Author | Hao Xin

Editor | Wang Pan

From July 8 to 9, Zhipu and MiniMax faced their first lock-up expiration after listing on the Hong Kong Stock Exchange.

Unlike the tumultuous AI 1.0 era, where lock-up expirations often led to sharp declines and halved stock prices, this round proved relatively calm, with the market even showing a hint of warmth.

On July 8, Zhipu's stock price experienced a rollercoaster ride. Opening at HK$1,563, nearly 3% lower, it quickly dropped to HK$1,450 within minutes, widening the loss to 10%. Just as the market anticipated a repeat of past scripts, buying surged, reversing the trend and touching a high of HK$1,689.

As more investors confirmed the lock-up intentions of cornerstone shareholders and noted no signs of long-term state-backed funds exiting, market sentiment warmed further. Zhipu closed at HK$1,825, up 13.35% on the day of lock-up expiration, with a total market cap of approximately HK$813.7 billion.

Affected by shareholder structure and lock-up proportions, MiniMax's stock price saw greater volatility. On July 8, it closed nearly 12% higher; by July 9, it briefly fell over 20% during intraday trading, hitting a near six-month low. However, overall, no major earthquake occurred.

The lock-up expirations for Zhipu and MiniMax were no surprise, as both companies were no longer at their peak valuations on the day of expiration. Before the official lock-up date, the market had already "rehearsed" for this liquidity release through consecutive stock price declines.

When the lock-up finally arrived, the market realized that those who intended to sell had likely already done so, and panic had largely been digested.

More importantly, the rapidly rising ARR curves injected excitement into the market, with both Zhipu and MiniMax setting ambitious targets of $1 billion in ARR for the year.

Why No Major Decline?

The temporary smooth landing of Zhipu and MiniMax can be attributed to at least three factors: market sentiment, lock-up strategies, and commercialization curves.

The underlying market sentiment is vastly different. During the AI 1.0 era, lock-up expirations coincided with a global retreat in AI investment enthusiasm, labeling the entire sector as a "bubble burst." The market voted with its feet, and panic spread like an avalanche.

In contrast, by 2026, large models have become a globally recognized sector, with optimism surrounding AI's future development. Even though both companies' stock prices had significantly corrected before the lock-up, panic remained largely at the expectation level rather than a collapse in sentiment. The market's long-term confidence in the AI sector remained unshaken, providing a psychological foundation for buyers to step in.

Differences in lock-up strategy design determined the intensity of selling pressure. The lock-up crashes during the AI 1.0 era largely stemmed from being caught off guard. A flood of shares hit the market simultaneously, with no preparation from buyers. This round was starkly different, as panic had already been repeatedly priced in.

In the 10 trading days before Zhipu's lock-up expiration, its stock price retreated about 39% from its peak; MiniMax saw an even steeper decline, falling over 70% from its March high of HK$1,330, with its market cap evaporating by nearly HK$300 billion by the eve of the lock-up.

When the lock-up day finally arrived, the market breathed a sigh of relief. The market fears bad news but hates uncertainty even more; once the greatest uncertainty was removed, there was ample reason for contrarian trading.

During the AI 1.0 era, up to 70% of shares were unlocked at once, instantly disrupting market supply and demand. In contrast, Zhipu's first batch of unlocked shares was only 5.76%, with large-scale unlocks not expected until next year, limiting and making the expansion predictable.

The shareholder structures of both companies provided crucial support. Nearly 70% of Zhipu's cornerstone investors were state-backed, with clear lock-up commitments; MiniMax's strategic shareholders, including Alibaba and miHoYo, collectively held over 20% of shares and publicly stated no intention to sell. The proportion of financial investors was far lower than that of the "Four Little Dragons" of the past, naturally reducing selling pressure. Lock-ups were not a sudden flood but a steady stream, giving the market ample time and space to digest.

Most importantly, commercialization curves have begun to emerge, with clear benchmarks—Anthropic. During the AI 1.0 era, companies facing lock-up expirations had unproven business models and shrinking revenues, leaving the market skeptical of their value. In contrast, while this round's large model companies have yet to turn a profit, their ARR growth curves are clearly visible.

Zhipu disclosed in its March earnings call that API calls had surpassed $250 million in ARR by March, aiming to reach $1 billion by year-end. Some external investors estimate its year-end ARR could be between $1.5 billion and $3 billion.

MiniMax revealed that its ARR exceeded $400 million in May and expressed confidence in achieving $1 billion in ARR by year-end.

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

History teaches us that the stock market never abandons a company for being unprofitable but for lacking visibility into its future. Zhipu and MiniMax landed smoothly because the market believes that while their growth curves are still climbing, the direction is becoming clearer.

From valuation to pricing, the evolution of this value system may be a more noteworthy signal than the lock-up expiration itself.

Diverging Valuations

Zhipu and MiniMax make an interesting comparison, summarized perfectly by the phrase "thirty years river east, thirty years river west."

At the time of their listings, MiniMax was the "east river" protagonist. On its debut day on the Hong Kong Stock Exchange, MiniMax surged 109%, overshadowing Zhipu, which listed on the same day, with its market cap exceeding HK$100 billion.

Zhipu was labeled as a B2B company, criticized for its outdated approach of pursuing projects and clients, far less glamorous than MiniMax's grip on the overseas C-end market. New stories offer more room for imagination, naturally commanding higher valuations. For a time, MiniMax overshadowed Zhipu.

