Who Can Break Through the Homogenization Bottleneck? Insights from a Survey of 300 Medical AI Large Models

After more than 1,000 healthcare-specific vertical models emerged within six months, the market has entered a cooling-off period.

NVIDIA, which peddles computing power; cloud service providers that allocate computing resources; and AI tool developers focused on the B2B sector are the biggest winners in this trend. Particularly the first two: as GPUs have transformed into “zero-risk” money-printing machines, they can profit from scarce computing resources simply by providing basic infrastructure services.

In terms of application, although developers of large medical models can create high-quality vertical models, they struggle to achieve successful commercialization. High investments have yielded negative returns, with very few enterprises managing to generate profits from LLM applications.

Developers face numerous questions: Which application scenarios should be prioritized as R&D directions? How can competitive barriers be established? Should a SaaS model or a perpetual license model be adopted? All these decisions require swift action at this stage.

Before discussing the aforementioned issues, let us first review the distribution of domestic vertical medical models.

Although the number of domestic medical large models has surged, existing models exhibit strong homogenization with significant functional overlap. Fundamentally, large models require extensive training data. From a cost perspective, enterprises prioritize the ease of data acquisition, starting with readily accessible medical datasets to develop corresponding medical AI solutions.

In May 2025, VCBeat conducted a statistical analysis of the functionalities of 288 domestic large medical AI models, covering 12 categories of application scenarios. The data revealed a total of 814 functionalities across all applications, with medical services alone accounting for 430 functionalities, representing over 50% of the total.

Distribution of Functional Capabilities Across Various Medical Large Language Models Among 288 LLMs

So-called medical services refer to AI vendors providing AI-powered services directly or indirectly to patients through large language models, including scenarios such as AI-assisted consultations, patient triage, telephone customer service, and intelligent health education.

These scenarios share a common characteristic: R&D enterprises typically have a large user base, making relevant medical data relatively accessible. Generally, internet healthcare companies and health informatics firms can integrate various types of routinely generated data through their related business operations to establish corresponding training datasets.

Large language models (LLMs) for medical research are also relatively easy to develop. Since most research findings are openly accessible, and given that numerous researchers have already constructed knowledge graphs related to scientific literature over the past few years, enterprises can rapidly build LLMs equipped with common functionalities such as literature retrieval and data analysis.

In contrast, there are fewer vertical models in the clinical domain. Although these models benefit from abundant data volume and a high degree of standardization (e.g., DICOM, HL7), their development requires not only processing textual data but also analyzing imaging data, resulting in higher construction costs and more demanding technical requirements.

Currently, most domestic clinical vertical models focus on the field of pathology. This is because pathological data are easy to preserve and standardize; pathological specimens from the past 20 years can be digitized using whole-slide scanners after simple cleaning. Consequently, the training datasets for large pathology models can easily reach millions or even tens of millions of cases.

Medical Vertical Models Involving Pathology in China (Incomplete Statistics; Data Source: VCBeat)

Although imaging data from CT, MRI, and other modalities are voluminous, significant variations exist in factors such as anatomical regions, resolution, scanner models, and lesion distribution. Consequently, very few models can provide auxiliary diagnostic support while maintaining low hallucination rates, and the datasets available for training vertical-domain models remain limited. Currently, imaging-specific vertical models primarily serve as tools to assist physicians in efficiently processing multimodal medical images, rather than being deployed at scale in clinical practice.

Currently, vertical models in the clinical sector are primarily led by hospitals with corporate assistance in development. These models are characterized by distinctly differentiated application scenarios, generally smaller model sizes, and functionalities focused on clinicians’ needs. Leading hospitals have participated in the research and development of applications for scenarios rarely addressed by enterprises, such as radiation therapy and infection prevention and control.

Medical Specialty Models Co-Developed by Leading Hospitals in China (Incomplete Statistics; Source: VCBeat)

Moreover, the number of large models in the biological field is also relatively scarce. Since genomic and clinical trial data involve trade secrets, are fragmented, and lack sharing mechanisms, it is difficult to build large-scale clinical datasets.

In addition to the difficulty of data acquisition, user scale is also a significant factor influencing the deployment of medical vertical models.

In the long run, enterprises should develop large language models (LLMs) for specialized clinical scenarios. These fields exhibit strong differentiation and face competitive pressures, making them more aligned with hospital needs. However, these scenarios also present pain points such as limited scale and significant challenges in data acquisition. Focusing solely on individual specialty areas is unlikely to generate sufficient returns to cover corporate investments.

Therefore, in the short term, enterprises developing large language models (LLMs) should focus on scenarios with a substantial user base. For instance, LLMs for healthcare informatics should first be deeply integrated into Hospital Information Systems (HIS) and data centers. After achieving initial success, companies can then collaborate with hospitals to develop specialty-specific models, thereby creating a “1+N” comprehensive solution and establishing a sustainable business model.

Overall, the development of vertical medical models is not determined by medical needs or competitive barriers. Instead, the difficulty (cost) of data acquisition and the scale of potential users (potential revenue) have had a more profound impact on the release of vertical medical models in this wave. There is an inverse relationship between the difficulty of data acquisition and the number of application-scenario models, while there is a positive correlation between user scale and the number of application-scenario models.

