China's AI Model Stocks Diverge Sharply as Lock-Up Expiries Force a Reckoning
Zhipu surges 13% while MiniMax crashes 18% on consecutive unlock days, exposing a fundamental split between B2B API models and consumer multimodal plays — and raising urgent questions about which monetization path survives the next phase of the sector's maturation.
The two-day window spanning July 8–9, 2026 delivered the first genuine stress test for China's publicly listed large language model companies, as lock-up expirations for Zhipu AI and MiniMax forced secondary markets to price these assets against real fundamentals rather than narrative momentum. The results were unambiguous: Zhipu closed July 8 up 13.35% at HK$1,825 per share, pushing its market capitalization to approximately HK$900 billion (US$125 billion); MiniMax closed July 9 down nearly 18% at HK$297.4, a level below its IPO-day closing price and representing a roughly 70% collapse from its March 2026 peak of HK$1,330.
The divergence was not accidental. Beneath the surface-level mechanics of float expansion and shareholder structure lies a deeper verdict: capital markets are rewarding AI companies that have built defensible, workflow-embedded B2B revenue, and penalizing those whose growth story rests on consumer multimodal products facing both commoditization and tightening regulation.
Unlock Mechanics Reveal Structurally Different Pressure Profiles
The scale of the two unlock events was asymmetric by design. For Zhipu, only 5.76% of total shares became freely tradeable on July 8, with the unlocked cohort comprising predominantly state-backed cornerstone investors — including JSC International Investment Fund, Taikang Life Insurance, GF Fund Management, Shanghai Gaoyi Asset Management, and WT Asset Management. In the current policy environment, these holders carry an implicit mandate to support strategic industrial development, making near-term liquidation politically and institutionally unlikely.
MiniMax faced a categorically different situation. Its July 9 unlock released 44.85% of total shares from escrow, expanding the freely tradeable float from under 3% to nearly 50% in a single session. The unlocked pool included venture capital firms — Hillhouse Capital, Sequoia China, and IDG Capital — alongside strategic investors Alibaba Group and miHoYo, plus several state-affiliated funds. While Alibaba and miHoYo publicly stated their intention to hold, the VC cohort faces fund lifecycle and IRR constraints that create genuine liquidation optionality. Critically, these early-stage investors entered at costs that make even the post-crash price a multiple-return exit.
Historical data from Hong Kong's technology sector reinforces why the market braced for impact: between 2022 and 2025, H-share tech stocks declined an average of 4% in the three months following lock-up expiration and 7% over six months, with the sharpest corrections hitting stocks that had appreciated most aggressively post-IPO.
That MiniMax's decline, while severe, did not trigger a disorderly rout reflects one mitigating factor: the stock had already shed more than 70% from peak, effectively pre-pricing much of the unlock anxiety. Markets, as the adage goes, fear uncertainty more than bad news — and by July 9, the bad news was largely known.
Business Model Divergence Drives Gross Margin Gap That Markets Cannot Ignore
The stock price split maps almost precisely onto a gross margin differential that tells the story of two fundamentally different competitive positions. For fiscal year 2025, Zhipu reported a gross margin of 41%; MiniMax recorded 25.4%, up from 12.2% the prior year but still lagging by 16 percentage points.
That gap reflects pricing power — or the lack of it. Zhipu raised prices on its general-purpose language model this year and sustained call volume growth, evidence that its enterprise client base has developed workflow dependency. By contrast, MiniMax launched its flagship M3 model at approximately double the price of its predecessor, only to announce a permanent 50% price reduction less than one week after launch — a forced retreat that signals inadequate demand elasticity and insufficient product differentiation.
The competitive positioning of each company's core model sharpens this picture. Zhipu's GLM series has maintained a position in the top tier of domestic model benchmarks while earning specific recognition from developers for performance on agentic and coding tasks. From February 2026 onward, as the market narrative rotated from multimodal video generation toward AI Agents and AI Coding — a shift accelerated globally by Anthropic's (Anthropic) outperformance and OpenAI's (OpenAI) decision to discontinue its Sora video product — Zhipu found itself holding the right hand. GLM-5.2 reportedly saw API call volume increase 400% following a price increase, a counterintuitive outcome that indicates genuine switching-cost lock-in among its approximately 250,000 active developers.
MiniMax, by contrast, bet heavily on multimodal video generation through its Hailuo AI product and on AI companionship through Talkie and Xingye. The multimodal thesis attracted a significant valuation premium at IPO — MiniMax listed at roughly twice Zhipu's market capitalization — but the narrative collapsed as the monetization challenges of standalone video AI became undeniable. OpenAI's Sora 2 achieved one million downloads within five days of its September 2025 launch before being shuttered as OpenAI rationalized loss-making verticals. ByteDance's Seedance and Kuaishou's Kling demonstrated that video generation is viable at scale only when integrated into an existing content and advertising ecosystem — a structural advantage unavailable to independent model companies.
Regulatory Headwinds Compress MiniMax's Largest Revenue Line
The pressure on MiniMax is compounded by a regulatory intervention directly targeting its highest-revenue product category. Following a widely reported incident in October 2025 in which a U.S. user died by suicide after extended interaction with Google's Gemini AI companion — prompting legal action by the family — China's regulatory authorities issued the Interim Measures for the Administration of Anthropomorphic Interactive AI Services. The rules take effect July 15, 2026.
