Breaking the "monetization difficulty" development dilemma! The "Lobster Fever" activates the entire industry chain

The sudden breakout of the “Lobster” AI agent—an open-source AI agent framework called “OpenClaw,” jokingly referred to in Chinese as “Lobster”—has sparked a nationwide industrial boom centered on “lobster farming.” Tech giants such as Baidu, ByteDance, Tencent, and others have also rolled out various “Lobster” agents. But as the heat gradually fades, the commercialization prospects of the AI industry—especially for large-model companies—have become increasingly clear. Hong Kong-listed companies such as MiniMax, Zhipu, and Diqup Technology that have been laying out large models are expected to see an earnings inflection point.

The “Lobster Rush” activates the entire AI industry chain

“In the United States, ‘lobster’ is used and discussed more by professional technical personnel, and I haven’t seen any phenomenon of people ‘farming lobsters’ on a mass scale. In China, ‘lobster’ is much more popular than overseas, and the reason is that there is an entire industry chain behind it supporting it.” Zhou Qi, who works at a tech giant in the U.S. and has long been tracking China’s AI industry development, told The Securities Times.

NVIDIA CEO Jensen Huang has previously broken down the AI industry into five closely connected layers. The lowest layer is energy (electricity), followed by chips. The middle layer is infrastructure represented by data centers. Then comes models—including large language models, world models, and others. At the top are various applications. Huang believes that a successful AI application will drive demand for all the layers beneath it—right down to the power plants.

“AI agents represented by ‘Lobster’ are a type of application. Their low-threshold deployment and open-source openness characteristics break down development barriers across different stages of the AI industry, driving comprehensive activation of the entire AI upstream, midstream, and downstream industry chain.” Zhou Qi said.

Zhou Qi explained that “Lobster” is an open-source approach that is not tied to any underlying large model. Users can choose from various large models—this directly stimulates global developers and users to enter quickly, especially China’s large-model companies that are primarily open source. Since the cost of China’s large models is lower than in the U.S., users are more inclined to use domestic large models. This is also the reason domestic large-model companies quickly became red-hot behind the current “Lobster” wave. The next step is directly igniting demand for upstream computing power: cloud service providers’ compute leasing and server ordering volumes have grown exponentially.

Major securities firms all have clear views. For example, Huatai Securities calculates that compared with chatbots, the token consumption (Token; the basic unit texts are processed by large language models) of agents may be increased by more than ten times, which would correspond to a more than one-hundredfold increase in demand for computing power. And this shift in demand will drive inference computing power to historically surpass training computing power, becoming the core support for computing power demand. Citic Securities believes that the sudden popularity of “Lobster” marks agents moving from concept to implementation; compute demand will shift from a pulse-like pattern to sustained growth, becoming the core engine for long-term growth in the computing power industry chain. CICC also said that the mass adoption of “Lobster” will quickly amplify the inference-compute shortfall, forcing upgrades in compute hardware and expansion of compute services.

“Although China is not leading in chip computing capability, domestic companies’ competitiveness in the AI industry chain is extremely strong thanks to relatively low electricity prices and stable power supply.” Zhou Qi said. In fact, in terms of energy prices, industry leaders such as Musk and Jensen Huang have expressed envy more than once about China’s highly competitive electricity prices.

Commercialization of large-model companies accelerates

Among the industry-wide chain dividend driven by the “Lobster” boom, domestic large-model companies have become the most direct and core beneficiaries.

OpenRouter, the world’s largest AI model API aggregation platform, released its latest data showing that from March 16 to March 22, the total number of global AI large-model calls reached 20.4 trillion tokens, up 20.7% month-over-month. Among the top ten AI large models on the list, the weekly calling volume of domestic AI large models was 7.359 trillion tokens, up 56.9% from the previous week; the weekly calling volume of U.S. AI large models was 3.536 trillion tokens, up 7.35% month-over-month. This means China’s weekly calling volume of AI large models has surpassed the United States for three consecutive weeks.

Looking at companies specifically, the top four by global calling volume last week were all domestic AI large models, including Xiaomi MiMo V2 Pro, Jietiao Xingchen Step3.5 Flash (free), MiniMax M2.5, and DeepSeek-V3.2. Zhipu GLM 5 also appeared on the list for a time earlier.

“This wave breaks the long-standing dilemma for domestic large-model companies—‘burning money with difficult monetization.’ Powered by three core drivers: a surge in token consumption, a rapid expansion in user scale, and upgrades to business models, it pushes the commercialization process of domestic large-model companies onto a fast track—moving officially from the technology investment period into the period of value realization.” Zhou Qi said.

The data also confirms this. Domestic large-model MiniMax M2.5 has claimed the global large-model calling volume crown for five consecutive weeks. At an earnings exchange meeting, MiniMax founder and CEO Yan Junjie disclosed that in February 2026, the company’s ARR (Annual Recurring Revenue; annual recurring revenue) exceeded $150 million. The Moonshot K2.5 model was released in January 2026. According to the company, within less than one month after launch, the company’s cumulative revenue over the past nearly 20 days has already exceeded its total revenue for all of 2025. This growth is mainly driven by a surge in paid users worldwide and a spike in API call volume.

Beyond this data surge, in addition to boosting the stock prices of two already-listed large-model companies, the valuations of companies not yet listed have also been rising steadily. Among them, Jietiao Xingchen completed over RMB 5 billion B+ round financing in January and is expected to list on the Hong Kong stock market. The Moonshot completed more than $700 million financing in February this year, and is also conducting a new round of $1 billion financing, with its valuation already exceeding $18 billion.

Earnings are expected to reach an inflection point

For large-model companies listed in Hong Kong, when they will become profitable has long been the focus of many investors. Based on two large-model companies that released performance recently, profitability is not far off.

MiniMax, which listed only in January this year, shows in its financial report that last year the company achieved total revenue of $79.04 million, up 158.9% year over year. Gross profit reached $204k, up 437.2% year over year, with gross margin rising to 25.4%, indicating a significant improvement in profitability. According to the company, in February 2026, the daily average token consumption of the M2 series text models increased by more than 6 times compared with December 2025, and the token consumption generated by the encoding scheme also increased by more than 10 times. JPMorgan’s report believes that such strong momentum in API demand provides high visibility for revenue to double in 2026.

Zhang Renqi, Managing Director of the Investment Department at cornerstone capital, an early investor in MiniMax, told reporters: “The biggest test after listing is financial performance. The Hong Kong market places great emphasis on commercialization capability and profit levels. For MiniMax, it may expand its hardware layout on the basis of its existing business and integrate interaction capabilities into specific hardware form factors. However, the large-model industry is still in an early stage; as model capabilities continue to improve, it will create more new applications and software form factors.”

Meanwhile, another company, Diqup Technology, an enterprise-level large-model AI application solutions provider, released its latest financial report showing that the company’s 2025 operating revenue was RMB 415 million, up 70.8% year over year, and it achieved operating profitability in the fourth quarter. For the company’s profitability outlook, Diqup Technology’s board chairman and CEO Zhao Jiehui is also very optimistic. At the earnings release meeting, he told investors that last year, after deducting non-operating items, the company’s adjusted net loss was RMB 27.54 million, narrowing by 71.4% year over year, and it significantly reduced losses for four consecutive accounting years. He believes the expectation for operating profitability in 2026 is clear.