Zhipu and MiniMax are surging, awaiting the 180-day ban lift

On July 8-9, Zhipu and MiniMax saw their first lock-up share expiration following their listings on the Hong Kong Stock Exchange.

Unlike the brutal era of AI 1.0, where "lock-up expirations triggered crashes and stock prices halved," this round of unlocks unfolded calmly, with relatively steady market sentiment and even a hint of optimism.

On July 8, Zhipu's stock price experienced a roller-coaster ride. Opening nearly 3% lower at HK$1,563, it rapidly plunged to HK$1,450 within minutes, extending losses to 10%. Just as the market expected a repeat of old-era patterns, buying flows surged in, reversing the price trend and pushing it to an intraday high of HK$1,689.

As more investors confirmed the willingness of cornerstone shareholders to hold their positions, with no signs of exit from state-backed long-term capital, market sentiment further recovered. Zhipu ultimately closed at HK$1,825, locking in a 13.35% gain on the unlock day, with a total market capitalization of approximately HK$8.137 trillion.

Affected by their shareholder structure and unlock ratio, MiniMax's stock saw greater volatility. On July 8, it closed nearly 12% higher; by July 9, its price tumbled over 20% at one point during trading, hitting a new low in nearly half a year. Overall, however, no major market disruption occurred.

Both Zhipu and MiniMax had pre-emptive rehearsals for the unlock, and by the expiration date, their valuations were no longer at their peak levels. Long before the formal unlock date, the market had already staged a "rehearsal" for this liquidity release through consecutive stock price declines.

When the actual unlock arrived, the market realized that those who intended to sell had seemingly already done so, and all panic-driven sentiment had long been fully digested.

More importantly, the sharply accelerating ARR curve injected a strong dose of optimism into the market, with both Zhipu and MiniMax setting firm targets to reach $1 billion in ARR this year.

Why No Major Crash Occurred?

The temporary soft landing for Zhipu and MiniMax can be attributed to at least three key factors: market sentiment, unlock strategy design, and their commercialization trajectories.

The underlying market sentiment is fundamentally different. During the AI 1.0 era, company lock-up expirations coincided with a global cooling of AI investment enthusiasm, the entire sector was labeled as a "bursting bubble," and investors voted with their feet in a wave of panic that snowballed.

In 2026 today, large language models have become a globally recognized core track, with widespread market optimism about AI's future development. Even though both companies' stock prices had already undergone significant corrections prior to the unlock, panic remained largely at the expectation level rather than escalating into emotional collapse. The market's long-term confidence in the AI sector remains unshaken, providing a psychological foundation for incoming buying support.

Differences in unlock strategy design determined the intensity of selling pressure. The sharp crashes during the AI 1.0 era were largely due to sudden, unexpected unlock events. Massive volumes of shares flooded the market simultaneously, with no preparation on the buying side. This cycle is completely different: the panic surrounding the unlock had been repeatedly priced in over time.

In the 10 trading days before the unlock, Zhipu's stock corrected approximately 39% from its peak; MiniMax saw an even larger correction. After hitting a high of HK$1,330 in March, it entered a months-long valuation reversion, with its price plummeting over 70% by the eve of the unlock, erasing nearly HK$300 billion in market capitalization from its peak.

When the unlock date finally arrived, the market breathed a collective sigh of relief. Markets do not fear bad news—they fear uncertainty; once the largest source of uncertainty is removed, there is full justification for contrarian trading.

In the AI 1.0 era, there were instances where approximately 70% of shares were unlocked at once, creating an immediate imbalance between supply and demand. This cycle, Zhipu's first unlock only accounts for 5.76% of shares, with large-scale unlocks not scheduled until next year, making the expansion limited and highly predictable.

The shareholder structures of both companies provide critical underlying support. Nearly 70% of Zhipu's cornerstone investors have state ownership backgrounds, with clear commitments to hold their shares; strategic shareholders of MiniMax including Alibaba and miHoYo collectively hold over 20% of shares, and have all publicly stated they will not reduce their positions. The proportion of financial investors is far lower than that of the "AI Four Dragons" back then, naturally reducing cash-out pressure. The unlock is not a one-time flood, but a gradual flow, giving the market sufficient time and space to absorb it.

