Commercial Value of Some AI Applications Emerges; Is the Domestic Large Model Industry Entering a "Profitability Tipping Point"?

From the explosion of API demand at the model layer for Zhipu and MiniMax, to the rapid rise in the proportion of generative AI for SenseTime, and further to Xunce and Mininglamp Technology validating their value in vertical data services and enterprise-level scenarios respectively, differentiation and acceleration constitute the most definitive narrative main line for the AI sector in 2026.

Since May, the AI sector has continued to heat up in the capital market.

Recently, the National Development and Reform Commission (NDRC) held its May press conference, explicitly guiding domestic large models to intensify their efforts in adapting to domestic computing chips, and promoting the deep integration of artificial intelligence with various industries across the economy and society. The Hang Seng Tech Index also announced the inclusion of large model-related enterprises, Zhipu and MINIMAX.

Brokerages Analyze "Two Main Lines"; Commercial Value of AI Applications Emerges

Behind the continuous explosion of the AI sector, on May 22, 2026, DeepSeek announced that the API price for its V4-Pro model would be permanently reduced to just RMB 0.025 per million input tokens. This represents a direct 75% discount compared to its originally planned price after reverting to the standard rate, making it the lowest-priced provider among domestic and global large model vendors.

As things stand, on one side, model vendors are actively slashing prices to enter the deep-water zone of the "token economy," while on the other side, the strategic horn at the national level has been blown. The domestic AI industry is undergoing a systematic paradigm shift from "stacking computing power" to "competing in implementation."

CITIC Securities (China Securities) stated in a recent research report that the core investment logic of the overseas AI industry presents two main lines. On the hardware and infrastructure side, the "two-phase structure" of large model inference and the popularization of Agent-orchestrated workflows are causing the marginal returns of general-purpose GPUs to decline progressively. The computing power arms race is evolving into a competition for energy and power grid elements, driving structural tilts in server architectures toward ASICs with high data reuse rates (such as Google's TPU), high-performance command-layer CPUs, and DDR5 DRAM, whose prices have skyrocketed due to rigid capacity expansion demands.

On the model and cloud ecosystem side, with the release of OpenAI's GPT-5.5, the download and installation volume of its Codex ecosystem is narrowing the gap with Anthropic's Claude Code. Computing power and cloud partners within the OpenAI camp, such as Microsoft, Oracle, and CoreWeave, are welcoming definitive narrative corrections and investment opportunities for strategic market share expansion.

The impact of DeepSeek's price cut within the industry ripples far beyond this. Previously, the API price for this model was scheduled to return to its original rate in June; the price cut means the promotional activity will be permanently extended. Behind this statement, it reflects that the pricing power of the general model layer is shifting from "technical scarcity" to "scale operational efficiency."

However, what deserves more attention than the price reduction of general tokens is that the value high grounds of the application layer and workflow layer are taking shape. Analysts at J.P. Morgan, in a China AI industry survey released on May 22, provided a core judgment: "Some AI applications are beginning to show initial signs of commercial value, particularly in vertical sectors that are workflow-intensive and data-rich."

The latest 2026 China AI Application Landscape Report published by QbitAI Think Tank shows that China's daily token invocation volume has surpassed 140 trillion, growing more than a thousand-fold within two years. The weekly invocation volume of AI large models has exceeded that of the United States for three consecutive weeks. Web-side visits have surpassed 900 million, and total downloads on the APP side have exceeded 240 million. The retention curves of leading products are shifting from "trial-and-error type" to "tool type," and the tipping point for large-scale paid conversion has emerged, while products deeply bound to workflows are becoming the key variable where the winner takes all.

Resonance of Policy and Capital; The "Money-Making Effect" Under Accelerated Industrialization

The NDRC press conference on May 22, 2026, released numerous policy signals, the most core of which was to "guide domestic large models to increase efforts in adapting to domestic computing chips."

This also means that the adaptation of the entire industrial chain—from computing chips to the model layer and then to the application layer—has been elevated to the level of national strategic coordination. The adaptation of domestic large models with domestic chips has achieved substantial progress.

At the capital level, the capital-attracting effect of the AI sector continues to amplify. In early May, Moonshot AI completed a new round of financing of approximately USD 2 billion, with its post-money valuation breaking past USD 20 billion. DeepSeek's first round of external financing has also entered its final stage, with a pre-money valuation of around USD 450 billion.

In addition, relevant large model concept stocks already listed on the Hong Kong stock market have also welcomed a tipping point in their performance since 2025. It is understood that representative Hong Kong-listed companies in the large model track, including Zhipu (2513.HK), MiniMax (0100.HK), Mininglamp Technology-W (2718.HK), SenseTime-W (0020.HK), and Xunce (3317.HK), have successively delivered their first annual reports after listing or their latest performance signals.

Through revenue growth rates, profitability tipping points, and leading indicators for 2026, it can be clearly seen that the AI sector is collectively moving from the "money-burning" phase toward a watershed of profitability tipping points.

Among them, Zhipu: In 2025, its API revenue reached RMB 190 million, a year-on-year explosion of 292.6%. As of March 2026, the ARR of its API platform had reached RMB 1.7 billion, representing a 6.4-fold increase year-to-date. The market expects its full-year revenue for 2026 to soar to approximately RMB 3.2 billion.

MiniMax: Adhering to the "full-modal self-development×ultimate cost efficiency" route, over 70% of its revenue comes from overseas markets. Its full-year revenue for 2025 reached USD 79.04 million, a year-on-year increase of 158.9%, with its gross margin leaping to 25.4%, and international revenue accounting for over 70%. In February 2026, its ARR broke through USD 150 million, an increase of over 50% compared to the last two months of last year.

