MiniMax's 90% Collapse Exposes the Brutal Economics of China's AI Model Wars

As Hong Kong's dual large-model listings face a reckoning, model iteration velocity—not revenue growth—is emerging as the decisive valuation driver in China's crowded generative AI landscape.

Shares of MiniMax fell below their IPO price for the first time on October 22, 2026, closing down 6.52% at HK$160.70 and slashing the company's market capitalization to HK$56.6 billion (US$7.3 billion)—a staggering 86% retreat from the HK$410 billion peak reached just seven months prior. The breach of the HK$165 IPO floor crystallizes a broader repricing of Hong Kong-listed AI pure-plays and raises pointed questions about which Chinese large language model developers can convert benchmark scores into durable business moats.

The selloff is not idiosyncratic. ZhipuAI, MiniMax's closest peer among Hong Kong's two publicly traded foundation model companies, dropped 7.10% on the same session to HK$460.80, bringing its cumulative decline from a June 2026 peak of HK$2,980 to approximately 85%. Yet a critical divergence persists: ZhipuAI's market cap of roughly HK$225 billion remains four times that of MiniMax, and its shares still trade nearly 3x above their IPO price—a gap that encodes the market's verdict on model quality.

Valuation Compression Signals a Structural Repricing, Not a Sentiment Blip

The numbers that defined the frenzy are instructive. At MiniMax's March 2026 high, the stock traded at a trailing price-to-sales multiple exceeding 600x. ZhipuAI's June peak implied a P/S ratio above 1,000x. Both figures were untethered from any conventional framework for software or platform businesses, reflecting instead a pure optionality bet on artificial general intelligence timelines.

That optionality premium is now being systematically stripped out. Investors are rotating away from narrative multiples toward a tighter lens: annualized recurring revenue, model iteration cadence, and the credibility of the path to breakeven. On all three dimensions, the gap between MiniMax and ZhipuAI has widened materially through the second half of 2026.

MiniMax's ARR stands at US$8 billion as of the latest disclosed figure. ZhipuAI management stated at its interim results briefing that ARR had reached US$16 billion as of end-August 2026—precisely double. For a sector where scale advantages in inference economics compound rapidly, a 2x ARR gap is not a rounding error.

Model Iteration Velocity Becomes the Decisive Competitive Variable

The core strategic fault line between the two companies lies in how quickly each can ship frontier-capable models. This is not merely a product roadmap question—it directly determines a company's ability to retain enterprise customers who benchmark providers continuously and switch with low friction.

MiniMax launched its flagship MiniMax M3 in June 2026, with the company claiming competitive performance on SWE-Bench Pro against GPT-5.5 and Gemini 3.1 Pro, placing it near Claude Opus 4.7 on coding benchmarks. That launch generated meaningful investor enthusiasm. The problem: in the four months since, MiniMax has not released a successor flagship. The only update, M3.1 Flash released October 19, 2026, scored marginally above ZhipuAI's GLM-5.3 Flash on the Artificial Analysis leaderboard but below ZhipuAI's flagship GLM-5.3. Two next-generation models—M3.1 and M3 Pro, the latter reportedly scaling to approximately 3 trillion parameters with expanded reinforcement learning—remain on an unspecified timeline.

ZhipuAI, by contrast, shipped two major model generations within a ten-week window: GLM-5.2 in June and GLM-5.3 in August 2026. At launch, GLM-5.3 scored 60 points on Artificial Analysis, tying Kimi K3 for top position among open-source models. The company's GLM-5.3 Flash variant—branded "Nulai"—generated additional downstream distribution momentum.

A July 2026 J.P. Morgan research note articulated the strategic logic succinctly: open-weight distribution expands a model's reach but monetization remains highly dependent on model quality, iteration speed, and endpoint reliability. For differentiated leaders, open weights enable scale through cloud providers, inference platforms, and private deployment. For undifferentiated models, broader access accelerates price comparison and traffic substitution. MiniMax's four-month iteration gap has, in effect, moved it closer to the undifferentiated bucket in the market's assessment.

