Lobster heat activation accelerates the entire industry chain; domestically developed large models commercialize at a faster pace

Securities Times Reporter Chen Xiachang

The explosive popularity of the “Lobster” AI agent (the open-source AI agent framework “OpenClaw,” colloquially called “Lobster” in Chinese) has sparked a nationwide “shrimp farming” industry boom. Tech giants like Baidu, ByteDance, Tencent, and others have also launched various “Lobster” AI agents. However, as the hype gradually subsides, the commercialization prospects of the AI industry, especially large model companies, are becoming clearer. Hong Kong-listed companies such as MiniMax, Zhipu, and Dipu Technology, which are deploying large models, are expected to see a turning point in their performance.

“The ‘Lobster’ craze” activates the entire AI industry chain

“In the US, ‘Lobster’ is mostly used and discussed by professional technical personnel, and there’s no sign of a nationwide ‘shrimp farming’ phenomenon. ‘Lobster’ is more popular domestically than overseas because there is a complete industry chain supporting it,” Zhou Qi, who works at a major US tech company and has long been paying attention to China’s AI industry, told Securities Times.

Nvidia CEO Jensen Huang once broke down the AI industry into five closely connected layers: the bottom layer is energy (electricity), followed by chips, then infrastructure represented by data centers, then models—including large language models and world models—and finally various applications. Huang believes that a successful AI application will drive demand across all these layers, from the top down to the power plants at the bottom.

“AI agents like ‘Lobster’ are a type of application. Their low deployment threshold and open-source nature break down development barriers across the AI industry chain, stimulating full activation of the entire upstream, midstream, and downstream AI industry,” Zhou Qi said.

He explained that “Lobster” is an open-source model that is not tied to any specific large model, allowing users to choose from various large models. This directly stimulates a rapid influx of global developers and users, especially Chinese large model companies that focus on open source. Since Chinese large models are generally cheaper than those in the US, users tend to prefer domestic large models. This is also why domestic large model companies have quickly become popular during this “Lobster” wave. Consequently, this has directly triggered a surge in upstream computing power demand, with exponential growth in cloud service providers’ computing rentals and server orders.

Major securities firms have clear opinions on this. Huatai Securities (601688) estimates that compared to chatbots (300024), the consumption of tokens (the basic units of text processed by large language models) by AI agents could increase more than tenfold, with corresponding computing power demand growing over a hundred times. This demand shift will push inference computing power to historically surpass training power, becoming the core support for computing needs. CITIC Securities (600030) believes that the “Lobster” craze marks the transition of AI agents from concept to implementation, with computing power demand shifting from pulsed to sustained growth, becoming a long-term growth engine in the industry chain. China International Capital Corporation (CICC) also stated that the widespread adoption of “Lobster” will rapidly expand the inference computing power gap, forcing hardware upgrades and capacity expansion in computing services.

“Although China is not leading in chip computing power, domestic companies are highly competitive in the AI industry chain due to relatively low electricity prices and stable power supply,” Zhou Qi said. In fact, industry leaders like Elon Musk and Jensen Huang have expressed envy over China’s highly competitive electricity prices multiple times.

Accelerated commercialization of large models

Amid the full industry chain benefits driven by the “Lobster” craze, domestic large model companies are the most direct and core beneficiaries.

According to the latest data from OpenRouter, the world’s largest AI model API aggregation platform, from March 16 to March 22, the total calls to AI large models worldwide reached 20.4 trillion tokens, a 20.7% increase week-over-week. Among the top ten AI large models, Chinese models had a weekly call volume of 7.359 trillion tokens, up 56.9% from the previous week; US models had 3.536 trillion tokens, up 7.35%. China’s AI large models have now surpassed the US for three consecutive weeks in weekly call volume.

Specifically, the top four in global usage last week were all Chinese AI models, including Xiaomi MiMo V2 Pro, Zheyu Xingchen Step3.5 Flash (free), MiniMax M2.5, and DeepSeek-V3.2. Zhipu GLM 5 also previously ranked among the top.

“This wave has broken the long-standing dilemma of domestic large models being ‘money-burning and hard to monetize.’ Driven by three core factors—rising token consumption, exploding user base, and upgraded business models—it has accelerated the commercialization process of domestic large models, moving from the technical investment phase to the value realization phase,” Zhou Qi said.

Data also confirms this. The domestic large model MiniMax M2.5 has ranked first in global model call volume for five consecutive weeks. Yan Junjie, founder and CEO of MiniMax, disclosed at an earnings conference that by February 2026, the company’s ARR (Annual Recurring Revenue) will exceed $150 million. The “Moon Shadow” K2.5 large model was launched in January 2026. The company revealed that within less than a month of launch, nearly 20 days of accumulated revenue already exceeded the total revenue for all of 2025, mainly driven by a surge in global paid users and API calls.

This explosive growth has not only boosted the stock prices of the two listed large model companies but also significantly increased valuations of unlisted firms. Zheyu Xingchen completed a B+ round of financing exceeding 5 billion RMB in January and is expected to list in Hong Kong. The “Moon Shadow” model raised over $700 million in February and is currently in a new round of $1 billion funding, with a valuation exceeding $18 billion.

Performance may see a turning point

For large model companies listed in Hong Kong, when they will turn profitable has been a key concern for many investors. Recent earnings reports from two such companies suggest that profitability is not far off.

MiniMax, which went public in January, reported that last year the company achieved a total revenue of $79.04 million, a 158.9% increase year-over-year; gross profit reached $20.08 million, a 437.2% surge, with gross margin rising to 25.4%. The company disclosed that in February 2026, the daily token consumption of its M2 series text models increased more than sixfold compared to December 2025, and the token consumption from encoding schemes increased over tenfold. A Morgan Stanley report believes that such strong API demand provides high visibility for doubling revenue in 2026.

Zhang Renqi, Managing Director of the cornerstone investment department at MiniMax’s early investor, said, “The biggest test after listing is financial performance. The Hong Kong market values commercialization and profit levels highly. For MiniMax, future plans may include expanding hardware deployment and integrating interaction capabilities into specific hardware forms. However, the large model industry is still in its early stages, and as model capabilities continue to improve, more new applications and software forms will emerge.”

Another enterprise-level AI application solution provider, Dipu Technology, recently released its financial results, showing revenue of 415 million RMB in 2025, a 70.8% increase year-over-year, with operational profit achieved in Q4. Chairman and CEO Zhao Jiehui expressed strong optimism about the company’s profitability prospects. He stated that after deducting non-operating items, the company’s adjusted net loss was 27.54 million RMB, a 71.4% reduction year-over-year, and the company has been reducing losses significantly for four consecutive fiscal years. He expects the company to be profitable in 2026.