The medical AI industry is generally loss-making. Yunzhisheng gained hospital trust by tapping into medical record data, expanded into commercial insurance and medical insurance bureau services, and achieved higher customer value and revenue growth.
Medical AI has always been a slow business.
No matter how hyped the general large language models get on the outside, anyone doing business with hospitals has to play by the rules of public institutions. Procurement cycles are measured in years, and no one can bear the responsibility if a system fails. The cruelty of medical AI is that the window of opportunity is closing. Companies that started working inside hospitals back in 2016, and players who only entered after the 2023 large language model boom, are not holding the same ticket.
Looking at peers' financial reports, everyone is generally struggling: Airdoc Technology, which does retinal screening, saw revenue drop 20% to 156 million yuan in 2024 and posted a loss of 255 million yuan, bringing its cumulative losses since 2019 to nearly 900 million yuan. Imaging AI companies like Infervision are generally trapped in a homogenized cycle of "getting certified, entering hospitals, and cutting prices," with distribution costs eating up most of their profits.
But Yunzhisheng just released its 2025 financial report, and a few data points don't match the overall temperature of the industry.
The report shows Yunzhisheng's smart medical business brought in 244 million yuan in a single year, up 22.3%. What deviates even more from the industry trend is that revenue from services extending outward from the hospital side quadrupled, which by calculation directly pushed the average revenue per medical customer up 53.2%. In an environment where peers are generally cutting prices to win deals and surviving by bundling hardware, an AI company sold its unit price at a 50% premium while revenue still grew at double-digit rates.
This math can't just look at the technical specs of large language models. Peeling back the wrapper of large models, we only look at the most concrete business ledger: who exactly did it sell its code to? Why can it sell at a premium? If this way of making money works, can it still be applied in another province?
The Peripheral Predicament: Peers Who Can't Command a Price
In the past few years, companies doing medical AI were basically losing money, with the core reason being that products couldn't sell at a good price.
The old standard playbook was: spend tens of millions to buy a batch of anonymized CT scans, train an algorithm to detect lung nodules, queue up at the National Medical Products Administration for a certificate, then have regional sales reps knock on hospital IT department doors.
But the IT director's math is very sharp. A top-tier hospital's annual IT budget is tightly capped. To buy a software suite worth several hundred thousand yuan, it either has to help the hospital make more money or save real money. A lung nodule AI at best saves an imaging doctor two minutes of reading time per scan, and the hospital can't charge patients an extra 50 yuan for an exam just because AI was used. This is what's called lacking a genuine motivation to pay.
Especially after health insurance bureaus across regions fully implemented diagnosis-related group payment reform, the amount a hospital can be reimbursed for treating a patient has been locked down completely. Every department in a hospital is desperately cutting costs. Since they can't generate revenue, these kinds of peripheral support tools are often the first to get their budgets cut, or AI companies are forced to become free add-ons for large MRI equipment manufacturers.
Even if a company manages to get its equipment into a hospital, the Damocles sword of policy can fall at any time. Take Airdoc Technology as an example: on July 1, 2024, a new regulation from the National Medical Products Administration banned the sale of laser myopia treatment devices that hadn't obtained Class III medical device certification. That single policy change forced Airdoc to take a one-time goodwill impairment of 43.2 million yuan plus a provision for losses of 23.7 million yuan, directly wiping out its entire annual profit. The unpredictability of policy regulation makes hardware-dependent AI medical companies like these walk on thin ice.
As long as they stay stuck on reading images and making simple triage chatbots, AI companies will never touch a hospital's most valuable asset: the doctor's clinical reasoning logic. If they can't get into the doctor's core working system, they can't obtain real medication habits and prescription data. Without new data, the algorithm can only stay in place. In the end, all companies doing point solutions can only compete on rock-bottom prices in a homogenized red ocean.
The Core Entry Point: Trust Built Over Time
Yunzhisheng's approach is different. It goes straight after the most sensitive and monopolistic asset in a hospital: the entry point for generating medical record data.
This isn't simple "informatization." Medical records are a hospital's data sovereignty, touching three red lines at once: patient privacy, clinical responsibility, and medical insurance compliance. For an external company to embed AI into this link, what's needed isn't just technology, but trust assets that make top-tier hospital IT departments willing to connect the system to their internal network. That can only be built up through years of on-site refinement and a track record of handling countless real medical disputes.
A hospital's internal network is physically isolated. Generic large language models on the outside immediately hit a wall when they encounter outpatient medical records. In real hospital wards, doctors often type in continuous streams, mixing Chinese and English, even using slang that only doctors in that particular hospital understand. For AI to comprehend this, it has to be fed with real patient cases happening every single day.
Yunzhisheng has been laying groundwork in this track for years. There are a few hard indicators in the report: it's connected to nearly 450 hospitals nationwide, nearly 85% of which are tertiary hospitals, and over one-third of its customers have been with it for more than three years.
Having been embedded for years in the systems of top-tier hospitals like Beijing Friendship Hospital and Peking Union Medical College Hospital, Yunzhisheng has accumulated 10.02 million medical relationships and 5.19 million medical terms. This isn't copied from Wikipedia definitions; it's real combat experience happening every day. For example, if an elderly patient with diabetic complications is hospitalized for a fracture, will the painkillers prescribed by the orthopedic surgeon conflict with the diabetes medication from endocrinology? AI can flag this before the doctor hits the enter key. This is tangible, quantifiable value.