Most importantly, the commercialization trajectory has begun to take shape, with a clear benchmark to reference—Anthropic. During the AI 1.0 era, when companies faced lock-up expirations, their business models remained unvalidated, revenues kept shrinking, and the market had lingering doubts about their corporate value. This cycle, while large model companies are still not profitable, their ARR growth curves are clearly visible.

Zhipu's March financial report revealed that its API call ARR exceeded $250 million as of March, with a target of reaching $1 billion in ARR by year-end. Some external investors estimate its year-end ARR will land between $1.5 billion and $3 billion.

MiniMax disclosed that its ARR surpassed $400 million in May this year, and the company is confident in achieving $1 billion in ARR revenue by the end of the year.

This means the market is no longer paying for pure concepts, but now has tangible metrics to measure the value of large model enterprises.

Historical lessons tell us that the stock market never abandons a company simply because it is unprofitable—it only discards companies whose future outlook becomes unrecognizable. Zhipu and MiniMax's soft landing reflects the market's belief that while their growth trajectories are still on the ascent, their directional path has become increasingly clear.

This evolution of the value system, from rough valuation to precise pricing, may be a far more noteworthy signal than the unlock event itself.

The Widening Valuation Gap

Zhipu and MiniMax form an interesting comparative pair, perfectly summed up by the saying "times change, fortunes turn."

At the time of their listings, MiniMax was the standout star. On its first day of trading on the Hong Kong Stock Exchange, MiniMax's price surged 109%, outshining Zhipu which listed on the same day, with its market capitalization breaking the HK$1 trillion mark.

Back then, Zhipu was labeled as a B2B-focused enterprise, with its model of pursuing projects and acquiring clients dismissed as outdated—far less appealing than MiniMax's control of overseas consumer-facing markets. New narratives carried greater imaginative potential, naturally driving higher valuations. For a period, MiniMax completely overshadowed Zhipu in market hype.

The reversal of this dynamic ultimately stems from the two companies choosing fundamentally different paths.

Zhipu followed a "focused" strategy, fully committing to code generation and agent directions as early as the start of 2025. It remained in a catch-up position until June this year, when Anthropic's service disruption event immediately elevated GLM-5.2 to the top position as the leading domestic alternative.

In hindsight, the viral "lobot" craze was fleeting and superficial. Various domestic consumer-facing AI applications could only capture partial consumer traffic, and relying on subscription fees to drive commercial revenue was unsustainable. In contrast, while Zhipu received less public attention during the lobot boom compared to MiniMax and Kimi, it precisely captured the massive opportunities in AI coding.

Compared to consumer-facing AI chatbots, AI coding has stronger exclusivity. Once development teams become accustomed to using GLM to write code, debug interfaces, and run agents, high migration costs create a certain competitive moat. The deep integration of 250,000 developers not only allowed Zhipu to capture traffic, but more importantly, embed itself into their core workflows.

This exclusivity granted Zhipu pricing power. Its price increase for GLM-5.2 did not drive away clients—instead, it drove a 400% surge in usage volume. For developers, the time and effort saved by using GLM-5.2 far outweigh the extra few dollars spent per million tokens.

This allowed Zhipu to rapidly form a self-reinforcing loop: price increases lifted revenue expectations, driving valuation growth, which attracted more capital inflows to support the development of next-generation higher-performance models. Driven by the principle of "scarcity creates value," further price increases for the model still find willing paying users.

MiniMax chose a more diversified path, adhering to its native multimodal strategy while simultaneously deploying across video generation, consumer-facing applications, and multiple vertical industry sectors.

At this stage, the market is revealing an extremely harsh reality: only two categories of large model vendors can break through and survive. They are either top-tier SOTA vendors that secure premium pricing through absolute technological leadership, or ultra-low-cost vendors that outcompete all others via cost advantages. Both groups can find viable survival logic, with the former capturing high profits and the latter capturing market share.

