Interview with Zhang Shaoting of SenseTime Healthcare: It's Hard to Break the Deadlock by Using New AI Technologies to Make Traditional Profits, and the Killer App in Healthcare Lies on the Consumer Side.

Text | Hai Ruojing

In this spring of "AI + everything", AI in healthcare has regained its popularity.

In the era of AI healthcare 2.0, which started with deep learning, there were many unicorn companies that received billions in financing and were valued at nearly tens of billions. However, limited by various factors, few companies have met the market's profit expectations.

As the timeline enters the era of generative AI (AI healthcare 3.0), the problems of high medical costs and difficult access to medical services still exist. What medical problems can the emerging new technologies change, and which companies can benefit from the business opportunities?

Recently, 36Kr interviewed Zhang Shaoting, the CEO of SenseTime Healthcare, to discuss the application forms, technology stacks, price wars, and payment models in the AI healthcare industry.

In February 2025, SenseTime Healthcare announced its first round of financing after the spin - off. Its pre - investment valuation was 1.5 billion yuan, and the financing amount exceeded 100 million yuan.

In the industry, SenseTime Healthcare has a unique position: Its team of over a hundred people has been quietly working in healthcare for seven or eight years. They have developed a large language model called "Da Yi", nearly 60 AI applications, and obtained multiple medical device certifications at home and abroad, achieving a certain scale of revenue.

However, when it was spun off as a startup, it was in the cold winter of healthcare capital, and the market's constraints on valuation were much stricter than before.

In Zhang Shaoting's view, it is not advisable to be too aggressive in the previous rounds of financing. Besides raising funds, the key is to build barriers with the help of shareholders. As a person with a technical background, he believes that technological advantages, business models, and channels are not sufficient to form long - term barriers. Only by leveraging the industry ecosystem can a company gain a head start.

How do you view the deployment of DeepSeek in many hospitals, and how can the problem of AI hallucination be overcome?

In the face of industry involution, price wars, and the lack of medical insurance coverage, how can the commercialization of AI healthcare break the deadlock?

How do you view the outside evaluation that "SenseTime is a last - generation AI company"?

Zhang Shaoting has shared his insights on all of the above questions.

Consensus and Anti - Consensus

36Kr: How do you view the deployment of DeepSeek in many hospitals? What impact does the open - source movement have on SenseTime Healthcare?

Zhang Shaoting: I once used an analogy of the three - stage launch model of a rocket for large models:

The first - stage rocket is the basic model, such as the "full - blooded version" of DeepSeek, Qianwen, and SenseTime's Shangliang, which are general - purpose models.

The second - stage rocket is to incorporate vertical knowledge, such as in healthcare. It's not just simple LoRA but includes full - parameter fine - tuning, instruction fine - tuning, and CoT (Chain of Thought) capabilities.

The third - stage rocket is to develop applications and integrate knowledge bases.

It is risky to directly develop third - stage rocket applications based on the first - stage rocket. For example, DeepSeek - r1, the so - called full - blooded version, has obvious hallucination problems. Moreover, a model with 671 billion parameters requires a large amount of computing resources for deployment. If the "deep thinking" function is enabled, a large number of token outputs will also lead to a decrease in concurrency, resulting in high usage costs for hospitals. I think from a cost - effectiveness perspective, vertical models with 20 - 100 billion parameters that have been quantized are more suitable for hospital deployment.

Since September 2022, SenseTime Healthcare has been focusing on the second - stage rocket. In February 2023, we brought our healthcare vertical model to hospitals for trials. The continuous emergence of basic models and the open - source movement are definitely good for us, as we have more excellent basic models to choose from to iterate our vertical models.

36Kr: SenseTime Healthcare emphasizes "multi - modal large models + platform - based deployment". Several healthcare AI companies with relatively high valuations in the industry have penetrated the market through single - point applications, such as lung nodules, brain tumors, and coronary heart disease. When did the idea of building a platform start?

Zhang Shaoting: In 2018, when I started the healthcare business at SenseTime, I proposed the concept of "empowering the whole - hospital clinical diagnosis, treatment, and recovery". The key words were clinical, whole - hospital, and diagnosis - treatment - recovery. It's not just about empowering the radiology department or providing auxiliary diagnosis.

At that time, many manufacturers were developing AI for reading radiology films, but we believed that such a business model would be difficult to succeed for most enterprises.

If it's about the radiology department, it's essentially about the business of large equipment such as CT and MRI, rather than artificial intelligence software and informatization. Software requirements usually need to be submitted through the hospital's information department. Therefore, auxiliary diagnosis software for the radiology department often has to be sold together with hardware equipment, and pure software companies have limited bargaining power.

Moreover, auxiliary diagnosis in the radiology department is just a small entry point for AI in healthcare. AI can play a role in many departments, including surgical treatment. Solving the clinical needs of multiple departments can better convince hospital procurement decision - makers.

