Enterprise AI Services Anchor the Core Business as a Second Growth Curve Emerges
“Why is nobody talking about OpenClaw anymore?” “OpenClaw is finished”…
Just as the internet began holding a “cyber memorial” for the viral product that once brought the Agent concept into the mainstream, one company has managed to turn its Agent business into—
Money!
What’s even more surprising is that Agent business at this company isn’t sitting at the “kids’ table.” It has already become the core engine driving the company’s growth.
Just now, Yunzhisheng, the “first AGI stock on the Hong Kong exchange,” released its 2026 interim financial report, revealing several striking figures:
- Revenue for the first half of the year reached 562 million yuan, up 38.7% year over year—nearly double the growth rate recorded during the same period last year;
- Revenue from intelligent agents reached 478 million yuan, up 35.7% year over year;
- Token business revenue approached 30 million yuan, up more than 500% quarter over quarter in Q2, with a gross margin of over 60%;
- Repeat purchases accounted for more than 60% of revenue, while the company held nearly 1 billion yuan in cash flow;
- ……
This creates a subtle contrast with OpenClaw’s current situation:
While OpenClaw is still being subjected to soul-searching questions about cost, stability, and practical value, Yunzhisheng has already brought Agents into core business scenarios including healthcare, insurance, and urban public services—and made real money from them.
Both are building Agents, so how exactly has Yunzhisheng turned them into a business worth hundreds of millions of yuan?
Enterprise AI Services Anchor the Core Business as a Second Growth Curve Emerges
The answer is right there in the interim report.
After reading through the financials, the most obvious change is this: Yunzhisheng is moving faster.
In the first half of 2025, Yunzhisheng generated revenue of 405 million yuan, up 20.2% year over year. In the first half of 2026, revenue jumped to 562 million yuan, representing year-over-year growth of 38.7%.
In just one year, its revenue growth rate surged from 20.2% to 38.7%—nearly doubling.
Where did this acceleration come from?
Let’s first look at the three major revenue segments Yunzhisheng newly defined in this interim report:
-
The first is enterprise AI services, generating revenue from implementation.
This includes intelligent-agent applications, intelligent-agent platforms, and integrated solutions. It is currently the company’s largest source of revenue.
-
The second is the Token business, added this year, generating revenue from model usage.
Customers directly access Yunzhisheng’s portfolio of large models through public-cloud APIs and pay based on Token usage.
-
The third is edge AI, generating revenue through the scale of terminal devices.
This mainly includes voice chips, modules, and in-vehicle solutions. Facing downward pressure across the industry, this mature business is seeking new growth through overseas markets, embodied intelligence, and other emerging scenarios.
Looking at the business mix and revenue quality, three signals are particularly noteworthy.
Highlight 1: Enterprise AI Services Anchor the Core Business
The most eye-catching segment is naturally Agent business.
In the first half of 2026, Yunzhisheng’s intelligent-agent business generated revenue of 478 million yuan, up 35.7% year over year and accounting for 85.1% of total revenue.
Breaking it down, this revenue is supported by intelligent-agent applications, intelligent-agent platforms, and integrated solutions.
More importantly, these three segments form a tightly connected chain.
The most direct approach is to send Agents into real-world scenarios to do actual work.
Yunzhisheng’s intelligent-agent applications have already entered more than a dozen sectors, including healthcare, insurance, transportation, and advanced manufacturing.
Once a scenario has been successfully implemented, mature capabilities can be incorporated into the intelligent-agent platform and expanded from serving a single enterprise to serving an entire region or industry.
When dealing with large customers with more complex systems, integrated solutions come into play, connecting models, data, and existing business systems as a whole.
In this way, applications penetrate deeper, platforms expand more broadly, and solutions become increasingly complete—naturally driving enterprise AI services to grow.
To date, Yunzhisheng has served more than 470 medical institutions, of which over 80% are top-tier tertiary hospitals. At the regional-platform level, it has completed data profiling and governance for 524,400 Xiamen enterprises. After entering advanced manufacturing, it also helped increase orders for a new-energy vehicle manufacturer by 10%.
Highlight 2: A New Growth Curve Is Already Emerging
With the core business anchoring the foundation, the Token business has also begun to scale rapidly.
In the first half of 2026, Yunzhisheng’s Token business generated nearly 30 million yuan in revenue, representing year-over-year growth of approximately 760%. Revenue in the second quarter alone exceeded 25 million yuan, up more than 500% quarter over quarter, with a gross margin of over 60%.