This reversal of fortune stems from the starkly different paths the two companies chose.

Zhipu pursued a "focused" strategy, betting heavily on code generation and intelligent agents in early 2025. It remained in a catching-up phase until the June 2026 "supply cutoff" incident involving Anthropic propelled GLM-5.2 to the forefront of domestic alternatives.

In hindsight, the "lobster craze" was false (false/short-lived), with domestic Claw apps only capturing partial C-end traffic. Relying on subscription fees for commercialization was unsustainable. In contrast, while Zhipu was less vocal during the lobster craze than MiniMax and Kimi, it precisely captured the AI Coding boom.

Compared to lobsters, AI Coding is more exclusive. Once development teams become accustomed to using GLM for coding, interface tuning, and running agents, migration costs create barriers. The deep Binding (binding) of 250,000 developers allowed Zhipu to capture not just traffic but "workflows."

This exclusivity gave Zhipu pricing power. GLM-5.2's price hike did not drive away clients but increased call volume by 400%. For developers, the time and effort saved by using GLM-5.2 far outweighed the few extra dollars per million tokens.

This allowed Zhipu to form a closed loop quickly: price hikes boosted revenue expectations, driving valuation gains and attracting more capital to fund next-generation model development. Under the principle of "rarity commands value," further model price hikes still found buyers.

MiniMax opted for a more diversified route, adhering to a native multimodal strategy while expanding into video generation, C-end applications, and vertical industries.

Currently, the market shows its ruthless side: only two types of large model vendors can break through: top SOTA vendors, locking in premiums through absolute technological leadership, or ultra-low-cost vendors, crushing competitors with cost advantages. Both find their survival logic—the former eats profits, the latter eats market share.

However, MiniMax's M3 sits awkwardly in the middle. Its coding capabilities lag behind Zhipu, and its cost-effectiveness is inferior to DeepSeek. In video generation, it is overshadowed by ByteDance's Seedance, struggling to compete in either SOTA level or iteration speed. Neither the best nor the cheapest, this "straddling" position has pushed MiniMax into a difficult squeeze.

Compounding the issue is MiniMax's C-end DNA, which naturally traps it in a price-sensitive market. Users come and go quickly, with loyalty far lower than B-end clients bound to workflows. A price hike attempt faced market resistance, forcing a rollback to stabilize prices. The lack of pricing power directly depresses revenue expectations and valuations.

JPMorgan summarizes this divergence as a winner-takes-all phenomenon: the stronger a model's capabilities, the more it can convert open distribution into paid monetization; those with undifferentiated models face faster price comparison competition and traffic diversion risks.

The market offers no survival space for the middle ground—perhaps the harshest reality in today's large model industry.

Next Stop: $1 Billion ARR

Upon closer inspection, large model vendors are marching toward $1 billion in ARR by year-end.

What does $1 billion in ARR represent? Roughly Anthropic's level in early 2025, but more importantly, the growth trajectory behind it.

Anthropic achieved a staggering leap in ARR from $9 billion to $30 billion in Q1 2026, driven by Claude Code's breakthrough in coding. In its revenue structure, the vast majority of ARR comes from usage-based API services, contributing 75%-85% of revenue, while subscription income accounts for only 15% of total ARR.

Zhipu's positioning in B-end API services aligns perfectly with this commercial logic. Domestically, Zhipu has locked in internet giants like Tencent and ByteDance, whose high token consumption makes them ideal sales channels. Higher usage directly translates to more frequent API calls and revenue growth.

Internationally, its partnership with Amazon Web Services further opens global distribution channels. Additionally, local deployment demands from state-owned enterprises and central enterprises provide another layer of stable, ongoing consumption. This three-tiered client structure ensures Zhipu's API business is not only large but also stable in quality. When combined with Anthropic's growth narrative, capital markets are more than willing to grant higher valuation premiums.

Anthropic's core growth driver was finding a killer application scenario, which hinges on three key elements: a scenario with massive demand and extremely high value, ultimate (ultimate) commercialization efficiency, and a flywheel effect from high revenue.

For Chinese large model companies to narrow the gap with overseas counterparts, the key lies in finding their own "Claude Code moment."

However, Anthropic's AI Coding may not be the standard answer. At least from MiniMax's perspective, code generation is just the first step in AI commercialization. The real opportunities lie in high-value, knowledge-intensive vertical industries like finance, law, consulting, healthcare, and security—with coding being merely one vertical among many.

If AI Coding enhances programmers' efficiency, vertical industry AI reconstructs how professionals make decisions. Financial analysts, legal advisors, doctors, and consultants are far more sensitive to productivity gains than average C-end users. Once AI is embedded in their core workflows, migration costs and willingness to pay reach another order of magnitude.

In video generation, Seedance has already validated market explosiveness and achieved a commercial closed loop . Content creation scenarios are far from reaching their ceiling. One notable data point: Kunlun Wanwei's AI short drama platform business exceeded $700 million in ARR in Q2 2026, while its AI tools business surpassed $100 million in ARR.

Thus, the "Claude Code moment" may not be a fixed answer for domestic large model vendors but an open-ended question.

While directions differ, the proposition remains the same: find the killer application scenario that can shift ARR from linear to exponential growth, then pull the trigger with full force.