Token going global may be the next hot trend! Kimi + DeepSeek + Zhipu + MiniMax focus on concentrated chip concept stocks

AI large model services are facing a severe computational power challenge.

By 2026, competition in the AI large model industry will enter a fierce stage. Before the Spring Festival, major companies like Tencent Yuanbao, Alibaba Qianwen, and ByteDance Doubao concentrated their efforts on the C-end market, launching various welfare activities to spark a “battle for users.” Since the holiday, financing news in the global large model track has also been frequent, with capital enthusiasm continuing to rise.

However, with the rapid increase in user scale and usage frequency, token consumption has grown significantly, and large model services are facing a serious computational power test.

During the Spring Festival, Doubao experienced computational pressure due to high-frequency interactions during the CCTV New Year’s Gala, temporarily suspending video call functions; companies like Zhipu AI also implemented API rate limiting measures due to high concurrency requests. Insufficient computing power supply and strained inference resources have become urgent core issues in the large-scale deployment of large models.

Domestic Large Models Dominate OpenRouter

Since DeepSeek’s emergence at the beginning of 2025, stunning the world, China’s large model industry has rapidly risen, continuously dominating globally with unique technology and scene advantages.

According to OpenRouter, the world’s largest large model API aggregation platform, by February, the top ten models on the platform had consumed over 27 trillion tokens, with domestic models contributing over 14 trillion, accounting for more than 50% of the market share.

Among them, MiniMax M2.5 led globally with over 5 trillion tokens consumed; Kimi K2.5 followed closely, with over 4 trillion tokens in February; DeepSeek V3.2 and GLM 5 ranked fourth and eighth respectively. These leading domestic large models generally focus on improving code capabilities and autonomous agent (Agent) automation tasks.

From the consumption during the Spring Festival period, from February 16 to 22 (Lunar New Year’s Eve to the sixth day of the first lunar month), the top ten models on OpenRouter consumed over 60% of tokens from Chinese large models. Notably, MiniMax M2.5 and GLM-5 saw significant month-over-month increases of 197% and 158%, respectively.

Globally, the token consumption growth rate of overseas models from companies like Google, xAI, and OpenAI is significantly lower than that of domestic models, with some even experiencing negative growth, indicating sluggish global market share growth.

Token Consumption

May Enter a Multi-Year Growth Mode

Tokens are the smallest units of information processed by AI. Each time a large model generates a token, the backend servers must operate at high speed, requiring not only powerful computing power but also substantial electricity. Industry estimates suggest that over 70% of the cost structure for token generation is electricity and computing power.

A JPMorgan report indicates that, based on user scenarios and application adoption, token consumption may enter a multi-year growth period. China’s AI inference token consumption is expected to grow from over 10 quintillion in 2025 to approximately 3,900 quintillion by 2030, an increase of nearly 369 times. This is driven by two main factors: first, AI becoming the default interface for search, content generation, customer service, and productivity, increasing penetration in both consumer and enterprise workloads; second, use cases expanding from conversational AI to autonomous agents and multimodal outputs (longer context, more tool calls, richer outputs). Even if user growth slows, this will practically increase tokens per session.

In specific fields, the proportion of inference demand is expected to change significantly over time. Conversational AI is projected to account for nearly half of total token consumption in 2025 but gradually decrease to high single digits by 2030; AI agents in life scenarios are expected to increase from single digits in 2025 to 10-20% by 2030.

Domestic Large Models Have

Dual Advantages of Power and Electricity

For developers, cost-effectiveness is key. From China, both electricity and computing power have significant cost advantages and are expected to reshape AI pricing power.

Regarding electricity, data from the National Energy Administration shows that by the end of 2025, China’s total installed power capacity reached 389 GW, a 16.1% increase year-over-year, accounting for one-third of global generation. China also became the first country worldwide to surpass 10 trillion kWh in total electricity consumption, with ample supply and cost advantages. Under the AI boom, US electricity prices have risen noticeably, further highlighting China’s competitive electricity costs.

In terms of computing power, as early as October 2025, CCTV reported that during the 14th Five-Year Plan, China launched a new round of digital infrastructure deployment, with a nationwide integrated computing network connecting east, west, south, and north. The total computing capacity has risen to the second largest globally, providing a new foundation for China’s digital economy growth and creating new opportunities.

Currently, many leading domestic computing power companies have established complete adaptation systems, effectively breaking overseas technical monopolies in inference. Key hardware components like servers, optical modules, and cooling systems are largely domestically controlled. Relying on large-scale deployment, low-cost electricity, and continuous algorithm iteration, domestic solutions have a clear cost advantage in token production for mainstream applications.

Industry experts generally believe that domestic large models, with more efficient architecture design, are gradually reducing reliance on high-end GPUs; widespread adoption by global developers further drives explosive growth in domestic computing power demand.

14 Major Concept Stocks Focused on Large Models

Although recent trading days have shown significant divergence in the computing power sector, it is undeniable that increasing computing demand is an inevitable trend. In weekly statistics, domestic large model token consumption once surpassed the US (February 9-15, 2026), with continued advantages expanding. Supported by multiple favorable factors, token exports are expected to become a trend.

According to Wind, Tonghuashun, and other public sources, over 140 concept companies in the A-share market are related to leading large models like Zhipu, Kimi, DeepSeek, and MiniMax.

From market performance, as of February 27, these over 140 concept stocks have an average increase of over 13% this year. Jiechuang Intelligent, BlueFocus, and Zhongkong Technology have all gained more than 60% since the start of the year, with Jiechuang and BlueFocus related to Zhipu concepts, and Zhongkong Technology related to DeepSeek. Capital Online, Jinhong Modern, and UCloud-W have each increased over 40%, all involved in Zhipu concepts.

Regarding shareholding changes, 14 concept stocks have seen a decrease in the number of shareholders since the end of 2025, with three stocks experiencing declines of over 10%, including Century Tianhong, Hengwei Technology, and YuanGuang Software.

Century Tianhong’s latest shareholder count has dropped nearly 24% from the end of last year. The company, based on its strategic positioning and resource endowment, has launched AI intelligent agents like “Xiao Hong Assistant” for teachers, based on large language models, and invested in AI-assisted writing products like “BiShen Composition,” focusing on primary and secondary school scenarios.

Hengwei Technology’s shareholder count has decreased over 21%. In February 2025, the company announced its launch of the Ascend and DeepSeek integrated machines, which are divided into training-inference all-in-one and inference all-in-one types, with core advantages of “ready to use + high concurrency low latency,” covering finance, government, scientific research, healthcare, and other vertical industries, helping industries achieve intelligent upgrades.

YuanGuang Software’s shareholder count has decreased by 10.6% since the end of last year. Currently, the company mainly integrates or adapts large models like Zhipu, Alibaba Qianwen, DeepSeek, and Pangu.

These 14 concept stocks have an average increase of over 9% this year, with Hengwei Technology, Century Tianhong, and Zhide Mai leading with over 15%. Taiji Shares, Inspur, and Yuxin Technology performed weaker, with stock prices declining this year; among them, Inspur leads globally in server market share and involves DeepSeek and MiniMax concepts.

(Article source: Securities Times)