Over the past 28 months, the market has completed a full stress test on "medical AI diagnosis" hot stocks: Tempus has retreated approximately 35% from its historical high; the significant valuation drawdown prompts us to re-examine the real commercial bottlenecks of "medical AI diagnosis."
This article is based on 212 publicly disclosed AI-related medical device registration certificates in China from early 2020 to September 2026, as well as policy original texts. Along the three commercialization gates of "productization—registration certificate—price project establishment," it analyzes the core commercial bottlenecks of AI diagnosis layer by layer.
Productization: Regulatory requirements dictate that the design logic of AI products must be based on medical guidelines and specific to vertical indication areas. General-purpose large model products following the DeepSeek pattern are difficult for regulators to approve, reflecting a reverence for medical seriousness and clear boundaries of responsibility.
Registration Certificate: Standalone software, or software components embedded in medical devices (where the registration entity remains the medical device), are the only two product registration forms recognized by regulators for AI products. Commercialization of AI diagnosis can only involve selling software or selling device hardware. The business model primarily based on API interfaces, typical of DeepSeek-style approaches, has not yet obtained an entry ticket to the medical industry.
Price Project Establishment: Without independent medical service output as support, regulators will not establish independent prices for AI diagnosis. In the pricing system, the highest treatment for AI diagnosis is merely an extension item with "no additional markup." The channel for AI diagnosis to create a second revenue curve for enterprises is closed.
Even setting aside the aforementioned policy restrictions, hospitals still find it difficult to pay separately for AI diagnosis because AI diagnosis does not provide new services "from 0 to 1" for departments, but rather represents an intelligent upgrade of existing software capabilities of equipment manufacturers: Hospitals view it as a negotiation chip, while manufacturers view it as a competitive barrier. The true role of AI diagnosis is as a "defensive asset" for manufacturers to maintain market share and price levels, rather than a revenue-generating tool.
The long-term breakthrough for medical AI diagnosis may still require waiting for a DeepSeek moment after more industry accumulation. However, in the short term, strategies could be adjusted to turn "insignificant efficiency gains" into "quantifiable productivity":
Either convert efficiency into revenue, such as significantly shortening report turnaround times and increasing examination throughput, so that AI's contribution directly enters departmental income statements; or go to scenarios where doctors are scarce to become "productivity itself," such as primary care, health checkups, and large-scale screening, where AI is not assistance but the only affordable productivity; or step outside the medical insurance payer framework and look for clients willing to pay for data and services within the budgets of health checkup institutions, pharmaceutical companies, and commercial insurance.
01 Background: Value Regression of AI Diagnosis Hot Stocks
Over the past 28 months, the capital market has used real money to conduct a complete stress test on the "medical AI diagnosis" business narrative of hot stock Tempus.
In June 2024, Tempus went public at $37, experiencing dramatic fluctuations in the first half year. In August, it doubled intraday to $77; in November, it approached $80; in December, it plummeted 46% in a single month.
The real trigger point was January 2025: The purchase of call options by the Pelosi family, known as the "Congressional Stock Gods," was disclosed on January 14, coinciding with the release of the AI health assistant Olivia. On January 21, the stock surged 35.5% in a single day, marking the largest daily gain since listing; on February 14, it touched $91.45 intraday before rapidly retreating to $56.
Subsequently, aided by the $200 million tumor foundation model agreement between AstraZeneca and Pathos, along with consecutive quarterly earnings exceeding expectations, the stock hit a historical high of $104.32 in October 2025, with a market cap of approximately $18 billion. Although positive news such as quarterly EBITDA turning positive and Duan Yongping building positions were disclosed, the stock price continued to fall. Only after gradually disclosing the PRISM2 paper published in Nature Medicine and partnerships with Moderna and Merck in September did the stock price rebound to around $68, still down approximately 35% from its historical high.
02 Current Status of Medical AI Applications: Significant Structural Differentiation Beneath the Glamour
A review of public information indicates that commercial application directions for medical AI mainly involve the following seven categories:
1.Medical Imaging AI: Imaging-assisted diagnosis;
2.Diagnostic AI: Detection-assisted diagnosis;
3.AI Drug Discovery: Target mining, etc.;
4.Medical Large Models: Medical record quality control, evidence-based Q&A, etc.;
5.Surgical Robot AI: Surgical positioning and navigation, etc.;
6.Office AI: Doctor/hospital office assistance, etc.;
7.Patient Chronic Disease Management: C-end health checkup interpretation, home chronic disease management, etc.