This business started from almost zero this year. Its current scale is still modest, but its growth curve is remarkably steep.
The reasons are not complicated.
On the model side, Yunzhisheng’s portfolio of large models, centered on Yunzhisheng U2, already offers the performance and stability required for direct use in production scenarios.
On the market side, more and more enterprises are beginning to access models through APIs and accept Token-based, usage-based pricing.
With the models ready and customers’ willingness to pay gradually maturing, the two forces have converged, naturally accelerating Token business revenue.
If the Agent business proves that Yunzhisheng can make money through implementation, the Token business further demonstrates that:
Customers have begun paying real money for Yunzhisheng’s model capabilities themselves.
Highlight 3: Repeat Purchases Account for More Than 60%, Making Growth More Resilient
Rapid revenue growth is only meaningful if that revenue can continue.
In the first half of 2026, repeat purchases accounted for more than 60% of Yunzhisheng’s total revenue, while average deal size also rose significantly.
The logic behind repeat purchases is easy to understand:
Once an Agent enters a core business process, customers tend to continue expanding its application scenarios, upgrading their intelligent-agent platforms, and using model capabilities. The relationship thus evolves from one-off delivery to ongoing services.
For an industrial AI company, signing a new project only proves that the product has been sold. Continued customer purchases demonstrate that the Agent is genuinely creating value.
More than 60% of revenue coming from repeat purchases speaks for itself. It also gives Yunzhisheng’s growth an additional layer of certainty.
Customers can also generate word-of-mouth referrals for one another. To date, Yunzhisheng’s contract value in 2026 has grown 65% year over year, with outstanding orders exceeding 1.5 billion yuan.
Taken together, these three changes make Yunzhisheng’s growth logic in the first half of the year quite clear:
Enterprise AI services support revenue scale, the Token business provides a second growth curve, and repeat purchases underpin the sustainability of the business.
Growth was not limited to revenue, either.
In the first half of 2026, Yunzhisheng generated gross profit of 186 million yuan, up 42% year over year—faster than its 38.7% revenue growth.
At the same time, loss attributable to owners of the company fell to 237 million yuan, narrowing by 20.2% year over year.
As revenue expanded, the net loss margin also narrowed by more than 31 percentage points year over year.
Gross profit is growing faster, while both the absolute loss and loss margin are declining. The trend of operational improvement is already quite clear.
That said, it is worth noting that Yunzhisheng has not yet achieved profitability.
This result is not particularly surprising. AI companies in China and abroad may be raising money at a frenetic pace and reporting seemingly healthy revenue, but profitability remains genuinely difficult.
OpenAI’s latest quarterly revenue rose 18% quarter over quarter, yet its operating loss expanded from $9.3 billion to $12.3 billion.
Anthropic charged ahead for years and only achieved positive adjusted operating profit for the first time in Q2 this year.
One is still burning through cash at a massive rate, while the other has only just reached the threshold of profitability. The performance of these two global AI leaders is enough to show that losses are nothing unusual in this industry.
Everyone understands that what matters more than whether a company is profitable right now is the direction of its loss curve:
If revenue growth is accompanied by expanding losses, growth remains expensive. If revenue accelerates while losses continue to narrow, commercialization is beginning to absorb the investments made in the early stages.
Yunzhisheng is showing the latter trend.
A Chinese AI Company Is Following a Palantir-Style Commercial Path
To briefly summarize the report, Yunzhisheng’s continued growth ultimately comes down to two things:
First, the models have to be powerful enough. If the models are not strong enough, it is difficult to generate sustained usage and payments, whether they are offered directly as APIs or packaged into specialized Agents.
Second, the company must deliver strong application services. There is a great deal of know-how involved in bringing models into a customer’s workflow and getting customers to keep using them.
These two points correspond precisely to Yunzhisheng’s strategy of “strong foundation models + deep applications.”
So what connects the models and applications into a scalable business?
The answer points to an industrial AI business model with a distinct Palantir flavor.
Palantir was founded in 2003 by Peter Thiel and others to help organizations of all kinds solve practical problems with data and AI.
It is mentioned here because it provides a classic example of how industrial AI can be commercialized.
Businesses of this kind face a common challenge:
Enterprise customers vary enormously, and delivery is inherently labor-intensive. How can this business scale?
Palantir’s development offers one possible answer:
- Step 1: Send engineers deep into the customer’s operations to understand its data, processes, and business rules;
- Step 2: Encode this experience in the Ontology, creating a “business map” that AI can understand and operate;
- Step 3: Use the Foundry platform to embed these capabilities into real workflows, turning data analysis into business action.
