prove increasingly durable as agentic use spreads from tech enthusiasts to more knowledge-intensive workers, he noted.

However, whether in chip production or data center building, computing power capacity growth follows a more linear path, while demand has been increasing exponentially, Xiong said. "Companies are facing a token demand that exceeds supply, with even their own computing power reserves being insufficient to meet the potential customer needs they have observed," he said, adding that the supply-demand imbalance "may last longer than expected."

Smaller Performance Gap

In the computing power race, Chinese AI model developers can offer better token prices compared to US peers, despite the smaller performance gap, thanks to their sustained focus on inference efficiency optimization, which allows them to "drive their cost advantage as low as possible," Xiong said.

China's computing costs are roughly half those of the United States, underpinned by a stable electricity supply, noted Kenneth Fong, head of China internet research at UBS Investment Bank.

Regarding the risk of a price war, Xiong said that in a supply-constrained market, cutting prices to grab a token share "doesn't make sense, because even if you win that demand, you may not have enough computing capacity to meet it."

Leading providers are actually rising token prices while discontinuing low-priced subscription tiers, launching premium plans, and migrating toward usage-based billing, he pointed out.

Roughly 78 percent of information technology decision-makers at Chinese firms have moved beyond the pilot stage into active project deployment, according to the findings of a UBS EvidenceLab survey with over 100 respondents. Generative AI spending accounted for about 12 percent of the IT budgets last year, with AI-related budget growth outpacing overall IT spending at a rate faster than the equivalent US survey. But the unclear return on investment remains the biggest constraint.

According to Xiong, the findings mirror results from US counterparts. "Both Chinese and US companies are showing the same pattern."

Logic Shift for Internet Firms

The internet industry's underperformance since the start of the year, including macro headwinds, rising AI capital expenditure, and uncertainty around disruption, has been fully reflected in stock prices, Fong said. "Low expectations, low valuations, fundamentals gradually bottoming out, all imply the stock price should have bottomed."

Among subsectors, Fong is optimistic about the high resilience of AI model companies and the low valuations of the gaming sector, where leading firms have a price-to-earnings ratio of only about 12 to 13 times.

More fundamentally, the industry's core logic has shifted, Fong noted. "In the past, you made money from time spent, with the more traffic you sold, the more advertising or transaction revenue you earned.

"Now, internet companies charge a share of the value they create for users," Fong said, adding that the transformation will push their monetization ceiling to new heights.

Editor: Martin Kadiev