After ten years of honing its blade, Xunce has finally reached its moment of realization.Article author and source: WeChat official account "New Intelligence Yuan"
A few weeks ago, a MiHoYo engineer forgot to set a circuit breaker while testing multi-agent collaboration, causing dozens of agents to enter a loop of mutual calls.
13 hours later, the bill came in: 2 million.
Uber across the ocean didn’t fare much better. By making AI programming tools available to 5,000 engineers, they burned through their entire 2026 AI budget in just four months.
Aren't all tokens down 99%? How can they still burn like this?
Great question. Just now, DeepSeek-V4-Flash slashed the cache hit price to ¥0.02 per million tokens. Xiaomi has followed suit with a permanent 99% price reduction. Large model providers are competing fiercely.
Logically, AI should become cheaper to use the more you use it.
But once businesses truly got up and running, they discovered something counterintuitive: the cheaper the generic token, the more heavily it was used; and the more it was used, the more they realized that "cheap intelligence" alone simply wasn't enough.
A large model can chat with you about philosophy until dawn, but when it comes to a multi-million-dollar transaction, would you dare let it make the call?
It can write flashy marketing copy, but can it detect a microscopic cold solder joint on a PCB?
It can memorize medical textbooks inside and out—but would it dare to look at a shadow on a CT scan and say, “Don’t worry, it’s benign”?
The reason is very simple—these things simply aren’t available in publicly accessible internet data. Your factory’s process parameters aren’t in textbooks, and the treatment protocols of top-tier hospitals aren’t posted on forums.
The market split in two: general-purpose tokens plummeted to near-zero prices, while scenario-specific tokens tied to industry expertise soared to 10 to 100 times the value of general-purpose tokens—and continue to rise.
One name, two directions.
The truth about this business is simply this: those who profit are not the ones selling cheap tokens, but those who continuously charge fees from others' token burning processes.
There is a company that just happened to be on the upside of the surge.
An unusual curve: 288 million, 530 million, 632 million, 1.285 billion.
This is Xunce Technology (3317.HK)'s revenue over the past four years. Revenue increased by 84% in 2023, rose by only 19% in 2024, and then surged by 103% in 2025.
What’s worth considering is that 19%.
2024 was the most intense year yet in the global large model arms race, with billions of parameters, tens of thousands of GPUs, and massive funding—all poured into buying chips and training models.
A company specializing in data infrastructure barely moved that year—which was perfectly normal, since its revenue depends not on who is building AI, but on who is using AI.
Then there's the recently released interim earnings forecast—the curve hasn't slowed down; instead, it's become even steeper.
First half of 2026: Revenue reached RMB 967 million, a year-over-year increase of 389%, setting a new all-time high for the same period; attributable net profit was RMB 72.51 million, compared to an attributable net loss of RMB 89 million in the same period last year, marking the first-ever semi-annual profit; adjusted net profit was RMB 67 million, compared to a net loss of RMB 105 million in the same period last year, demonstrating a significant turnaround and the release of profitability.
A decade in the making! From continuous investment to achieving its first-ever profit in the first half of the year; from technological accumulation to validating its business model; from market skepticism to proven performance. While the entire Hong Kong-listed AI sector remains focused on valuation debates around computing power, models, and applications, this company has already pioneered a new path to AI profitability.
Did you notice? This is essentially a thermometer for measuring whether AI has truly been implemented—
In 2024, it won’t rise because it’s still in training; in the second half of 2025, it will surge dramatically because AI inference is being deployed and companies are truly starting to use AI.
In simple terms, Xunce helps businesses organize scattered data into a format that AI can use.
The hospital has tens of thousands of CT scans that require annotation to identify lesions. The factory has operational logs from over 200 types of equipment that need to be organized into structured data. The power company receives real-time data streams every second that require cleaning, alignment, and feeding into models.
For ten years, Xunce has mastered end-to-end data processing across all nine stages and five major data tokenization workflows, covering the full stack from data acquisition, cleaning, and standardization to labeling, modeling, real-time stream computing, AI agent development, and finally, large model optimization and hybrid model integration. This has refined three core technological advantages: millisecond- to second-level real-time responsiveness, 100% accurate data processing, and robust data tokenization output capabilities—ensuring direct accountability for client business outcomes.
With these core capabilities, Xunce has grown into China's leading provider of AI-powered real-time data infrastructure and analytics services.
After the model was successfully implemented, Xunce took a key step: instead of just helping clients process data, they began charging based on Token usage. TokenONE, released in May 2026, marked this turning point.
The real turning point: the revenue formula has changed.
In the past, Xunce made money by selling projects. The revenue formula was simple: number of customers × project value.
Clients were acquired one by one, projects were signed one at a time, payment was collected upon delivery, and growth was linear. This is the fundamental reason why growth was only 19% in 2024—the ceiling of the linear model has been reached.
Until May 2026, Xunce officially launched the world’s first TokenOS operating system—TokenONE. In simple terms, it refines a company’s private data into AI-ready scenario tokens and charges based on usage volume.
The revenue formula directly became four factors:
Token revenue = Number of customers × Number of modules × Token price × Number of calls
The first three factors have ceilings. The number of customers is limited, the number of modules is countable, and although the token price is high, it falls within a specific range. Only the fourth factor—number of calls—has no upper limit.