Therefore, in 2018, SenseTime developed the SaaS platform SenseCare to empower the whole hospital. It integrates modules such as auxiliary diagnosis, auxiliary treatment, and surgical planning on the AI platform, sharing the underlying technology. This is how the idea of platform - based deployment came about. In 2018, SenseCare made its debut, equipped with 3D modeling for preoperative planning.

36Kr: In 2018, you judged that the auxiliary diagnosis business in the radiology department would be difficult to succeed. Didn't competitors at that time realize the problems such as inconsistent demand and decision - making parties for film - reading software?

Zhang Shaoting: In those years, most manufacturers didn't really worry about the difficulties in the radiology business. On the one hand, the capital market environment was relatively good at that time. On the other hand, they thought the market needed an education process. However, by 2020, when they found that this path really didn't work, everyone started to emphasize clinical treatment and surgical planning.

What is now a consensus was often an anti - consensus a few years ago.

36Kr: Do hospitals prefer a "comprehensive" diagnosis and treatment platform or "specialized" products for trial use first?

Zhang Shaoting: In addition to clinical value, hospitals also need to consider budget constraints. Many hospitals don't allocate a large budget at the beginning. They hope to verify the value brought by AI first, so they will choose one or two AI products for trial use. However, they also hope that the selected suppliers can provide incremental products in the future. Otherwise, it will be inconvenient for hospital management if they have to install a dozen software products from a dozen different manufacturers.

For example, Kiang Wu Hospital in Macau has continuously purchased nearly 20 artificial intelligence products from SenseTime Healthcare in the past five years since 2019.

Involution and Barriers

36Kr: Do you think there is a consensus in the 2B healthcare AI industry in terms of technology routes and commercialization?

Zhang Shaoting: There should be a consensus on the commercial difficulties. It is very difficult to turn a single software product into a commercial hit. As long as it is a standardized software product, it will face price wars. As long as it is a project (custom - made), it will involve intense competition in labor costs. Neither of these situations meets the development needs of healthcare AI companies.

Some healthcare AI companies are expanding horizontally, from lung nodules to head, chest, and abdomen. Others are going deeper, from preoperative diagnosis and surgical planning to intraoperative navigation. However, it is still difficult to achieve high - speed revenue growth, high - quality payment collection, and cost control simultaneously.

So, I say that neither technology nor business models can be regarded as long - term barriers. They can only bring short - term advantages. Of course, this is the case for most 2B industries in China.

36Kr: Can neither technology nor business models be regarded as barriers?

Zhang Shaoting: First of all, without a technological first - mover advantage, we wouldn't even be able to enter the game. However, in the 2B industry, this won't be a long - term barrier.

Investment in technology can certainly bring advantages and may become a key factor for cost - reduction and efficiency - improvement in the future. However, it is unrealistic to rely on technological advantages to keep late - comers out in the long run, especially in an environment where IP is not highly protected.

Even though OpenAI has such advanced and excellent technology, its models can still be distilled, and it has recently been impacted by many companies, including the impact of open - source on closed - source models. In the 2B industry, product design and detailed refinement are as important as technology.

36Kr: What about the commercial aspect? Some investors believe that if healthcare AI products can be charged on a per - use basis frequently, there will be opportunities.

Zhang Shaoting: In China, it is difficult to get medical insurance to pay for per - use charges, as medical insurance also faces cost - control pressure. If patients are guided to pay, there will also be risks for manufacturers. Even if it can be done, it is likely that operational companies will benefit rather than technology - product - oriented enterprises.

Of course, it is also not easy to get hospitals to directly purchase software modules, as hospitals will assess the profitability.

SenseTime Healthcare has been promoting the "whole - hospital digitalization" concept. Through platform construction and the integration of dozens of AI applications, it actually follows the key - account (KA) strategy. We have done well in serving KA customers in the past. After being spun off from SenseTime Group, we will also leverage the power of distribution channels to expand from large tertiary hospitals to primary - level hospitals and increase our revenue scale.

36Kr: If neither technology nor business/price can form long - term barriers, what do you think is the long - term barrier for competition among healthcare AI manufacturers?

Zhang Shaoting: I think the "ecosystem barrier" and a rich product matrix are relatively difficult to replicate.

Here is an interesting example. On the day when the popularity of DeepSeek caused a sharp drop in NVIDIA's stock price, several technical experts in our large - model technology team bought NVIDIA stocks at the bottom.

Their basic judgment was that writing communication codes with PTX was not enough to shake the CUDA ecosystem. Developers still highly rely on the CUDA ecosystem. Therefore, the emergence of DeepSeek should not affect NVIDIA's expected performance. Instead, it would be beneficial to the sales of NVIDIA's H100 cards, as they are the most suitable inference chips. Later, the stock price quickly recovered. The same principle applies to the Android ecosystem.

