Source | Benyuan Finance

Author | Li Youshan

A noteworthy new variable in this year's "Double 11" is the debut of tokens as standardized digital commodities on e-commerce promotion shelves. Three major platforms—Tmall, JD.com, and Douyin—have all incorporated AI inference services and token packages into their promotional benefits, signaling that the "Token Economy" is transitioning from a buzzword to market reality. The explosion in token usage stems from the large-scale deployment of inference applications and intelligent agents (Agents). As the fundamental unit of information processing for large models, every time we use AI to generate content or intelligent agents to make decisions, it corresponds to real token consumption. As AI evolves from a novelty application to an infrastructure of production and daily life, tokens have transformed from an internal cost item for model manufacturers into a measurable, tradable, and circulable general means of production.

Policy-level endorsements have further accelerated the industrialization of the token economy. After the National Data Administration officially proposed the concept of the "Token Economy" in May this year, more than 20 cities, including Zhejiang, Beijing, Hubei, and Guangdong, swiftly followed suit, establishing token factories, formulating token measurement and monitoring standards, and setting trading rules. They view the token industry as a new growth pole for the digital economy. Just as the industrial era saw the establishment of power plants and the internet era witnessed the construction of data centers, the digital infrastructure capabilities of the AI era determine the efficiency and cost of large model training and inference. AI Infra has thus emerged as a key trend.

DataCanvas, which is advancing its Token Factory strategy, plans to build a 100,000 P computing power cluster called the "Training Factory" and a "Token Factory" capable of producing 10 trillion tokens daily. Its transition from AI infrastructure software to intelligent computing operations aligns perfectly with the entire rhythm of the token economy's evolution from concept to industrialization, providing an excellent case study for dissecting the business logic of token factories.

From AIFS to Token Business

According to China Insights Consultancy, DataCanvas ranks ninth among China's intelligent computing service providers and is the only "New Intelligent Computing Cloud" among the top ten, making it China's largest New Intelligent Computing Cloud (based on 2025 revenue). DataCanvas founder Fang Lei graduated from the Department of Electronic Engineering at Tsinghua University and later earned a Ph.D. in Electrical Engineering from Virginia Tech in the United States. He previously worked at Microsoft, where the Aether large-scale collaborative data analysis platform inspired him to recognize the need for a unified data science platform in the enterprise market that could enable hundreds of data engineers to work collaboratively.

Thus, in March 2013, Fang Lei returned to Beijing to establish DataCanvas, named after the ancient Chinese mathematical classic "The Nine Chapters on the Mathematical Art." For the first decade after its establishment, DataCanvas primarily focused on AIFS (AI Infrastructure Software Services), with software licensing and project delivery as its main business. This business offered stable growth and high gross margins but faced a size limit constrained by the pace of enterprise digitalization.

After ChatGPT ignited the large model wave, DataCanvas made a strategic transformation decision. Its core resources shifted from AI software to intelligent computing infrastructure and cloud services, transitioning from "selling software" to "building and operating computing power." Since 2023, the company has begun undertaking one-stop construction projects for intelligent computing centers (AIDCs). In 2024, it released its self-developed Alaya NeW OS operating system and launched intelligent computing cloud services, pioneering the "1 Degree of Computing Power (DCU)" precise measurement model, which charges for computing power based on usage, similar to utilities like water, electricity, and gas.

During the GTC 2026 keynote speech, NVIDIA CEO Jensen Huang proposed the "Token Factory Economics," prompting DataCanvas to further extend its business into token services. It introduced the Alaya Token service, which packages its proprietary computing clusters and third-party large language models into standardized API interfaces. Clients can call these interfaces to obtain inference results and pay based on the actual number of tokens consumed.

On September 29, 2026, DataCanvas submitted its application for listing on the Hong Kong Stock Exchange, aiming to become the "First Stock of the New Intelligent Computing Cloud." According to Tianyancha's financing information, DataCanvas has completed 14 rounds of financing since its inception, raising a cumulative total of approximately 1.62 billion yuan.

Can Token Factories Be Profitable?

The China Academy of Information and Communications Technology disclosed that China's daily average token usage was approximately 100 billion at the beginning of 2024, but by June 2026, this figure had surpassed 500 trillion. Capital has flooded in, with companies like Silicon Flow, Infinite Core, and DataCanvas all rushing towards IPOs with their "token-selling" narratives.

Why does the industry still describe this business as "a sexy story but a lean balance sheet"? Taking DataCanvas as an example, from 2023 to 2025, its total revenue surged from 106 million yuan to 1.098 billion yuan, a tenfold increase over two years, with a compound annual growth rate of 221.5%. In the first half of 2026, it achieved revenue of 588 million yuan, a year-on-year increase of 92%, maintaining a high level. Behind this growth lies a shift in revenue structure. Intelligent computing infrastructure business typically involves large-scale hardware procurement and integration projects, making it easier to rapidly boost revenue.

In 2023, AIFS services contributed 76.5% of revenue, while intelligent computing infrastructure business accounted for 23.5%. By 2025, the proportion of AIFS business had plummeted to 1.7%, with intelligent computing infrastructure business becoming the main revenue pillar, contributing over 90% of revenue. In the first half of this year, the revenue shares of the two businesses were 73.8% and 0.1%, respectively.

The change in business structure caused the company's comprehensive gross margin to follow a "V-shaped" trend.

From 2023 to 2025, the data were 40.6%, 11.7%, and 17.4%, respectively. In the first half of 2026, the company's comprehensive gross margin rebounded to 31.0%, attributed to the delivery of several projects during this period that involved a significant proportion of software components, as well as the realization of economies of scale in cloud services.