The first project is indeed labor-intensive. But by the time the company reaches the second and third customers in the same industry, the accumulated industry knowledge and functional modules can be reused directly.
Use intensive delivery to build industry expertise, then use a platform to turn that expertise into a repeatable business.
When we apply this logic to Yunzhisheng, some seemingly contradictory data points begin to make sense.
An AI company with fewer than 500 employees can handle several hundred million yuan in intelligent-agent business. Clearly, it is not relying on a “human-wave strategy.”
Behind it is a three-layer architecture that likewise emphasizes accumulation and reuse.
△Image generated by AI
At the bottom is the foundation-capability layer.
Regardless of whether the final product is Tokens, APIs, or a complete industry solution, the quality and usability of the model itself is always the first threshold.
To determine whether an industry model is truly useful, at least two things must be considered:
Is it intelligent enough, and does it genuinely understand the business?
The former tests the capabilities of the foundation model. Since launching its first self-developed general-purpose large model in 2023, Yunzhisheng has continued upgrading this “brain.”
Upgrades naturally require substantial spending.
In the first half of 2026 alone, Yunzhisheng invested 284 million yuan in R&D, accounting for 78.5% of the combined total of its R&D, sales, and administrative expenses.
Its 327 R&D employees account for 67.7% of the total workforce—meaning roughly two out of every three employees are engaged in R&D.
This investment and this team ultimately support the latest flagship model, U2.
As a natively intelligent-agent model, U2 has entered the global first tier in several widely recognized industry Agent evaluations.
More importantly, thanks to its sparse MoE architecture, U2 has 260 billion total parameters but activates only around 10 billion per inference, balancing speed and cost while maintaining strong capabilities.
Using U2 as its foundation, Yunzhisheng has further developed a model portfolio covering healthcare, speech, and vision.
Whether a model understands the business depends on accumulated industry expertise. Understanding a business is not simply a matter of feeding a model a few industry documents or connecting it to a knowledge base.
The model must also know what entities exist in the business, how processes operate, how rules constrain them, and what action should be taken next.
This is precisely what Yunzhisheng has accumulated through more than a decade of deep involvement in healthcare, IoT, transportation, and other scenarios.
Today, this experience has been further broken down into 17 subsets and 61 detailed categories, transforming knowledge previously scattered across business systems, operating manuals, and employees’ minds into business assets that models can compute and invoke.
In Palantir’s terminology, this amounts to drawing a “business map” for each industry.
Once the map is in place, the next step is to assemble the Agent. This brings us to the intelligent-construction layer.
This layer is jointly supported by the business operations platform and the intelligent-agent platform.
The business operations platform connects the customer’s business processes, data, and rules, defining the boundaries within which the Agent can act. The intelligent-agent platform selects models, connects knowledge bases, invokes tools, and orchestrates a sequence of tasks.
It is also at this layer that capabilities accumulated from one project become modules that can be directly invoked in the next.
For example, when developing a medical-record quality-control Agent for a hospital, general-purpose modules such as medical-record parsing and issue identification can first be accumulated. When the next hospital comes along, there is no need to reinvent the wheel.
To date, Yunzhisheng’s intelligent-agent platform, UniAgentOS, has accumulated 1,773 instantiated Agents. A single module has been reused up to 119 times, while project implementation costs have fallen from 21% of revenue during the same period last year to 17%.
At the top is the value-operations layer.
After the Agent has been built, it still needs to enter the customer’s workflow, collaborate with doctors, reviewers, and case handlers to complete tasks, and continue making adjustments based on the results of repeated use.
At this layer, the evaluation criteria suddenly become very straightforward:
How well is the work being done? How much time has been saved for the customer? How much cost has been reduced? Is the customer still willing to keep paying?
These questions can only be answered by putting the Agent directly into the customer’s workflow.
With the stage set this far, it would be a shame not to run through a representative real-world case (doge).
Medical-record quality control has long been a major headache for hospitals.
A large top-tier tertiary hospital may generate hundreds or even thousands of discharge records every day, each often running to more than 30 pages.
In the past, several quality-control staff members had to conduct manual spot checks, going through the records page by page and item by item. Being unable to keep up was the norm, and in the end, they could not cover even half of all records.
What happened after Yunzhisheng came in?
The company first used a medical large model and a knowledge graph covering 475,000 medical entities to help the Agent understand medical records and standards. It then connected the Agent to hospital systems such as HIS, EMR, and LIS, enabling it to read records, check for issues, identify defects, and provide revision suggestions on its own.