However, MiniMax's M3 model is trapped in the awkward middle ground. Its coding capabilities lag behind Zhipu, its cost-effectiveness cannot compete with DeepSeek, and in video generation it is outperformed by ByteDance's Seedance. Whether in SOTA performance or iteration speed, Hailuo Video cannot match its rivals. Neither the best nor the cheapest, this "no-man's-land" positioning pushes MiniMax into the most uncomfortable competitive squeeze.

To make matters worse, MiniMax's inherent consumer-facing DNA naturally traps it in a price-sensitive market. Users arrive quickly and depart just as fast, with loyalty far lower than B2B clients integrated into core workflows. A previous attempt to raise prices was met with market pushback, forcing the company to cut prices to stabilize operations. This lack of pricing power directly suppresses revenue expectations and drags down its valuation.

JPMorgan Chase describes this divergence as a winner-takes-all dynamic: models with stronger capabilities are far better at converting open distribution channels into paid monetization, while models with insufficient differentiation face faster price-comparison competition and the risk of user traffic shifting away.

The market offers no survival space for mid-tier players, which may be the cruellest reality facing the large model industry today.

Next Stop: $1 Billion in ARR

If you observe closely, large model vendors are all advancing toward the $1 billion ARR target by the end of this year.

What does $1 billion in ARR represent? It roughly matches Anthropic's level at the start of 2025, but more importantly, it signals the underlying growth trajectory.

Propelled by the explosive adoption of Claude Code in the coding space, Anthropic achieved an astonishing leap from $9 billion to $30 billion in ARR in the first quarter of 2026. In its revenue structure, the vast majority of ARR comes from usage-based API services, contributing 75%-85% of total revenue. In comparison, subscription revenue only accounts for 15% of total ARR.

Zhipu's strategic positioning in B2B API services perfectly aligns with this commercial logic. Domestically, Zhipu has secured major internet giants including Tencent and ByteDance as clients—high-token-consuming customers that serve as ideal organic sales channels. The more they use the service, the more frequent API calls become, driving direct, immediate revenue growth.

Internationally, its partnership with AWS further opens up global distribution channels. Additionally, local deployment demand from central and state-owned enterprises forms another stable, continuous source of consumption. The three-tier customer structure ensures Zhipu's API business not only has large scale, but also stable quality. When paired with Anthropic's high-growth narrative, capital markets are highly willing to assign it a higher valuation premium.

The core driver of Anthropic's growth was identifying a killer application scenario, which relies on three critical elements: a scenario with massive demand and extremely high value, exceptional commercialization efficiency, and a flywheel effect generated by high revenue.

For Chinese large model enterprises to narrow the gap with global leaders, the key lies in finding their own "Claude Code moment."

However, the AI coding path represented by Anthropic may not be the only correct answer. At least from MiniMax's perspective, they believe code generation is only the first step in AI commercialization, and the true future opportunities lie in high-value, knowledge-intensive vertical industries such as finance, law, consulting, healthcare, and cybersecurity—with coding itself being just one of these verticals.

If AI coding enhances programmers' productivity, then vertical industry AI will completely redefine how professional workers make decisions. Financial analysts, legal advisors, doctors, and management consultants are far more sensitive to productivity improvements than ordinary consumer users. Once AI is embedded into their core workflows, migration costs and willingness to pay will rise to an entirely different level.

In the video generation sector, Seedance has already proven the market's explosive potential and demonstrated a fully functional commercial loop. The content creation scenario is far from reaching its ceiling. A notable data point: Kunlun Tech's AI short drama platform business exceeded $700 million in ARR in the second quarter of 2026, while its AI tool business surpassed $100 million in ARR.

Thus, for domestic large model vendors, the "Claude Code moment" is not a fixed, pre-determined answer—it is an open-ended question.

While paths may differ, the core challenge remains the same: identify that killer application scenario that can shift ARR from linear growth to exponential explosion, then pull the trigger with full force.

This article originates from the WeChat public account "guangzi0088" (ID: TMTweb), written by Hao Xin, edited by Wang Pan, and published with authorization from 36Kr.