The capital markets are set to impose increasingly stringent standards on AI model developers. On September 14, Hong Kong stocks opened lower, with the Hang Seng Tech Index dipping as AI-related shares faced selling pressure. Zhipu (2513.HK) fell 5.67% in pre-market trading, with losses widening to 8.83% after the bell and briefly surpassing 10% during the session. MiniMax (0100.HK) followed suit, sliding more than 6%.

The selloff was triggered by a confluence of factors. On September 13, Zhipu announced a roughly US$5 billion fresh financing round. That same day, US model leaders OpenAI and Anthropic issued "slowdown" signals, emphasizing that in light of the trend toward recursive self-improvement (RSI) in AI models, developers need to temper their pace and prioritize investments in observability and alignment. These developments combined to spark a broad correction across Hong Kong's AI sector.

Persistent Capital Raises Stoke Funding Concerns

According to official announcements, Zhipu is placing up to 21.965 million new H-shares at HK$714 per share, a discount of approximately 9.96% to the prior day's closing price of HK$793. Concurrently, the company is issuing zero-coupon convertible bonds with a total principal of RMB 20.14 billion, featuring an initial conversion price of HK$892.5. Together, these measures are expected to raise net proceeds of approximately HK$39.274 billion.

In terms of fund allocation, around 60% will be directed toward the development of next-generation GLM foundational models, the buildout of a fully self-training system, and the deployment of computing infrastructure. The announcement defines "full self-training" as training next-generation GLM models within the environment established by their predecessors, forming a closed loop of RSI that encompasses three dimensions: self-generation of data, self-creation of environments, and self-optimization of infrastructure. Founder Tang Jie noted at the earnings conference that the core challenge lies in whether the model can autonomously determine when to stop training and when to correct its own errors.

While building up a war chest is intended to accelerate R&D, the pace of cash burn has drawn scrutiny. Zhipu completed multiple private funding rounds totaling approximately RMB 8.344 billion prior to its listing; its January 2026 IPO raised roughly HK$4.896 billion, which was fully consumed within eight months; a July placement brought in HK$31.375 billion, with HK$10.955 billion deployed over 50 days, averaging about HK$220 million per day. The September round adds another HK$39.274 billion, bringing cumulative post-listing financing to approximately HK$75.5 billion. The burn rate has accelerated from roughly HK$612 million per month during the IPO phase to approximately HK$6.57 billion monthly after July — a nearly tenfold increase in just two months.

Capital markets now expect these investments to translate into concrete financial results. Official data shows Zhipu's first-half revenue reached RMB 954 million, up 399.7% year-over-year, yet overall gross margin contracted from 50.0% to 26.4% compared to the same period last year. R&D spending remains 2.2 times the company's revenue for the period.

MiniMax is also confronting funding pressures. After listing in January at an IPO price of HK$165, the company unveiled a roughly HK$16 billion placement and convertible bond financing plan in July. First-half revenue stood at US$117 million, up 283.1% year-over-year; R&D expenses reached US$297 million, approximately 2.5 times revenue. Overall gross margin improved year-over-year to 17.9%, though it remains relatively low, and the adjusted net loss was US$293 million.

The financial results of both companies point to the same structural issue: the capital intensity of large-model development is rising sharply, while the ability to generate sustainable commercial cash flow has yet to keep pace.

Capital Markets Tighten the Screws

The market consensus has long been that AI entails high costs and heavy investment, with profitability unlikely in the near term. However, recent industry developments have introduced new dynamics. According to market reports, Anthropic generated over US$11.5 billion in second-quarter revenue with its first positive adjusted operating profit, and expects a second consecutive positive quarter in Q3. The company has entered substantive IPO preparation, targeting a Nasdaq listing in October with a potential valuation approaching US$2 trillion.

On one hand, there is an upcoming listing backed by improving profit signals; on the other, there are accelerating capital consumption rates and still-unproven commercial monetization capabilities. This contrast is steadily eroding the tolerance that capital markets have shown toward Hong Kong-listed model companies.

At the industry level, the "slowdown" consensus is also shaking the valuation foundation of the AI sector. Over the weekend, Anthropic CEO Dario Amodei called for a more measured pace in advancing model capabilities. OpenAI CEO Sam Altman echoed the sentiment the same day, announcing that OpenAI would not pursue an IPO this year, reasoning that "under the current safety landscape, going public would be unwise." Tesla CEO Elon Musk also voiced support.

Amodei acknowledged in his post that recursive self-improvement is already occurring across the industry, and predicted that within six to twelve months, uncontrolled clusters of AI agents could seize control of the entire internet. This framing has disrupted the traditional narrative within the model market. Historically, AI stock valuations rested on the assumption that massive capital infusions would drive continuous, rapid improvements in model performance. When industry leaders publicly advocate for slowing down, investors are forced to reassess whether those expectations still hold.

Analysts at Zheshang Securities' media and internet team argue that Anthropic's "slowdown" call will pressure AI hardware companies and model developers in the short term, while benefiting AI applications and terminal-device makers over the medium to long term. Additionally, the medium-to-long-term trajectory for AI computing infrastructure — particularly domestic computing power — remains unchanged. Global token usage will continue its explosive growth, rapidly driving up demand for inference computing. This suggests that the pace of frontier model competition may converge temporarily, leaving companies whose valuations depend on "capability narratives" subject to repricing.

When underlying models no longer iterate on a quarterly basis, smaller AI application companies with deep vertical focus actually see improved prospects, as the B2B sector prioritizes model stability and accuracy far above the pursuit of being "state-of-the-art."

The collective weakness in Hong Kong's AI sector is not an isolated event. AI chip-related stocks in Japanese and South Korean markets also opened sharply lower, with memory semiconductor, optical communications, and PCB stocks all declining — market sentiment was already under pressure before Hong Kong's session began. As Anthropic's financial performance and the AI safety narrative break with conventional logic, capital markets will hold model developers to increasingly exacting standards.

For the two Hong Kong-listed model companies, the next test is not merely whether token call volumes can keep growing, but whether those calls can solidify into stable revenue, sustainable gross margins, and positive cash flow.