Furthermore, given the unique privacy and security concerns associated with medical data, hospitals will play a pivotal role in the development of clinical vertical applications. Clinical data harbors immense, yet often untapped, value. Enterprises with higher levels of digitalization—reflected in superior ratings for electronic medical records (EMR) and health information interoperability—are better positioned to develop vertical models that truly meet clinical needs.

As for which category of applications is expected to achieve commercial viability first, WeDoctor believes that the commercialization path of large language models will exhibit a characteristic of “B2B first, followed by C2C penetration.” Scenarios such as assisted diagnosis and treatment, health management, and drug R&D are likely to be the first to achieve successful commercialization. However, in the long run, if a company aims to achieve multidimensional breakthroughs in technology, payment systems, and ecosystem collaboration, it must build diverse application scenarios, continuously accumulate high-quality data, and constantly enhance its AI capabilities.

When it comes to intelligent software applications, enterprises often engage in intense debates over choosing between SaaS and perpetual licensing models, seeking the optimal commercial solution. However, for large language models, we may need to explore new pricing models that better align with this technology.

Unlike traditional scenarios, evaluating an AI model requires a comprehensive assessment of its understanding of the healthcare industry, problem-solving capabilities, interactive abilities, and user experience. The more feature-rich large language models become, the more challenging it is to evaluate their competitiveness.

Therefore, in the era of large language models, AI models that can be developed into standalone products for hospital scenarios—such as accelerated image acquisition, quality control, follow-up, and chronic disease management—can still achieve rapid commercialization through SaaS or perpetual licensing.

Taking Fuxin Kechuang as an example, the company approaches the entire workflow of “identification–acquisition–guidance–delivery” by deploying data governance agents to manage and mine hospital data. It leverages fully automated proactive health management to help hospitals automatically acquire patients and increase revenue, enhances patient services through AI-powered consultation accompaniment and AI-based pre-consultation to ensure a positive patient experience, and utilizes products such as generative electronic medical records to enable AI to drive the entire hospital workflow, rather than merely assisting in hospital operations.

Currently, the implementation of this system has yielded significant operational benefits for the hospital. On one hand, through efficient data mining, the hospital can identify valuable data and cases relevant to its operations, scientific research, and teaching. This capability enables the hospital to attract more patients, diversify its patient base, and improve the utilization rates of outpatient services and inpatient beds. On the other hand, continuous follow-up and management extend the duration of the hospital-patient relationship, thereby directly generating revenue and enhancing the hospital’s economic performance. In other words, this project not only focuses on patient health management but also becomes an integral part of the hospital’s sustainable development, helping it maintain competitiveness and innovation capabilities.

However, for scenarios such as R&D investment in pharmaceuticals and medical devices, and hospital management, short-term ROI is difficult to measure; neither SaaS nor perpetual licensing represents the optimal payment model in healthcare settings.

In this era of top-down cost-cutting and efficiency improvement, “selling outcomes” is more attractive than “selling tools.”

“Selling outcomes” refers to a model where payers do not pay, or only partially pay, the costs associated with large language model deployment during the initial service procurement phase; instead, the primary payment is determined based on “final outcomes.”

Take Deepwise Medical, one of the top 50 generative AI companies globally, as an example. The company has multiple product lines that have achieved sustained commercialization worldwide through subscription-based and per-case pricing models. Under this model, the final payment amount is determined by value-driven ROI metrics, such as the number of patients helped daily in clinical settings and patient throughput.

In this process, DeepWisdom Medical did not merely provide tools; instead, it established business logic and clinical procurement standards centered on delivering ultimate commercial and clinical value, serving as a reference for user payment decisions. Currently, the company has deployed its solutions in over 700 hospitals and imaging centers worldwide. It has also achieved commercialized collaborations with pharmaceutical companies such as Bayer, Bracco, and Telix to empower the value of AI in imaging agents, promoting AI-generated models to clinical contrast agent application scenarios.

Overall, the emergence of large language models has ushered in a diversified era for the commercialization of medical software. Through SaaS, perpetual licensing, “pay-for-results,” and hybrid business models, healthcare institutions may find it easier to prioritize efficiency improvement as their core objective, while aligning software pricing more closely with its true value.

Based on the development of AI (deep learning) in recent years, cases where independent products have been successfully commercialized and achieved profitability in healthcare settings are extremely rare.

Most AI-enabled products that are currently profitable were already generating profits before the integration of AI (e.g., HIS, PACS, and DRG systems in hospital informatization; EDC and eCOA platforms in pharmaceutical digitalization). In these scenarios, AI’s impact on product commercialization is limited; its value lies in enhancing user experience and raising the competitive barriers for products or solutions.

Today's large language models also face the same problem.

However, unlike all previous innovative technologies that attempted to disrupt the healthcare industry, large language models (LLMs) have distinguished themselves by penetrating the healthcare system within just a few months, prompting hospitals to proactively, rapidly, and at scale adopt the necessary infrastructure and engage in the development of clinical applications.

Clinical data harbors immeasurable value, yet the vast majority of existing vertical models have failed to enter clinical practice. When physicians become the primary developers of these applications, we may witness large language models achieving dual breakthroughs in application scenarios and commercial value within months, thereby reconstructing the healthcare system with a new generation of digital-intelligence applications.