The above is merely an application direction perspective based on case classification; switching to the certification perspective yields very different results.
We compiled statistics on medical device registration certificates publicly disclosed from early 2020 to September 2026 with product names or contents containing AI (or Artificial Intelligence), totaling 212 certificates (including 174 Class III certificates)—diagnosis-related certificates accounted for 91%,with 177 applied to imaging diagnostic AI, monopolizing84%, all being software products; followed by in-vitro diagnostic AI applications, with 14 certificates accounting for 7%.
03 Three Commercialization Gates for Medical AI Diagnosis
In the previous chapter, we observed that from a case classification perspective, medical AI has seven tracks, bustling and dispersed; but switching to the "certification" perspective, 177 out of212 registration certificates fall under imaging diagnostics, monopolizing 84%. This 84% is not an accidental technical preference but a result of policy guidance in the medical industry.
To understand this, one must first understand the logical chain of marketing medical products.
The medical sector is an industry with strong policy intervention,and the commercialization path of any medical product cannot escape the standardized path already framed by the system—Productization (Enterprise & Regulator) → Product Certificate (Drug Administration) → Price Project Establishment (Healthcare Security Bureau) (Price project establishment does not necessarily synchronously enter the healthcare security list; entering healthcare security to expand commercialization scale is not elaborated in this article and does not affect the analysis of the industry's underlying logic).
The value of medical AI must be carried by standardized medical products and pass through three gates to achieve scalable monetization in the medical industry:
Productization is the prerequisite for all commercial promotion; medical services without standardized packaging cannot enter subsequent product registration and price project establishment processes (ICL is outside the scope of this discussion);
Only when a product obtains a medical device registration certificate approved by the drug administration can enterprises legally sell to hospitals;
Only when a product is included in the healthcare security bureau's price project establishment guidelines can hospitals legally charge patients and reasonably procure enterprise products.
I. Productization: Auxiliary Product Design Logic Following Vertical Domain Medical Guidelines
In terms of productization, the drug administration consistently upholds the auxiliary positioning of AI diagnosis and the product design principle of "supported by medical guideline logic."
In June 2026, the Center for Medical Device Evaluation of the National Medical Products Administration (NMPA) once again clarified in the "Guiding Principles for Clinical Evaluation Registration Review of AI-Assisted Diagnostic Medical Devices (Draft for Comments)": "(Artificial intelligence) software can only assist physicians in diagnosis... Physicians are responsible for clinical diagnostic results," and "product design must be based on medical guidelines."
The seriousness of medicine does not allow something with ambiguous responsibility boundaries and unclear functional connotations to become a product within the medical system—AI capabilities must be encapsulated as standardized products with clear capability boundaries, controllable liability risks, and repeatable delivery, rather than being a service, a model interface, or a prompt.
Figure: Original Text ① of the "Guiding Principles for Clinical Evaluation Registration Review of AI-Assisted Diagnostic Medical Devices (Draft for Comments)" issued in June 2026, Source: Center for Medical Device Evaluation, NMPA
Figure: Original Text ② of the "Guiding Principles for Clinical Evaluation Registration Review of AI-Assisted Diagnostic Medical Devices (Draft for Comments)" issued in June 2026, Source: Center for Medical Device Evaluation, NMPA
II. Registration Certificate: The Only Two Forms of Standalone Software or Embedded in Medical Devices
In terms of registration certificates, the drug administration has always limited the only two product registration forms—standalone software, or software components embedded in medical devices (where the registration entity remains the medical device).
As early as March 2022, the NMPA's "Guiding Principles for Registration Review of AI Medical Devices" clearly defined the application scope for AI medical device declarations—"including Class II and Class III AI standalone software and medical devices containing AI software components (including in-vitro diagnostic medical devices)." The certified forms of AI products were thus locked into two types: declaring as standalone software, or declaring as part of a medical device with the medical device as the main entity.