Intelligent computing cloud services related to tokens have shown a positive growth trend. They accounted for only 0.1% of DataCanvas's total revenue in 2024 but increased to 7.1% in 2025 and rose to 26.1% in the first half of this year. Using debt to scale up and time to capture the market, the issue of its "computing power leverage" has also become a focal point of market attention.

DataCanvas does not invest heavily in building its own data centers or procuring GPUs. Instead, it signs long-term (3-5 year) locked-volume leases with intelligent computing center owners and computing power operators, incorporating third-party computing power resources into its own scheduling pool. It then attempts to convert computing power scale into revenue and profits through software, scheduling, and cloud services.

Computing power leverage has led to explosive growth in lease liabilities, rising from 187 million yuan at the end of 2024 to 3.087 billion yuan by the end of July 2026, an increase of more than 15 times in less than two years. This leverage has directly propelled the company's industry standing. In 2025, DataCanvas ranked ninth in China for intelligent computing service revenue and was the only independent "New Intelligent Computing Cloud" vendor among the top ten. The top three vendors held market shares of 22.1%, 15.2%, and 10.5%, respectively, while DataCanvas had a market share of 1.6%.

Under this model, project payments can partially offset the upfront investment in computing power leasing, which is also the reason for the phenomenon of "overlapping customer and supplier identities" in the prospectus. For example, the largest supplier in 2024 became the company's largest customer in 2025.

Due to its reliance on large projects and a few partners, DataCanvas also faces the issue of overly concentrated customers and suppliers.

From 2023 to 2025 and the first half of 2026, the revenue contributed by the company's top five customers accounted for 54.8%, 85.7%, 88%, and 80.8% of total revenue, respectively. On the supplier side, during the same periods, the procurement amounts from the top five suppliers accounted for 47.1%, 74.8%, 84.8%, and 74.5% of the total, respectively.

Another issue with the high-leverage model is that the company's daily operations and computing power expansion are highly dependent on external financing, and its self-generated cash flow capacity has not yet formed, as evidenced by the following:

First, persistent net outflows in operating cash flow. From 2023 to 2025 and the first half of 2026, net operating cash outflows were 251 million yuan, 38.44 million yuan, 82.43 million yuan, and 86.5 million yuan, respectively, and have not yet turned positive.

Second, tight short-term debt repayment capacity. At the end of 2025, the company's current ratio once dropped to 0.24. After completing approximately 500 million yuan in financing in the first half of 2026, its cash and cash equivalents at the end of the period were 381 million yuan, and the current ratio rebounded to 1.10, still low compared to the general reference level of 2.0.

Third, rapidly increasing interest burdens. Interest on lease liabilities surged from 360,000 yuan in 2023 to 35.88 million yuan in 2025 and reached 32.46 million yuan in the first half of 2026. At this pace, annual interest expenses could indeed exceed 60 million yuan, posing a real cash outflow pressure on a company that has not yet achieved profitability.

It is worth noting that intelligent computing cloud services are not equivalent to token sales. The revenue scale of the Alaya Token service, which is directly linked to token sales, has not been separately disclosed in the prospectus. The intelligent computing services sector offers immense imagination, but beyond high revenue, it often manifests as high losses and high investments.

From 2023 to 2025 and the first half of 2026, DataCanvas accumulated a net book loss of nearly 2.3 billion yuan over three and a half years. Excluding the fair value changes of 1.759 billion yuan in preferred shares, the adjusted net loss narrowed to 91.27 million yuan in 2025 and further decreased to 14.884 million yuan in the first half of this year, approaching break-even.

In terms of R&D expenditure, it was 112 million yuan, 103 million yuan, and 122 million yuan in 2023, 2024, and 2025, respectively, and 66.785 million yuan in the first half of 2026, with a total R&D investment of approximately 400 million yuan over three and a half years. Computing power utilization increased from 56.0% in 2025 to 79.6% in the first half of 2026.

Beyond Positions

In the past few years, independent AI Infra vendors have benefited from two rounds of "technology timing" dividends: the boom in intelligent computing center construction and the demand for domestic chip adaptation. Companies like DataCanvas have sought differentiated space by positioning themselves as "New Intelligent Computing Cloud" providers with a "neutral stance."

However, as of June 2026, there are over 500 intelligent computing centers in operation, nearly 400 under construction or planned, bringing the total to nearly 1,000, with an intelligent computing power scale of 2,185 EFLOPS (FP16).

Entrants include the three major basic telecom operators, as well as leading cloud and internet companies like Alibaba Cloud, Tencent Cloud, Huawei Cloud, Baidu Intelligent Cloud, and Volcano Engine. The trend of large-scale and centralized computing power supply is becoming increasingly evident.

Efficiency metrics reveal another side of the industry. The average GPU utilization rate in domestic intelligent computing centers is only about 30%. In 2025, the national supply of intelligent computing power was approximately 3.8 billion card-hours, but the actual "utilized computing power" was only 1.4 billion card-hours, meaning nearly two-thirds of the supply capacity was not converted into effective paid demand. The shape of a "structural bubble" is clearly visible.

As the construction boom subsides, pure networking adaptation and hardware optimization projects continue to decrease, and the space for one-time technical service fees is constantly shrinking. Independent AI Infra vendors are being forced to transition from "project-based technical service providers" to "long-term computing power operators," seeking the next growth driver.

From selling software to selling tokens, from "1,000 specialized models" to "token costs reduced by 1,000 times," DataCanvas still needs to prove how much real profit its efficiency premium can ultimately deliver.

Operations/Yu Shuya Design/Yan Weier

*All rights reserved. No reproduction without authorization.

end