MiniMax's B2B Pivot Produces Revenue Growth but Widens Losses

MiniMax's financial trajectory through H1 2026 tells a story of genuine commercial traction undercut by the cost structure required to sustain it. Total revenue grew 283.1% year-on-year to US$117 million (approximately RMB 778 million / US$108 million at reference rate). The growth engine was the open platform and enterprise services segment, which surged 703.1% to US$74 million, lifting its share of total revenue from 30.3% to 63.4%—a structural shift from consumer app dependency toward B2B infrastructure.

However, adjusted net losses widened 52.7% year-on-year to US$293 million (approximately RMB 1.96 billion). The loss-to-revenue ratio remains deeply negative, reflecting the capital intensity of training next-generation parameter-scale models while simultaneously building enterprise sales infrastructure. Until MiniMax demonstrates a credible inflection toward loss narrowing—either through revenue acceleration or training cost efficiency—the widening loss profile will continue to suppress the multiple investors are willing to assign.

Founder's Capital Commitment Tests Retention Strategy Amid Wealth Erosion

MiniMax founder Yan Junjie, 37, holds 79.1025 million shares representing approximately 22.65% of the company. At the March 2026 peak, his stake was valued above HK$100 billion. It now stands at approximately HK$12.7 billion—a destruction of roughly HK$92.5 billion in personal wealth over seven months.

In July 2026, Yan published an internal letter announcing he would forgo all salary until the company achieves AGI. More operationally significant: he pledged to allocate shares equivalent to 4% of total equity over four years as long-term employee incentives, and a further 1% to support open-source community development. The gestures signal conviction in the long-term thesis, but they also reflect the competitive reality that retaining top AI research talent in China's overheated talent market requires equity compensation that cash-burning startups cannot substitute with salary.

Multimodal Assets Offer a Differentiation Anchor, But the Window Is Narrow

MiniMax retains one meaningful competitive asset that pure-text model metrics do not capture: multimodal capability built on sustained consumer product investment. In July 2026, the company released MiniMax H3, an open-source video generation model positioned on price-performance efficiency. Within three weeks of launch, H3 accumulated 24 million downloads, establishing it as a primary competitor to ByteDance's Seedance 2.0.

Video generation is a capital-intensive capability that few Chinese AI labs have productized at scale. If MiniMax can translate H3's distribution momentum into enterprise video API revenue and bundle it with its text model offerings, it creates a differentiated stack that ZhipuAI does not yet match. The strategic question is whether MiniMax can execute that bundling strategy before its text model gap relative to ZhipuAI widens further—and before better-capitalized competitors such as Baidu or Tencent-backed startups commoditize video generation entirely.

What the Divergence Reveals About China's Foundation Model Endgame

The MiniMax-ZhipuAI valuation split is a live experiment in how public markets price AI infrastructure businesses in the absence of profitability. The market's current answer: iteration velocity and ARR scale matter more than consumer app download counts or single-point benchmark victories.

For institutional investors assessing the Hong Kong AI model complex, the near-term catalyst framework is relatively clear. MiniMax needs to ship M3.1 or M3 Pro with benchmark results that close the gap against GLM-5.3 and demonstrate that its enterprise ARR can accelerate toward ZhipuAI's level. Absent those catalysts, the stock's breach of IPO price is unlikely to be the floor—it is more probably a waypoint in an ongoing multiple compression cycle.

ZhipuAI, meanwhile, faces its own test: whether a 1,000x P/S compression to current levels has created genuine value or merely a less extreme version of the same structural problem. At HK$225 billion and a trailing P/S that remains elevated by any conventional software benchmark, the company still requires the market to price in substantial future revenue growth and eventual margin expansion that no Chinese LLM developer has yet demonstrated at scale.

The AI model sector in China is consolidating around a harder-edged question than it faced eighteen months ago: not whether large language models will matter, but which specific companies can build the iteration discipline, enterprise sales infrastructure, and cost efficiency to survive a competitive cycle that now includes both domestic rivals and increasingly capable open-weight global models.

Related Coverage:

MiniMax M3.1 Flash Sparks China’s Budget AI Rivalry

MiniMax Triples Alibaba Cloud Spending Cap to $1.2 Billion as AI Compute Demand Surges