After deployment, the review time per medical record was reduced to under 10 seconds, quality-control coverage reached 100%, and staff efficiency improved by more than 80%.
Ten seconds, 100%, and 80%—these are business outcomes that can be directly measured and verified, and they are also the reasons customers continue paying.
Once capabilities such as medical-record parsing and rule verification can be reused across more hospitals, this labor-intensive delivery business gains the potential to scale.
After walking through this case, it is clear how Yunzhisheng’s model is implemented.
Taking out the Palantir yardstick once again, the biggest difference between the two also comes into focus:
Palantir does not place its bets on any single foundation model. It integrates whatever customers need and focuses on connecting the models, data, and business processes.
Yunzhisheng, by contrast, also keeps the model “brain” in-house. Beneath the stack is the U2 general-purpose foundation model; above it are industry models for healthcare and other sectors, multimodal speech and document capabilities, and even edge-side chips.
The two approaches each have their advantages and disadvantages. But in terms of the depth of coordination between models and industries, keeping the models—and even the entire technology stack—in-house creates better conditions for project feedback to flow directly into model iteration, generating long-term compounding returns.
△Image generated by AI
Clearly, Yunzhisheng is validating an industrial AI path with Palantir-like characteristics, driven by self-developed models.
Following this path, the answer to how Agents make money is also becoming increasingly clear.
OpenClaw Has Not Disappeared; Agents That Cannot Make Money Will
Making money is both difficult and not difficult (runs away screaming.jpg).
At least from Yunzhisheng’s experience, several practical lessons can be summarized:
First, find scenarios worth paying for.
What makes a scenario worth paying for? It generally needs to meet three golden criteria:
The problem occurs frequently enough, the cost of errors is high enough, and the results of solving it can be clearly measured.
Medical-record quality control, insurance claims processing, and medical-insurance audits are all typical examples.
Second, there need to be enough ways to monetize.
Once the right scenario has been found, the next question is: How should customers be charged?
Customers with sensitive data need private deployment. Those seeking rapid deployment are better suited to platform subscriptions. Customers with flexible usage needs are willing to pay by the Token.
It may sound like a matter of simply opening up a few more payment channels, but putting it into practice is not easy.
Models, platforms, and Agents must be capable of being sold separately while also being assembled into complete solutions at any time. This tests whether the entire capability set can truly be turned into products.
It is no wonder that Yunzhisheng has repeatedly emphasized in its public disclosures:
The company has established an end-to-end, full-chain service loop from foundational capabilities to value delivery. Each layer of capability can be delivered either as part of an integrated solution or as a standardized product sold independently. The company supports both private deployment and public-cloud API and Token access.
Third, make every project the starting point for the next one.
The compounding effect in the Agent business can be summarized in two directions:
Horizontal replication and vertical expansion.
Horizontal replication means mastering an entire industry. With every completed project, knowledge, processes, and functional modules continue to accumulate, making the next project more efficient.
Vertical expansion means deepening relationships with individual customers. Once an Agent enters a customer’s core processes, cooperation often expands from one task to more departments and scenarios.
One force spreads delivery costs across more projects, while the other increases customer lifetime value. Together, they get the compounding flywheel of the Agent business truly turning.
Of course, these points can only be considered a provisional summary. There is no standard formula for Agent commercialization; the rest must be validated in one real-world scenario after another.
But regardless, one thing is becoming increasingly certain:
It is time for the AI industry to start talking seriously about money.
In the large-model era, people looked at parameter counts. In the Agent era, they look at tool calls. Now that the novelty has worn off, customers are asking increasingly practical questions:
Did the work actually get done? Was time saved? Was the money well spent? Will the subscription be renewed next year?
Model capabilities remain important, but they have become the starting point. This is also what makes Yunzhisheng’s financial report worth watching.
It uses self-developed models to ensure that Agents can do the work, industry expertise to teach Agents how to do it, and platforms to replicate delivery experience—ultimately generating recurring revenue through solutions, subscriptions, and Token usage.
“Strong foundation models + deep applications” has thus formed a closed loop, transforming from a strategic slogan into the revenue and repeat-purchase curves reported in the financials.
Looking back at the “cyber memorial” for OpenClaw mentioned at the beginning, it is actually quite interesting.
Does the end of the hype mean the end of the Agent story?
Not so fast.
Once the novelty fades, Agents truly leave the trending charts and enter the examination room of commercial value.
OpenClaw has not disappeared, and Agents will not disappear either.
What will disappear are the Agents that merely consume Tokens without creating value.