The Hong Kong AI stock sector underwent a concentrated stress test in July 2026, with KNOWLEDGE ATLAS (02513.HK) and MINIMAX-WP (00100.HK) exhibiting completely divergent price movements following the lifting of post-IPO lock-up restrictions.
On July 8th, KNOWLEDGE ATLAS saw the lock-up expire for its cornerstone investors, involving approximately 25.68 million shares, or about 5.76% of the total share capital. While such expiries typically trigger sell-off expectations, the company's stock price rose by about 15% that day. Subsequently, the company announced a placement to raise around HK$31.4 billion net for AI R&D and computing power expansion.
In stark contrast, MINIMAX-WP faced a much more challenging situation. On July 9th, a concentrated lock-up expiration of roughly 44.85% of its shares caused its free float to surge dramatically, leading to a steep 18% drop in its share price. Since its January IPO at HK$165, the stock had soared to a high of HK$1,330 within three months, only to see its market capitalization fall sharply from its peak.
Both companies swiftly initiated new fundraising rounds. MINIMAX-WP raised approximately HK$15.96 billion, with about 80% earmarked for strengthening AI infrastructure and model development. While the stated capital uses are similar, the secondary market's reaction was markedly different.
Divergent Market Narratives
The surface-level reason for the differing reactions could be the scale of the lock-up expiries. However, on a narrative level, KNOWLEDGE ATLAS is perceived as a more "native" model company, associated with keywords like domestic foundational models, MaaS, coding, and enterprise clients. Following the release of its GLM-5.2 model, it has even been dubbed a Chinese version of Anthropic.
Conversely, MINIMAX-WP, due to its early impressive growth in consumer-facing products and overseas revenue, was more easily framed within the "productization and user-growth driven" internet company paradigm. Notably, Alibaba is its largest external institutional shareholder with a roughly 13.7% stake, potentially reinforcing this early "internet" label.
Core Business Logic Differences
The core logic of internet platforms revolves around traffic and network effects, converting user scale into revenue. The logic for large language model companies is different; they sell "intelligent services," with revenue stemming from API calls, enterprise MaaS, subscriptions, and token consumption.
Internet platforms pursue economies of scale post-traffic monopoly, while model companies pursue high-value calls based on model capability. The market's focus has shifted towards heavier, more difficult hard tech areas.
Since early 2026, as models like Anthropic demonstrated rapid performance gains in coding scenarios, global AI company valuations have increasingly centered on metrics like Annual Recurring Revenue (ARR), coding capability, token consumption, and model performance.
Current Valuation Disparity and Positioning
Despite both companies' founders recently emphasizing a long-term AGI vision in internal letters, this does not explain their current valuation gap. As of July 14th's close, KNOWLEDGE ATLAS had a market cap of around HK$745 billion, while MINIMAX-WP was valued at approximately HK$72.14 billion—a tenfold difference.
KNOWLEDGE ATLAS's story is relatively straightforward: a To-B and To-G model company originating from Tsinghua University research, with revenue primarily from localized deployments. The market naturally prices it around themes of domestic self-reliance, foundational model capability, and MaaS.
MINIMAX-WP presents a more complex picture. In 2025, over 67% of its revenue came from AI-native products, with international markets contributing over 70% of total income. These figures paint a picture of a global consumer AI platform company rather than a domestic enterprise model delivery firm.
The Imperative for a Narrative Shift
Following Anthropic's rise, the market realized that an AI company's ceiling is determined not by a single product but by underlying model capability, inference costs, long-context handling, multimodality, agents, coding, developer ecosystems, and MaaS revenue.
KNOWLEDGE ATLAS naturally fits this current narrative. Its founder emphasized that real commercial opportunities lie in leaps in "intelligence upper bounds." Notably, the two companies' formulas for AGI value differ: MINIMAX-WP emphasizes "intelligence density x token throughput capacity," while KNOWLEDGE ATLAS focuses on "intelligence upper bound x token consumption scale."
Shareholder Structure's Influence
Beyond business models, shareholder structure shapes market perception. KNOWLEDGE ATLAS has a relatively dispersed ownership with various industry capital investors. MINIMAX-WP's close ties with Alibaba—as a major shareholder, key financier, and computing power supplier—have, at times, led the market to view it as part of Alibaba's AI ecosystem.
While such relationships are advantageous during the growth phase, providing capital, cloud resources, and industry backing, they can become a double-edged sword in the public markets, especially during lock-up expiries, as the market has not consistently rewarded the AI narratives of major Chinese internet conglomerates.
Three Angles to Gauge MiniMax's Narrative Transition
MINIMAX-WP needs to execute a narrative shift: from a "productization, user-growth driven AI company" to a "platform company centered on model capability and token consumption." Evidence of this shift may appear in upcoming financial reports and can be observed from three key dimensions.
First, whether its consumer products can serve as a data flywheel for its models. With 236 million cumulative users and over 70% overseas revenue, these must generate high-frequency intelligent calls to be considered model commercialization assets, not just internet traffic.
Second, its open platform and enterprise services must maintain rapid growth. With ARR growing significantly in early 2026, achieving a stated year-end ARR target would bring its valuation metrics closer to industry anchors like Anthropic.
Third, model efficiency must translate into commercial gross margins. The improvement in its gross margin from 12.2% in 2024 to 25.4% in 2025 indicates progress in managing inference costs and infrastructure efficiency, a trend that must continue.
Comments