China’s AI Model Duo Defy Lockup Expiry with Surge; Over 80% of Core Shareholders Hold, Wall Street Turns Bullish

Hong Kong's AI large model sector staged a historic counter-trend rally on July 8. Zhipu (2513.HK) and MiniMax, facing their largest lockup share expiries since listing, not only defied the widely anticipated "lockup expiry sell-off" curse but surged sharply. Zhipu spiked over 19% intraday, while MiniMax gained nearly 17%. Behind this rare price action lies a confluence of over 80% of core shareholders explicitly committing to hold and an unusually synchronized chorus of bullish calls from top-tier Wall Street investment banks, signaling a profound shift in how capital markets are valuing China's large model enterprises.

Market anxiety ahead of the expiry was not unfounded. Zhipu's unlocking involved approximately 25.68 million shares held by 11 cornerstone investors, representing 5.76% of total shares outstanding. Based on the previous day's closing price, the unlocked shares were valued at over HK$40 billion (approximately $5.1 billion). MiniMax faced an even sterner test, with its first tranche of unlocked shares reaching 146 million shares, accounting for roughly 63% of its Hong Kong-listed equity. Financial investors comprised a relatively high proportion, theoretically creating significant selling pressure. However, on the eve of the expiry, a series of public statements from core shareholders completely reversed market sentiment.

For Zhipu, JSC International Investment Fund SPC, an international investment fund under Beijing Financial Holdings Group, took the lead in stating its willingness to continue holding the company's shares based on a long-term positive outlook on Zhipu's development prospects. The core state-backed investors behind the fund include the Beijing Artificial Intelligence Industry Investment Fund, Beijing Jingneng Green Energy M&A Investment Fund, Beijing Information Industry Development Investment Fund, and Beijing Zhongguancun Science City Phase III Technology Growth Equity Investment Partnership. Subsequently, professional market-oriented investment institutions WT Asset Management and Optimas Capital Limited, along with early shareholder and cornerstone investor Lingyun Guang Technology Co., Ltd., also expressed their intention to hold for the long term.

MiniMax's shareholder structure similarly emitted strong stability signals. Its largest strategic shareholder, Alibaba (holding 12.65%), and miHoYo (holding 5.11%) clearly stated in late June their long-term optimism and commitment to holding without reducing their stakes. MiniMax's founding team voluntarily set a 12-month lockup period, exceeding the industry-standard six-month arrangement, meaning this initial unlocking involved no shares from the founding team or employee stock ownership plans. Among external shareholders, over 80% of pre-IPO and cornerstone investors have explicitly stated their long-term bullish view and intention to continue holding, including institutions such as Aspex, Boyu Capital, IDG Capital, as well as state-backed entities like China Life Investment and Xuhui Capital.

These concentrated lockup commitments effectively dispelled fears of a liquidity crunch, laying a foundation of confidence for the counter-trend share price surge.

While the shareholder structure stabilized, the collective bullish stance of top-tier Wall Street banks provided more solid fundamental support for the rally.

JPMorgan raised its target price for Zhipu from HK$1,800 to HK$2,000, maintaining an "Overweight" rating. The core logic of its research note is that Zhipu's newly launched GLM-5.2 has achieved global competitiveness, and its open-weight strategy is poised to leverage external infrastructure such as cloud service providers, inference platforms, and enterprise private deployments to further expand model usage scale, forming a growth flywheel of "broader distribution—larger usage scale—stronger paid conversion." JPMorgan noted that the market has largely priced in Zhipu's year-end $1 billion ARR guidance, but the scalable growth potential brought by the open-weight model still holds "option value" that has not been fully priced in.

For MiniMax, three international financial institutions—Goldman Sachs, Bank of America, and Citi—simultaneously issued "Buy" ratings, a rare occurrence in the current market environment of intensifying AI sector divergence. While each institution's focus varied, their conclusions all pointed in the same direction: MiniMax possesses the ability to sustainably enhance long-term value.

Goldman Sachs set a target price of HK$860 per share, with its report focusing on changes in the pricing environment of China's AI industry. Goldman noted that DeepSeek V4 is about to introduce differentiated peak/off-peak pricing, with peak-hour API prices set at twice the off-peak rate—an early signal of the industry's aggressive price war, ongoing since late April 2026, moving toward rationalization. Against this backdrop, MiniMax's M3 model, with its higher proportion of self-built optimized computing power and efficient architecture featuring smaller activation parameters, boasts gross margins significantly higher than peers.

Bank of America set a target price of HK$500 per share. Its report disclosed a significant shift in MiniMax's revenue structure: the company's revenue has pivoted from consumer-end products accounting for roughly 70% last year toward a higher proportion of enterprise and cloud API business, with enterprise/cloud API now assigned a higher strategic priority. On profitability, the previous-generation M2.7 model ultimately achieved an inference profit margin exceeding 40%, and Bank of America expects long-term margins to remain stable through continuous infrastructure efficiency improvements.

Citi set a target price of HK$533 per share, pointing to 53.8% expected upside from the current share price and forecasting that MiniMax's revenue growth will remain high. The bank noted that the upcoming next-generation video model could serve as a key catalyst to reverse market sentiment. Citi also highlighted recent developments including the MiniMax Code programming product and M3 model's integration with Alipay Token Pay, suggesting MiniMax is extending from single-model competition into more real-world business scenarios.

Technological capability remains the core focus for all institutions. Both Goldman Sachs and Citi emphasized the competitive advantages brought by MiniMax's newly released M3 model. Public data shows that M3 has entered the top tier of global open-source models across multiple international benchmarks, including Artificial Analysis, and achieved leading results in evaluations such as GDPval-AA. In terms of call volume, M3 ranks third globally on OpenRouter, having garnered 17.3 trillion token calls on that platform alone. Regarding generation speed, M3's output speed has improved from roughly 30 TPS at launch to 80 TPS, with MiniMax expecting a further 30%-40% improvement in the future.