Mininglamp Technology-W: It has taken the lead in turning its adjusted net profit positive, with AI-empowered operating leverage becoming the core engine of growth. Driven by the effect of AI-empowered operating leverage, the company achieved an adjusted net profit of RMB 42 million in 2025, successfully turning a profit, and its Agentic Services, which represent its new business, saw its first-year revenue break through RMB 100 million.

SenseTime-W: Its total revenue for 2025 exceeded RMB 5 billion, a sharp year-on-year increase of 33%. Revenue from its generative AI business reached RMB 3.63 billion, up 51% year-on-year, with its proportion soaring to 72.4%. In the second half of the year, its EBITDA turned positive for the first time to RMB 376 million, and the company expects its adjusted EBITDA to potentially turn a profit for the full year of 2026.

Xunce: Serving as a vertical token data hub, it has secured a differentiated position based on scarce industry data barriers. Its total revenue for the full year of 2025 was as high as RMB 1.285 billion, doubling with a year-on-year growth of 103.28%, and its customer ARPU leaped from RMB 2.72 million to RMB 5.59 million. The profitability tipping point is drawing near: it achieved an adjusted net profit of RMB 50.13 million in the second half of the year, and its adjusted net loss for the full year narrowed significantly by 33.4%. In April, its token invocation ARR exploded by 300% quarter-on-quarter, with the goal for 2026 being to achieve full-year profitability, as the token-based payment model enters a period of accelerated volume release.

Taken together, the 2025 financial reports and the latest 2026 signals of the five listed AI Hong Kong enterprises jointly point to a clear direction: the AI industry as a whole is transitioning from its past high-intensity "money-burning" expansion phase into a new stage of performance realization characterized by approaching profitability tipping points and accelerated differentiation in implementation.

From the explosion of API demand at the model layer for Zhipu and MiniMax, to the rapid rise in the proportion of generative AI for SenseTime, and further to Xunce and Mininglamp Technology validating their value in vertical data services and enterprise-level scenarios respectively, differentiation and acceleration constitute the most definitive narrative main line for the AI sector in 2026.

Hong Kong-listed Large Model Concept Stocks:

KNOWLEDGE ATLAS (02513.HK)

Knowledge Atlas Technology Joint Stock Company Limited (02513.HK; commercially known as Zhipu AI) is a leading player in China's independent large language model (LLM) sector (recognized for its capabilities and market position by Frost & Sullivan). The company has released its next-generation flagship model, GLM-5, achieving open-source State-of-the-Art (SOTA) performance in coding and agentic capabilities. It has also open-sourced the multimodal image generation model GLM-Image in collaboration with Huawei. Focusing on novel model architecture design, generalized reinforcement learning paradigms, and autonomous model evolution, its business layout closely aligns with the trend of enterprise-level AI productivity transformation.

Mininglamp Technology-W (02718.HK)

As a large model concept stock, Mininglamp Technology is recognized by the market as the "first Agentic AI stock on the Hong Kong Stock Exchange." Its self-developed model portfolio includes the DeepMiner large model product line (the underlying engine), models such as Mano/Cito, and the Cider inference acceleration framework. All these are interconnected via the Octo platform layer, which serves as a central hub for human-Agent collaboration. Ultimately, they are delivered in the form of Agentic Services, empowering the implementation of decision-making AI agents in industries such as marketing and mass consumer goods.Rather than pursuing the "Scaling Up" path of monolithic large models, the company adopts a "Scaling Out" approach through the collaboration of multiple specialized small models, achieving accuracy that surpasses general-purpose models in vertical scenarios. Its core competitive barriers lie in niche scenario data, specialized models, and continuous learning. Furthermore, the company has built an open-source, private-deployable, and white-box auditable Private AI infrastructure.

XUNCE (03317.HK)

A leading data infrastructure provider for financial asset management in China, recognized by the market as the "First Token Stock." The company continues to deepen the integration of LLMs into the core business workflows of financial asset management, reconstructing investment research, trading, and risk control pipelines through underlying data governance and AI technologies. Its underlying structured data cleaning capabilities closely align with the digital transformation of financial institutions and the compliant implementation trends of LLMs.

MINIMAX-W (00100.HK)

A top-tier player dedicated to omni-modal Artificial General Intelligence (AGI). The company continuously iterates its video generation model matrix and has partnered with mainstream domestic and overseas chip manufacturers and inference platforms to complete underlying adaptations. The total monthly active users (MAU) of its B2C and B2B LLM applications, along with multimodal API call volumes, maintain rapid growth, further broadening its computing power ecological layout.

SENSETIME-W (00020.HK)

Accelerating its transition towards Model-as-a-Service (MaaS) and generative AI services. The company has released and open-sourced the SenseNova U1 series, a native unified understanding and generation model, and the lightweight multimodal model SenseNova 6.7 Flash-Lite. Leveraging the computing reserves of its AI infrastructure (AIDC), it translates foundational computing power into omni-modal generation pipelines tailored for vertical industries such as automotive and healthcare.

NewTimeSpace Disclaimer: All content herein is the original work of NewTimeSpace. Any reproduction, reprinting, or use of this content in any other manner must clearly indicate the source as "NewTimeSpace". NewTimeSpace and its authorized third-party information providers strive to ensure the accuracy and reliability of the data, but do not guarantee the absolute correctness thereof. This content is for reference only and does not constitute any investment advice. All transaction risks shall be borne by the user.

×
Share to WeChat

Open WeChat, use the "Scan", and share to my Moments.