On July 20, 2026 (Beijing Time), Zhipu (02513.HK), known as the "first AI foundation model stock" in Hong Kong, suffered another sharp intraday decline. The stock closed at HKD 890.50, down 19.56% for the day. Over the past five trading sessions, its cumulative loss widened to 46.23%. Following a 28.49% plunge on Friday, July 17, Zhipu continued its downward trend, hitting an intraday low of HKD 920 and officially breaking below the critical HKD 1,000 mark.
This is not an isolated incident. Since hitting a record high of HKD 2,980 on June 22, Zhipu’s share price has pulled back more than 66% in just a month, wiping out over HKD 800 billion in market capitalization. During the same period, MiniMax (00100.HK), another leading Chinese AI foundation model company, saw its stock price drop more than 85% from its peak, approaching its IPO price. The simultaneous collapse of these two domestic AI giants signals a fundamental shift in how the market values the AI sector—from chasing model capabilities to verifying commercial returns.
Financing, Share Unlocks, and Supply Glut: A Sudden Shift in Market Dynamics
One of the triggers for Zhipu’s latest correction came from a surge in supply.
On July 8, Zhipu saw its first large-scale post-IPO share unlock, with 25.6816 million restricted shares (5.76% of total share capital) becoming tradable, mainly held by 11 cornerstone investors. Before this event, the company had only 11.74 million freely tradable shares. The unlock more than doubled the float, creating significant short-term supply pressure and weighing heavily on the share price.
Shortly after, on July 13, Zhipu completed a placement of 19.78 million new H-shares. This further increased the available shares in the market, and the placement price was below the prevailing market price, intensifying selling pressure as investors faced unrealized losses. Market data shows most placement participants are currently underwater.
What worries the market even more is the long-term pressure. In January 2027, a much larger share unlock is scheduled, when about 40% of the original shares will become tradable. For AI model companies that have yet to establish a stable profit model, ongoing financing and market expansion are often seen as signs of high capital dependency, indicating they are still far from achieving self-sustaining positive cash flow.
AI model training is inherently capital-intensive. GPU procurement, data center construction, compute leasing, and top-tier talent compensation all require massive, ongoing expenditures. Until revenue reaches scale, any form of equity dilution is seen by the market as a signal that valuations are under pressure.
The Kimi K3 Shock: A New Competitive Landscape for Chinese AI Foundation Models
While share unlocks and placements create "quantity" pressure on the supply side, the launch of Kimi K3 has triggered a "price" revaluation on the demand side.
In the early hours of July 17, Moonshot AI announced the release of its next-generation foundation model, Kimi K3, with a staggering 2.8 trillion parameters—making it the world’s largest open-source model by parameter count. Kimi K3 topped the authoritative Frontend Code Arena leaderboard with a score of 1,679, surpassing Fable 5 (1,631) and GPT-5.6 Sol (1,618), and became the first open-source model to outperform all overseas closed-source models on a major programming benchmark.
This event carries industry-wide significance far beyond a single product launch. Code generation is one of the most commercially promising AI applications. By surpassing closed-source giants in core capabilities through open-source means, Kimi K3 directly challenges the high-pricing business model that overseas closed-source models have maintained through technological moats. Following this news, the Philadelphia Semiconductor Index fell 12.5% in a week, entering a technical bear market, while the US AI sector lost about $470 billion in market value within 72 hours.
The shockwave quickly hit Chinese AI peers. On the day K3 was released, Zhipu plunged 28.49% and MiniMax dropped 15.62%. In a research note on July 18, Goldman Sachs pointed out that this reflects market concerns over the competitive landscape and long-term winners in China’s AI model sector, noting that "leadership and sustainability among model companies remain highly volatile."
The competitive matrix for Chinese AI foundation models is being redrawn. Zhipu is known for its GLM series and enterprise market presence, but the market is increasingly focused on its commercialization pace. Moonshot AI is rising rapidly with the Kimi ecosystem and long-context capabilities, making user growth a key metric. DeepSeek is building a differentiated position through open-source influence and cost advantages, while Baidu’s Wenxin leverages its enterprise ecosystem for real-world deployments. As the technological gap between these players narrows, the lifespan of technical advantages shortens, and model capabilities alone are no longer the sole basis for valuation.
CICC notes that Kimi K3 is built on hybrid linear attention mechanisms and residual attention technology, natively supporting visual understanding and million-token context windows. This means Chinese AI models are not only breaking through in parameter scale but also accelerating innovation at the architectural level. As the "technology lead" window shrinks from years to months, the market naturally questions the valuation premiums of individual companies.
AI Valuations Enter the "Revenue Validation Phase": A Systemic Shift in Global Investment Logic
Zhipu’s plunge is not just a stock-specific event in China—it is a microcosm of a systemic shift in global AI investment logic.
By 2026, the industry’s long-held Scaling Law is hitting limits: simply amassing compute and data to boost model performance is yielding diminishing returns. Spending billions to train the next generation of super-sized models no longer delivers commensurate breakthroughs. Both public and private markets are shifting their AI valuation anchors from technical prowess to commercial viability.