More exciting for the market, according to a report by foreign media outlet The Information, MiniMax may launch its next-generation flagship model, M3 Pro, in the third quarter of this year. The model reportedly features 2.7 trillion parameters, with further enhancements in Agent capabilities, multimodality, and complex reasoning. If true, M3 Pro could become the largest model by parameter count ever developed in China, and potentially the largest open-source model globally.

Video models also represent an important product direction for MiniMax worth watching. According to Goldman Sachs' report, MiniMax is expected to release its next-generation H3 video generation model within weeks. The new model is based on an upgraded DiT architecture and further integrates large language model capabilities, promising significant improvements in video generation quality, feature richness, and understanding of complex scenes involving human motion and physical relationships.

The strong performance of the two AI model leaders also reflected a broader explosion in Hong Kong's tech sector. On the same day, the Hang Seng Tech Index opened and stayed high, surging nearly 5%. Alibaba jumped over 12%, Hua Hong Semiconductor rose over 10%, Lenovo Group gained over 9%, SMIC and Kuaishou climbed over 8%, and Tencent Holdings briefly spiked more than 4%. Analysts believe this round of strength in Hong Kong stocks is driven by two factors: first, low valuations in Hong Kong make them attractive; second, the global memory chip sector has continued to suffer heavy losses, with funds flowing out of those sectors into Hong Kong. Meanwhile, People's Bank of China (PBOC) Governor Pan Gongsheng delivered a speech at the "Hong Kong Fixed Income and Currency Summit cum Bond Connect Forum," outlining plans across four directions—deepening financial market connectivity, supporting the prosperity of Hong Kong's capital market, consolidating its offshore yuan hub status, and maintaining Hong Kong's financial stability—providing policy-level support for market sentiment.

Zooming out to the broader industry makes it easier to understand why international institutions chose this moment to collectively turn bullish on China's large model enterprises. For an extended period, in the race for developers and market share, China's domestic large model industry had fallen into a fierce price war, with many vendors offering API services at near-cost or even below-cost prices, continuously compressing overall industry profit margins. Recently, however, multiple domestic large model vendors have begun adjusting API pricing strategies, pausing price reduction initiatives. DeepSeek announced that after the official launch of DeepSeek V4 Pro in mid-July this year, it will implement differentiated peak and off-peak pricing, with peak-hour prices doubling off-peak rates. A growing number of vendors are making the same choice, beginning to rebalance price, cost, and service capability, with the competitive focus gradually shifting from a pure race-to-the-bottom on price toward a more constructive battleground of comprehensive competitiveness encompassing model performance, inference efficiency, commercialization capability, and profitability.

MiniMax is a direct beneficiary of this trend. Its M3 model holds clear advantages in inference cost and computing power utilization, with current API blended pricing at approximately $0.22 per million tokens—lower than DeepSeek V4 Pro's $0.35 per million tokens, but higher than DeepSeek V4 Flash's $0.12 per million tokens. While maintaining top-tier model capabilities, MiniMax has not relied on extreme low pricing to capture market share, instead achieving a favorable balance between model performance and cost.

However, the lockup expiry is not without risk. Industry analysis points out that financial investors account for a relatively high proportion of MiniMax's first tranche of unlocked shares, with stronger exit motivation and fewer constraints, implying some residual selling pressure post-unlocking. In comparison, Zhipu's unlocked share proportion is relatively small, and the largest unlocking entity by holding size is a cornerstone investor with state-backed background, making the pressure relatively manageable.

From a broader perspective, the two AI model leaders had already experienced significant corrections. Zhipu listed on the Hong Kong Stock Exchange at an IPO price of HK$116.2, with its market capitalization once surging past HK$1.33 trillion, but by the time of the lockup expiry, it had evaporated over HK$600 billion from its peak. MiniMax's market cap had surged to HK$410 billion in March before subsequently shrinking by nearly HK$300 billion. Combined, the two companies have seen roughly HK$900 billion in market value erased from their respective peaks.

Shen Meng, director at Chanson Capital, analyzed that the earlier global AI frenzy had fueled irrational rallies in large model company stocks, while the recent phased cooling of AI enthusiasm in US and South Korean markets has, to some extent, pricked the valuation bubble. Coupled with lockup expiry anticipation pressure, this has accelerated the decline of Hong Kong-listed AI concepts. UBS Securities China internet industry analyst Xiong Wei previously noted that large model companies had enjoyed two layers of special premiums—scarcity premium and liquidity premium—and both of these important valuation supports are now gradually fading.

What will truly determine long-term valuation going forward remains model capability and commercialization delivery. Lighthouse Capital partner Lou Yang observed that since December last year, a landmark shift in the industry has been large models completely breaking out of the previous simple chatbot form of chatting and Q&A, formally entering real-world Cowork collaborative work scenarios. Legend Holdings Vice President Ji Zhaofeng noted that the overall technical paradigm for large language models is currently relatively established, but this comes with diminishing marginal returns—meaning more investment no longer yields the same magnitude of returns as before.

For Zhipu and MiniMax, the lockup expiry is merely a short-term event. What will truly determine future competitiveness remains whether the next few generations of models can continue iterating, whether new products such as Agents and video generation can open up more application scenarios, and whether commercialization capability can be continuously delivered. For China's entire large model industry, this may also serve as a new signal—the industry is moving from "proving models can be built" to "proving models can create sustained value." As competition begins to return to technology, products, and business fundamentals, the large model industry has truly entered a new phase of competing on long-term capability.