China's Zhipu AI shares plunge 23% amid computing shortage

Shares in Chinese artificial intelligence firm Zhipu AI fell nearly 23 per cent on February 23, erasing more than HK$70 billion (US$9 billion) in market value, as mounting concerns over computing capacity constraints overshadowed the company’s recent stock market surge.

The Hong Kong-listed company, known internationally as Z.ai, closed at HK$560. Despite the sharp drop, the shares remain more than 380 per cent above their January 8 listing price, South China Morning Post writes. 

The sell-off followed an unusual public appeal last week in which Zhipu sought partnerships with global and domestic providers of inference computing power, offering to share revenue in exchange for access to high-performance infrastructure. It marked the first time a Chinese AI firm has publicly called for additional computing resources.

“Inference” refers to the deployment phase of trained AI models — the stage at which users interact with the system. Unlike training, which requires massive but often one-off computational investments, inference computing represents a recurring cost directly tied to user demand and revenue generation. Consultancy McKinsey & Company projects that inference workloads will account for more than half of total AI computing demand by 2030.

Globally, Zhipu is seeking “high-performance compute providers” operating “advanced GPU [graphics processing unit] clusters”, said Carol Lin, CEO of Z.ai International.

The appeal comes amid mounting service complaints from users of the company’s flagship “GLM coding plan” product. On February 21, Zhipu issued an apology after customers in China and overseas reported response delays and rate limits. It was the company’s second public apology since its initial public offering in early January.

“We feel very ashamed after reading all your feedback messages on various platforms,” Zhipu said in its most recent statement. “Please give us a little more time, and we will take real action to prove ourselves.”

Following its IPO, Zhipu restricted new user registrations in an attempt to stabilise service quality for existing customers. However, complaints have continued across community platforms including Reddit and Discord.

The company operates separate inference clusters for domestic and international users. Surging overseas demand — particularly for AI-assisted coding tools — has strained its available capacity, complicating efforts to expand revenue beyond China.

Zhipu’s challenges have been compounded by geopolitical pressures. Blacklisted by Washington earlier this year, the firm faces restrictions on access to advanced US semiconductor technology, including chips from Nvidia. In response, Zhipu has deepened cooperation with China’s domestic chipmakers as Beijing pushes to reduce reliance on US suppliers.

According to its latest technical disclosures, the company collaborates with several Chinese semiconductor firms to adapt its GLM-5 model for domestic hardware. Yet in a February 16 statement announcing a new “computing power partnership” initiative, Zhipu acknowledged that “repeated efforts to expand domestic chip supply … [were] unable to completely resolve the current supply shortage”.

Despite operational headwinds, Zhipu has attracted strong investor interest amid enthusiasm for China’s expanding AI sector. However, its IPO prospectus flagged dependence on third-party computing providers and a limited commercialisation track record as key risk factors.

By Sabina Mammadli