Back to healthcare AI, unfortunately, a product with a score of 99 developed at a huge R & D cost may not sell better than a product with a score of 90 developed at a low cost, especially in a price war. It is very likely that bad money will drive out good money. Therefore, while pursuing high accuracy in AI, we also need to control costs. This "having it both ways" relies on the accumulation of pre - trained models in the early stage.

SenseTime invested heavily in building a supercomputer in 2017 and spent 6 billion yuan in 2018 to build Asia's largest single - entity computing power center. Now, it has 54,000 GPU cards. Computing power is one of the core advantages for pre - training large models, but this was also an anti - consensus seven or eight years ago.

36Kr: What about in terms of medical resources?

Zhang Shaoting: In this round of financing, among many options, we decided to accept the investment from Midea's industrial fund, Yingfeng Holdings, and Renwei Technology (the investment platform under People's Medical Publishing House) because of the consideration of ecological resources.

Midea and Yingfeng have long - term layouts in the large - health industry. For example, the Heyou Hospital, in which they have invested over 10 billion yuan, is equipped with proton and heavy - ion equipment, and there are many scenarios where AI can be applied. We have already started multi - aspect cooperation with Heyou Hospital.

People's Medical Publishing House serves over one million medical students and young doctors in China. The application of AI in simulated medicine and education and training is also a huge opportunity. AI manufacturers focusing on 2C applications will all value this market. Introducing Renwei as a shareholder of SenseTime Healthcare makes it easier to carry out suitable business together.

SenseTime has been in the healthcare field for seven or eight years and has accumulated a rich product matrix through previous investments. In this spin - off, SenseTime Healthcare is positioning itself as a Pre - A - round company for financing. The consideration is to introduce shareholders with resources. Financing is not just about raising funds but also about building a cooperative ecosystem.

AI Healthcare 3.0 and Killer Applications

36Kr: There is a large amount of existing medical data in hospitals, but the quality varies. What technical capabilities does SenseTime Healthcare have in data governance, and how can it balance the cost of data annotation and the generalization ability of models?

Zhang Shaoting: There are two research paradigms for dealing with the problems of data annotation and model generalization in specific scenarios. One is to search for data, annotate data, and train models for a specific problem.

The other is to have a pre - trained model with good generalization ability. Only a small amount of data needs to be annotated, and it can be adapted to different downstream tasks. This is the ability of pre - trained foundation models that is often mentioned now.

In terms of algorithms, SenseTime started researching the application of pre - trained models to downstream tasks early on. In the smart city scenarios of the group, there are many niche applications that are difficult to address using the traditional big - data - driven approach (i.e., annotating a large amount of data for each segmented task).

We refer to scenarios with low frequency, small data volume, and difficult annotation as "long - tail requirements". There are also many long - tail requirements in healthcare scenarios. Therefore, we have been using pre - trained models at the algorithm level early on. Although their accuracy may not be as high as that of big - data - driven models, they solve the problem of availability and are actually sufficient.

A recent representative case is the medical big - data training facility project won by SenseTime, which was also called "the first large - scale order for medical large models" by some public accounts.

This is an AI project initiated by Shanghai Shenkang in 2023 with a project amount of 43.96 million yuan. It involves 20 years of medical data from dozens of municipal hospitals in Shanghai. Model training needs to be carried out on 24 sets of multi - modal data, such as CT, pathology, ultrasound, and MRI, to achieve the full - process from semi - automatic data annotation, model training, to intelligent evaluation of pre - trained models.

We completed the project in about half a year and passed the acceptance last year, thus accumulating the ability of a "model production platform". Combined with our "model application platform", the product system is closed - loop.

Early on, we launched a "digital human" project with Ruijin Hospital, which involved fully segmenting more than 200 organs of the human body, including complex blood vessel networks. Therefore, when developing new applications later, such as surgical planning systems for the liver, chest, urinary system, and pancreas, we could use many basic modules, making it as easy as building with Lego bricks.

36Kr: Can you give an example of a healthcare application scenario? Compared with traditional small models, what problems are foundation models more suitable for solving?

Zhang Shaoting: For example, AI - assisted pathological diagnosis.

There are some technical difficulties in it, including:

Pathology belongs to IVD (in vitro diagnosis). The pathological sections from different hospitals may vary greatly, making quality control difficult. In addition, different hospitals use different scanners. This requires the pathological large model to have good generalization ability so that it doesn't need to be fine - tuned for each hospital.

In terms of application, the pathological AI model has performed well in screening based on small specimens, that is, determining benign or malignant conditions. However, it still faces challenges in diagnosing cancer subtypes based on large specimens. Because this is equivalent to the AI having to make more than a dozen different judgments. Even if the accuracy of each judgment reaches 95%, the probability of all judgments being correct