At Beijing Friendship Hospital, after the system was embedded into the workflow, the changes were very direct. It increased medical record generation more than tenfold in a single campus, raised medical record review coverage from under 5% to 100%, and achieved a defect detection accuracy rate above 90%. The erroneous records that manual random audits used to miss are now being screened in full volume by machines. When AI becomes a productivity tool that doctors must open every day at work, it becomes a routine item that the hospital IT department must renew the second year. That's the core of how it holds down its base at the hospital end.
The Pricing Power Shift: Commercial Insurance and Medical Insurance Footing the Bill
Relying on hospital IT budgets alone can't support a 53.2% surge in average revenue per customer. Individual hospitals in China have strict caps on IT budgets; a software suite tops out at a few hundred thousand yuan. Where Yunzhisheng actually makes its money is by taking the business logic proven inside hospitals and selling it to commercial insurers and medical insurance bureaus, which have much deeper pockets.
That's where the "4x growth in extension revenue" really comes from.
Even more critical is that the bargaining power of the payer has undergone a fundamental reversal. When a hospital IT department buys software, there's layer upon layer of price comparison and government procurement processes. When an insurance company buys risk control capability, it's calculating return on investment: as long as AI can block even 1% more insurance fraud, the money saved covers three years of service fees. From having its budget strangled to being paid based on outcomes, Yunzhisheng is no longer selling software licenses but risk-pricing capability.
Since AI can read top-tier hospital medical records and know what medications should be prescribed for what conditions, it can conversely act as the most stringent "underwriting reviewer." In recent years, commercial insurance companies have been losing heavily every year to excessive claims and fraud. A few dozen human claim reviewers flipping through hundreds of pages of hospitalization lists is slow and prone to missing things.
Yunzhisheng connected its large model to commercial insurance systems, and it can scan hundreds of claim documents in a second. A routine conservative treatment for a fracture, but the claim form lists a bunch of expensive traditional Chinese medicine injections and high-priced physiotherapy beyond the approved indications? AI immediately blocks it. In 2025, Yunzhisheng's commercial insurance intelligent agent platform reviewed more than 2.6 million claim orders, improved the cost control rate to approximately 20%, and achieved over 1 billion yuan in incremental cost control for partner insurance companies. Insurance companies protected their profits, so they pay willingly, and that's how the average revenue per customer got pushed up. Additionally, for small and medium insurance institutions that can't afford large full-scale systems, they also launched a pay-per-order flexible mechanism, turning a one-off transaction into a recurring revenue stream.
By early 2026, they had also won the first provincial-level vertical large model project for medical insurance in China: the Jiangsu Province medical insurance large model project. The nature of this transaction changed. It used to be ToB, earning a few hundred thousand yuan in procurement fees from one hospital. Now it's ToG, doing the underlying regulatory review for the medical insurance fund pool of an entire province. The client's spending power has jumped an order of magnitude, and the total revenue pool naturally swells.
The Labor Barrier: The Cost of Delivery
Using data from large hospitals to train models and selling them to medical insurance bureaus for cost-containment margins. The logic is sound, but in actual business execution, there are unavoidable hurdles.
Winning the first provincial-level medical insurance large model project is a start. But across more than 30 provinces, every local medical insurance system is a different flavor. Jiangsu might run on a foundation built by Neusoft, while Shandong might be running on legacy systems coded by local vendors decades ago. The DRG deduction weights and disease coding standards in each city vary wildly. A model from Jiangsu can't be directly deployed in Sichuan.
That means every time it enters a new province, Yunzhisheng has to rebuild a localized delivery capability, from engineers who understand local medical insurance rules to implementation teams that integrate with legacy systems. This is a common cost across the medical informatization industry, not something any single company can bypass.
This delivery model that heavily depends on human labor shows up in financial reports as a typical industry cost structure. Any company transitioning from standardized software to customized solutions has to go through periodic gross margin fluctuations. Yunzhisheng's 2025 gross margin declined from 38.8% to 36.1%, which precisely confirms that its business upgrade is actually happening.
The non-standard nature of medical insurance systems means every player has to rebuild localized delivery capabilities for each new province. The implementation team Yunzhisheng has built up, with familiarity in each province's medical insurance rules, is a first-mover advantage accumulated over ten years, not a hastily assembled outsourced crew. The true value of the 2026 Jiangsu project is not just the order from a single province, but validation of a transferable methodology: which fields can be standardized and packaged, which interfaces can be reused, and which local medical insurance bureaus are willing to pay a premium for efficiency.
Yunzhisheng's operating cash flow in 2025 was negative 212 million yuan, but this is typical of the medical AI industry's transition from a "project-based" model to a "platform-based" model. Two post-IPO placements (197 million yuan in January 2026 and 312 million yuan in February) have provided sufficient capital reserves for this strategic transformation.
Final Thoughts
At a time when peers are generally losing money, Yunzhisheng's financial report does represent a phased advance. It proves that medical AI doesn't have to be trapped in procurement packages worth a few hundred thousand yuan. As long as you can identify clients like commercial insurers and medical insurance bureaus that are willing to pay real money for cost reduction and efficiency gains, code can still command a healthy price.
Yunzhisheng has already figured out the first step of the medical AI equation: spend ten years to earn a ticket into core systems, use the ticket to get data, use data to build models, and use models to gain pricing power.
The ledger for the second step is opening now. 2026 will tell whether the Jiangsu project can distill a transferable methodology and turn first-mover advantage into economies of scale. While peers are still evaluating when to get in, Yunzhisheng is already defining the rules of the game in this track. Whether the first-mover's time gap turns into a real gap in distance will be decided in 2026.