Until 2025, when the NMPA released the "Announcement on Measures to Optimize Full Lifecycle Regulation and Support High-End Medical Device Innovation Development," regulatory attitudes softened slightly—"conduct research on management attributes and categories of new functions, technologies, and modalities such as medical large models, AI medical devices, pre-processing software for medical imaging, and process optimization software, and dynamically adjust product management categories timely according to technological maturity."
However, softening attitude does not equal softening principles—according to the latest "Guiding Principles for Registration Review of AI Medical Devices (2026 Revised Edition) (Draft for Comments)" released in September 2026, the registration forms for AI products remain locked into standalone software and embedded components.
Figure: Original Text of the "Guiding Principles for Registration Review of AI Medical Devices (2026 Revised Edition) (Draft for Comments)" issued in September 2026, Source: Center for Medical Device Evaluation, NMPA
III. Price Project Establishment: Extension Item Status Without Independent Charging Qualifications
In November 2024, AI was first written into price project establishment guidelines, but the document explicitly did not grant it the status of independent charging or additional markup.
At that time, the National Healthcare Security Administration (NHSA) simultaneously issued three trial guidelines for medical service price project establishment: "Radiology Examinations," "Ultrasound Examinations," and "Rehabilitation," writing AI content into the establishment guidelines for the first time—all three guidelines uniformly added "AI-assisted diagnosis" as an extension item under the main project, clarifying that "extension items refer to sub-items that expand the applicable scope of price projects without additional charges when provided in different ways or applied in different scenarios under the same project; the price of sub-items follows the main project."
Figure: Original Text of the "Trial Guidelines for Medical Service Price Project Establishment for Radiology Examinations" issued in November 2024, Source: NHSA
Concurrently, the NHSA published an official interpretation article stating the reason: "In the absence of independent medical service outputs and uncertain quality utility of assisted diagnosis, it is inappropriate to charge patients additionally for AI-assisted diagnosis after diagnostic examination fees have already been collected."
Figure: Official Interpretation Article by the NHSA on Radiology Examination Establishment Guidelines in November 2024, Source: NHSA
In December 2025, in the "Trial Guidelines for Medical Service Price Project Establishment for Pathology" released by the NHSA, "AI-assisted diagnosis" was similarly listed as an extension item for pathological diagnosis.
It is worth noting that in the final price guideline for in-vitro diagnostics released in August 2026—the "Trial Guidelines for Medical Service Price Project Establishment for Laboratory Tests"—"AI-assisted diagnosis" was initially listed as an extension item during the draft for comments stage but was removed in the final version, instead being incorporated into the price composition of the main project, losing its status as a separate "extension item"—meaning that from the perspective of price project design, AI-assisted diagnosis failed even to maintain its "extension item status" in broader in-vitro diagnostic projects and was integrated into the basic capability composition of the main project.
Looking back at the "unique position" of imaging AI diagnosis in obtaining certifications, this can be explained by policy logic:
First, the "extension item" status of AI-assisted diagnosis was first confirmed by price project establishment guidelines for imaging examinations; second, all 177 imaging AI registration certificates are software products—because current policy only recognizes two registration forms: standalone software, or software components embedded in medical devices (the latter's registration entity remains the medical device, not AI).
04 Commercial Bottlenecks of Medical AI Diagnosis: Policy Seriousness vs. Market Practicality
After sorting out the policy chain for commercializing medical AI diagnosis, it is evident that the core limitations on the policy side lie in product design forms and charging qualifications; however, even disregarding policy impacts, medical AI diagnosis faces its own bottlenecks in the hospital-side marketization path.
I. Policy Dimension: Restrictions on Product Form and Charging Capability
At the productization gate, the drug administration locks the auxiliary positioning and vertical product design principles of AI diagnosis. The drug administration requires enterprises to design vertical domain AI products based on medical guidelines, even subdividing vertically to specific indications. General-purpose AI capabilities like DeepSeek cannot be directly transferred and applied—regulators need standardized products with clear capability boundaries, controllable liability risks, and repeatable delivery, rather than a service, a model interface, or a prompt.
At the price project establishment gate, the healthcare security bureau prices based on "independent medical service outputs," while AI diagnosis precisely lacks independent outputs and therefore does not possess independent charging qualifications; medical enterprises can only embed AI into medical devices attempting to "increase revenue" via the price barcodes of device hardware.