Moreover, this factor is determined by the intensity with which customers use AI. The more frequently customers use and invoke the AI, the higher Xuncè’s revenue becomes. Previously, selling software was a push model; now, collecting Token fees is a pull model.
It’s like Xunce has installed an electricity meter for the client’s AI usage. The meter doesn’t turn on its own—it turns based on the client’s actual usage.
Data is validating this logic: In April 2026, Token call ARR experienced a quarter-over-quarter growth of 300%; by June, the proportion of Token-paid revenue had risen from approximately 5% in April to a quarter-over-quarter growth of 410% in Token call ARR, with the company targeting a year-end revenue contribution from Token services of 20%–30%.
At this rate of growth, token revenue is visibly contributing to income and becoming a new growth engine.
More notably, the pricing range: Xuncè’s vertical scenario token invocation costs between $10 and $100 per million tokens, more than ten times the price of general-purpose large models.
Why? Because General Tokens address "making AI understand language," while Scenario Tokens address "making AI do the right thing"—such as trading logic in financial risk management, manufacturing process expertise, or diagnostic experience behind medical imaging. Each Token encapsulates compressed industry knowledge that cannot be learned from public data alone.
This brings us back to the initial scissors gap: general-purpose tokens are standardized commodities that become cheaper the more they’re competed over; scenario-based tokens are tied to irreplaceable industry barriers and become more valuable the more they’re used.
Why should it be that?
Anyone can talk about tokenization, but only those who deeply understand "data" can truly deliver on it.
Xunce began in the asset management industry. Anyone who has worked in financial data knows this is the most demanding industry in terms of real-time performance and accuracy—requiring millisecond-level responses and 100% precision, where a single incorrect number could lead to losses of tens of millions. Since 2016, the company has spent nearly a decade refining its expertise in this field.
Using the same set of capabilities, over five years we expanded into 11 industries—including telecommunications, power, energy, advanced manufacturing, biomedicine, robotics training platforms, and commercial aerospace—accumulating 230 clients and over 300 standardized modules.
In 2025, the revenue share from non-asset management businesses increased to approximately 80%. Customer retention rates have consistently remained above 90%.
The industry-specific data know-how—which data to scrape in real time, what annotation standards are needed for each field, and which scenarios have irreversible consequences if something goes wrong—takes at least one to two years to accumulate for each industry. It cannot be achieved by simply throwing more computing power or tuning model parameters.
A foundational large model can reach a score of 80 using public data. Reaching 95 requires industry-specific proprietary data; advancing from 95 to 99 depends on real-time feedback loops. These last two stages are precisely where Xunce excels.
What I’ve accumulated over the past decade isn’t the data itself, but the eligibility to install that meter.
The table is installed—what’s next?
TokenONE is just the first step. Xunce's roadmap is clear:
Phase one is already underway. Converting enterprise private data into AI-consumable scenario tokens, charged based on usage volume. Token factories across healthcare, advanced manufacturing, finance, energy, and countless other industries are being implemented one after another. For example, Xuncè is collaborating with Gaochuang Dongzhi to build a manufacturing token factory, partnering with Botai Connected and Samoo Technology to develop intelligent vehicle scenarios, and working with Turing Quantum to enter the quantum computing field.
Phase two, scheduled to launch in the second half of this year, is TokenRouters—a nationwide token exchange platform. Previously used internally by individual enterprises, TokenRouters will now enable cross-enterprise token circulation. Non-competing businesses can securely and compliantly invoke, price, and exchange scenario-based tokens with one another. The company has established strategic partnerships with the Shenzhen Data Exchange and the Beijing International Data Exchange, laying the regulatory compliance and ecosystem foundation for global business expansion.
The revenue model has also changed: from individual client usage volume to the total circulation within the entire ecosystem. This is like the internet era, where a single website had limited earnings, but when websites connected into a platform and data began to flow, value multiplied exponentially.
The third stage is the endgame—enterprise small models. The goal is not for every company to rely on general-purpose large models, but for each company to have its own AI.
Compared to large models, smaller models offer lower costs, faster deployment, and easier debugging, providing businesses with a cost-effective gateway to artificial intelligence. In particular, data staying within local boundaries is a critical requirement for highly sensitive industries such as finance, healthcare, and aerospace.
At the national level, momentum is also accelerating. Liu Liehong, Director of the National Data Bureau, chaired a symposium on the token economy this year, explicitly proposing to integrate the promotion of the token economy into the official work system. In July, at WAIC, the token economy was listed as a core agenda item for the first time. Xunce, branded as the “first token company,” has now positioned itself at the starting line of this national-level initiative.
Just one last thing
The AI industry rests on three pillars: computing power, models, and data. The computing power segment is already dominated by giants. The models segment is also dominated by giants.
This position for data remains empty as of today.
The U.S. data company leader Palantir has surpassed a $1 trillion market capitalization. Until Xunce appeared on the Hong Kong Stock Exchange, there was no sizable listed player in this sector in China.
Meanwhile, the general token is still declining, while the utility token is still rising.
The company’s revenue is no longer determined by how much it sells, but by how much AI its customers use.
Xunce's recent performance surge is not merely a cyclical financial increase, but rather a business model transformation enabled by tokenization, built upon a decade of data infrastructure accumulation, at the critical juncture when AI is shifting from training to real-world deployment.
The table has been installed. What remains is a matter of speed, and the speed is not determined by Xunce, but by the curve of AI usage across all Chinese enterprises.
This curve has just begun to rise.