Key provisions prohibit virtual companion services for minors, mandate enhanced content review and user privacy protections for emotional interaction products, require information sharing with regulators for platforms above defined user-scale thresholds, and impose special restrictions on behaviors deemed psychologically manipulative or designed to induce user dependency.
MiniMax's Talkie and Xingye products — which the company's prospectus identified as the single largest revenue-contributing product line through the first three quarters of 2025 — fall squarely within the regulatory perimeter. The compliance burden arrives precisely as the company is attempting to rebalance its revenue mix toward its open platform and Hailuo AI, both of which are growing rapidly but from a smaller base.
MiniMax's Overseas Footprint Offers a Partial Counterweight
The bear case on MiniMax is real but not uncontested. The company's international revenue represented 73% of total revenue in 2025, with cumulative service to more than 236 million users and 214,000 enterprise clients across more than 200 countries and territories. This global distribution — built on competitive pricing that resonates in cost-sensitive markets across Southeast Asia and the Middle East — provides both revenue diversification and partial insulation from domestic regulatory pressure.
The financial efficiency trajectory is also notable. In 2025, MiniMax grew revenue 158.9% year-over-year while simultaneously cutting marketing expenditure by 40.3%, a combination that suggests the company is moving toward a more capital-efficient growth model even as absolute losses persist. Management has committed to reaching US$1 billion in annualized recurring revenue by year-end 2026, having disclosed that ARR surpassed US$400 million in May 2026.
At current valuations — MiniMax's market cap has contracted to approximately HK$93 billion (US$12.9 billion) against Zhipu's roughly HK$900 billion (US$125 billion) — the implied price-to-ARR multiple for MiniMax is substantially compressed relative to its domestic peer. Whether that compression represents value or a value trap depends on whether the company can credibly navigate regulatory constraints, close the gross margin gap, and identify a product wedge that is neither dominated by Zhipu on capability nor by DeepSeek (DeepSeek) on cost.
The $1 Billion ARR Race Frames the Sector's Next Valuation Catalyst
Both companies have publicly committed to crossing US$1 billion in ARR by December 2026 — a threshold that carries symbolic weight as the approximate level Anthropic occupied in early 2025, before Claude Code drove an extraordinary acceleration from US$9 billion to US$30 billion ARR in the first quarter of 2026 alone. Anthropic's revenue structure — with API usage contributing 75%–85% of total ARR against a 15% subscription share — validates the B2B API model that Zhipu has systematically built.
Zhipu's disclosed ARR reached US$250 million as of March 2026, with some external investors projecting a year-end range of US$1.5 billion to US$3 billion. The company's client stack — anchored by high-volume API consumers including Tencent and ByteDance domestically, expanded internationally through an Amazon Web Services distribution partnership, and supplemented by local deployment contracts with state-owned enterprises — provides a layered revenue base with relatively low churn characteristics.
The compute cost environment, however, introduces a structural risk for all API-dependent businesses. Infrastructure provider Baseten disclosed in June 2026 that NVIDIA B200 GPU rental rates will increase approximately 94% at October renewal — from US$2.63 to US$5.10 per hour. Simultaneously, the Silicon Data LLM Token Expenditure Index has declined nearly 20% from its May 2026 peak, suggesting that token consumption growth may be plateauing even as per-unit compute costs rise. Margin compression from this cost-revenue squeeze remains a sector-wide risk that neither Zhipu nor MiniMax has fully addressed in public guidance.
A Larger Unlock Looms, Resetting the Timeline for Zhipu
Investors focused on Zhipu's July 8 performance should note that the current unlock was structurally benign by design. The far more consequential event arrives January 8, 2027, when approximately 39.99% of Zhipu's total shares — held by Meituan, Ant Group, Tencent, Sequoia China, Hillhouse, and employee stock programs — become eligible for sale. This cohort includes early-stage financial investors with substantial unrealized gains and no stated lock-up extension commitments.
The six-month window between now and that event is, in effect, the period during which Zhipu must demonstrate that its ARR trajectory, gross margin expansion, and developer ecosystem depth are sufficient to absorb a materially larger supply shock. The July 8 performance — interpreted by some market participants as a vote of confidence — buys credibility but not permanence.
As J.P. Morgan has framed the emerging dynamic in its coverage of the sector, the large language model market is exhibiting winner-takes-most characteristics: models with sufficient capability differentiation can convert open distribution into paid monetization, while undifferentiated models face accelerating price competition and audience migration. MiniMax's M3, priced between DeepSeek's cost leadership and Zhipu's coding capability, currently occupies the least defensible position in that framework.
July 8 and July 9 were the opening moves in a repricing process that will extend well into 2027. The terminal outcome — whether these valuations represent the foundation of a durable AI infrastructure cycle or the partial deflation of a speculative overhang — depends on whether China's large model companies can generate the kind of workflow-embedded, recurring enterprise revenue that transforms AI from a research narrative into a compounding business.
Related Coverage:
Zhipu AI's 10x Rally Exposes Hong Kong's AI Narrative Premium Over MiniMax