From a more macro perspective to understand the deep intent of the policy, the policy is not actually "suppressing" medical AI diagnosis—it strictly adheres to the general principle of "people-centered health, clinically value-oriented, controlling the burden of medical expenses on the people" (2021 Eight Departments' "Pilot Program for Deepening Medical Service Price Reform"), steadily promoting technological progress in medical AI. Especially after the anti-corruption campaign in the medical sector in 2023, healthcare security fund income no longer maintained past high-speed growth; in the era of stock resources, the allocation of healthcare security resources must be more prudent.
II. Market Dimension: AI Diagnosis Is Not a New Payment Target, But a Basic Component of Enterprise Competitiveness
Special Note: The following content reflects only the author's personal experience in contacting the real hospital market. Limited by the scope of contact, there may be obvious biases and strong subjectivity; readers are advised to discern carefully.
If the policy dimension answers "why they cannot collect money," the market dimension answers—even if they could, few hospitals are willing to pay extra for AI diagnosis.
First, look at the promotion methods. There are two current commercialization paths for medical AI diagnosis: standalone software sales and embedded medical device sales, with the latter absolutely dominant—the former lacks a charging channel (as described in Chapter 3), so most can only enter hospitals by "piggybacking" on devices. The certificate is for software, but the money comes from devices.
Next, see how hospitals view AI diagnosis. In fact, before AI emerged, imaging and laboratory equipment came with accompanying analysis software—films required software reconstruction and processing, and sample test results required software analysis of basic bioinformatics. Hospitals purchasing equipment never just buy hardware, but an integrated software-hardware solution. After AI appeared, manufacturers upgraded their long-accumulated software analysis capabilities intelligently and applied for Class III certificates for AI software based on this. From a supply perspective, this is undoubtedly an increment in manufacturers' service capabilities; but from the demand side (hospital) perspective, this is not a change from 0 to 1—in the past, buying equipment included analysis software; now manufacturers upgrade analysis software to AI, and hospitals mostly view this as an enhancement of manufacturer service capabilities, rather than a newly added service deserving separate payment.
Combined with the market competition environment, this hospital-side attitude is further reinforced. For mature medical devices, once the industry enters intense competition, price reductions are inevitable, sometimes even vicious price cuts—strategies like giving away devices in exchange for reagent consumption are common. In such a market, hospitals naturally tend to interpret the AI and automation value-added capabilities of manufacturers' devices as service enhancements made by manufacturers to cope with fierce competition and maintain market share and prices. Buyers treat it as a negotiation chip (without AI, I can give the market to manufacturers who include AI), and sellers treat it as a competitive barrier (I followed industry trends to add AI value, allowing me to negotiate to maintain original device prices and defend existing share)—neither party can treat it as an independent paid commodity.
Therefore, from the actual experience on the hospital side, even without policy intervention restrictions, it is difficult to push current AI diagnosis capabilities into a revenue-generating phase in present market negotiations; it can only serve as an important tool for enhancing manufacturer competitiveness, used to maintain current equipment and reagent price levels and avoid excessive price cuts.
05 Immature Thoughts on Breakthroughs
The long-term breakthrough for medical AI diagnosis may still require waiting for a DeepSeek moment after more accumulation in the medical industry. However, there are some referenceable ideas in the short term: turning "insignificant efficiency gains" into "quantifiable productivity."
Either convert efficiency into revenue, significantly shortening report turnaround times and increasing examination throughput, so that AI's contribution directly enters departmental income statements; or go to scenarios where doctors are scarce to become "productivity itself," such as primary care, health checkups, and large-scale screening, where AI is not assistance but affordable productivity for "scaling forward medical gateways"; or step outside the healthcare security payer framework and look for clients willing to pay for data and services within the budgets of health checkup institutions, pharmaceutical companies, and commercial insurance.
Only when AI's value can be presented in reporting language understood by hospitals will willingness to pay shift from "hesitation over nice-to-have additions" to "firm commitment to actively buy."
This is also the other side of the opening sentence of this article: Returning to the business itself, becoming productivity itself, policy and market will both re-price AI.