This shift is clear: the market is moving from "AI story trading" to "AI cash flow trading." Investors who once paid high multiples for AI companies now focus more on capital expenditure efficiency, margin resilience, and the quality of profit growth. As one market analyst put it, AI trading is moving from "buying the future narrative" to "proving investment returns."
Global AI investment is now clearly divided into stages. Stage One was infrastructure investment—building out GPUs, cloud computing, and data centers, with primary beneficiaries like Nvidia and cloud giants such as Amazon, Microsoft, and Google. Stage Two is commercial validation—whether AI SaaS, enterprise applications, and AI agents can turn that infrastructure investment into real revenue and profit.
This logic shift will be further tested in the upcoming earnings season. Alphabet (Google) and Microsoft are expected to release their Q2 2026 earnings on July 29, with Amazon following on July 31. The market’s focus has shifted from "AI investment scale" to "can massive capex deliver tangible returns." Some analysts note that in 2026, the world’s top five tech giants are expected to spend over $690 billion on AI capex, while direct AI service revenue may only reach about $25 billion. This stark contrast is forcing investors to reassess the entire AI value chain’s valuation logic.
In this environment, model companies face especially acute valuation pressure. Unlike infrastructure providers, model companies must answer the question: "Can technological investment translate into revenue?" Zhipu’s API platform reported annual recurring revenue (ARR) of about RMB 1.7 billion as of March 2026—a 60-fold increase over the past 12 months. By July, ARR reportedly reached $1 billion. However, for a company once valued at over HKD 1 trillion, the market clearly demands faster and more verifiable commercial results.
Not the Disappearance of AI Demand, but a Restructuring of Valuation Logic
Zhipu’s recent plunge does not signal a collapse in AI demand, nor does it mean the outlook for China’s AI foundation model industry has reversed.
The fundamentals of China’s AI sector are still evolving rapidly. Goldman Sachs forecasts that API and subscription revenue for Chinese AI models will grow from an estimated RMB 35 billion in 2026 to RMB 879 billion in 2030—an increase of about 25 times. AI foundation models have already crossed the "production-grade inflection point," moving from labs into real-world production environments. These long-term trends remain unchanged despite share price volatility.
What’s really changing is the way the market prices these companies. As the AI industry shifts from "burning cash to tell stories" to "running the numbers for returns," investors are paying closer attention to profitability, cash flow improvements, and return on investment. As one market observer noted, whenever a company’s profit realization lags market expectations, or when AI commercialization slows, high-valuation assets may face correction pressure—"This is simply how capital markets operate, not a sign of long-term weakness in the AI industry."
For model companies, future valuations will increasingly depend on verifiable business metrics: AI API revenue growth, the number and quality of enterprise customers, the trajectory of inference cost reductions, gross margin improvements, and the real-world scale of commercial deployments. The importance of model size itself is giving way to "how much revenue these models can actually generate."
Zhipu’s drop from HKD 2,980 to HKD 890 in just one month is a microcosm of the market’s shift from "AI story trading" to "AI cash flow trading." For the industry as a whole, this may not be the end of a bubble, but the beginning of a healthier and more sustainable valuation system.
FAQ
Q: What are the main reasons behind Zhipu’s stock plunge?
A combination of factors: The July 8 share unlock and July 13 new H-share placement significantly increased market supply; on July 17, Moonshot AI released the 2.8-trillion-parameter open-source model Kimi K3, prompting a revaluation of the competitive landscape among Chinese AI foundation models; and, globally, AI investment logic is shifting from "story-driven" to "revenue validation."
Q: How exactly did Kimi K3 impact Zhipu?
Kimi K3, with 2.8 trillion parameters, became the world’s largest open-source model and surpassed all overseas closed-source models on key programming benchmarks. This shattered expectations for the sustained technical lead of early movers like Zhipu, signaling that the technology gap among Chinese models is closing rapidly and that model capability is no longer the sole basis for valuation.
Q: How is the AI industry’s valuation logic changing?
The market is shifting from "model scale → high valuation" to "revenue growth → high valuation." Investors are moving their focus from parameter count and model capability to verifiable business metrics like AI API revenue, enterprise customer numbers, inference costs, gross margin, and the scale of commercial applications.
Q: Does Zhipu’s plunge mean the outlook for China’s AI industry is bleak?
Not at all. Goldman Sachs forecasts that China’s AI model revenue will grow from RMB 35 billion in 2026 to RMB 879 billion in 2030. The plunge reflects a clearing of valuation bubbles and a correction in pricing logic, not a disappearance of industry demand.
Q: What stage is AI investment currently in?
Global AI investment is moving from the first stage (infrastructure investment: GPUs, cloud computing, data centers) to the second stage (commercial validation: AI SaaS, enterprise applications, AI agents). Model companies face greater valuation pressure, while application-layer companies that can prove commercial returns will be revalued by the market.
