Why Medical AI Diagnosis Struggles to Become an Independent Business: The Three Commercialization Gates Behind 212 Registration Certificates

Over the past 28 months, the market has put the popular "medical AI diagnostics" stocks through a complete stress test: Tempus has pulled back about 35% from its all-time high; the sharp valuation drawdown has prompted us to re-examine the real-world commercial bottlenecks of "medical AI diagnostics."

Based on 212 publicly disclosed AI-related medical device registration certificates in China from early 2020 to September 2026, as well as the original policy texts, this article analyzes the core commercialization bottlenecks of AI diagnostics layer by layer along the three commercialization gates of "productization—registration certificate—price listing."

Productization: Regulators require that the design logic of AI products must be based on medical guidelines and be specific to vertical indication areas. General-purpose large-model products in the DeepSeek style are unlikely to be recognized by regulators, 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 subject remains the medical device), are the only two AI product registration forms recognized by regulators; the commercialization of AI diagnostics can only sell software or sell device hardware. The DeepSeek-style business model, with API interfaces as the mainstream, still cannot obtain a ticket into the healthcare industry.

Price listing: Without independent medical service output to support it, regulators will not establish an independent price for AI diagnostics; in the pricing system, the highest treatment for AI diagnostics is merely an add-on item that is "not subject to additional charges." The channel for AI diagnostics to create a second revenue curve for companies has been closed.

Even setting aside the aforementioned policy restrictions, in the market's "perception," hospitals still find it relatively difficult to pay separately for AI diagnostics, because AI diagnostics do not provide departments with a new "from 0 to 1" service, but rather an intelligent upgrade of device manufacturers' existing software capabilities: hospitals regard it as a bargaining chip, manufacturers regard it as a competitive moat, and the true role of AI diagnostics is a "defensive asset" for manufacturers to protect market share and price levels, rather than a revenue-generating tool.

The long-term breakthrough for medical AI diagnostics may still require waiting for a DeepSeek moment after more accumulation in the industry. In the short term, however, the approach may be adjusted to turn "insignificant efficiency gains" into "quantifiable productivity":

Either convert efficiency into revenue, using it to significantly shorten report turnaround time, increase examination throughput, and so on, so that AI's contribution directly enters the department's revenue statement; or go to scenarios where doctors are scarce and make it "productivity itself," using it in primary care, health checkups, and large-scale screening, where AI is not an assistant but the only affordable productivity; or step outside the medical insurance payer and look to the budgets of health checkup institutions, pharmaceutical companies, and commercial insurers for customers willing to pay for data and services.

01 Background: The value regression of popular AI diagnostics stocks

Over the past 28 months, the capital market has used real money to put the popular stock Tempus's "medical AI diagnostics" business narrative through a complete stress test.

In June 2024, Tempus went public at $37 and saw sharp swings in its first half-year. In August, it doubled intraday to $77, in November it approached $80, and in December it plunged 46% in a single month.

The real trigger was January 2025: the disclosure that the family of Nancy Pelosi, the "Capitol Hill stock god" of the United States, bought call options on January 14, and the release of the AI health assistant Olivia in the same period, sent the stock surging 35.5% in a single day on January 21, the largest single-day gain since its listing; on February 14 it touched $91.45 intraday, then quickly gave back the gains to $56.

After that, thanks to AstraZeneca and Pathos's $200 million oncology foundation model agreement and consecutive quarters of better-than-expected results, it climbed to an all-time high of $104.32 in October 2025, with a market value of about $18 billion. Later, despite disclosure of positive information such as quarterly positive EBITDA and Duan Yongping's building of a position, the stock price continued to fall, and only after the gradual disclosure in September of the PRISM2 paper being published in Nature Medicine and cooperation with Moderna and Merck did the stock price recover to around $68, still about 35% below its all-time high.

Figure: Tempus AI historical closing prices. Source: drawn from public information

The above are only application directions from the perspective of case classification; if we switch to the perspective of certificates obtained, we will see a very different result.

We counted medical device registration certificates publicly disclosed from early 2020 to September 2026 whose product names or product content contain AI (or artificial intelligence), totaling 212 (including 174 Class III certificates)—diagnosis-related certificates account for 91%, with 177 applied to imaging diagnostic AI, taking 84% alone, all of them software products; next is in vitro diagnostic AI applications, with 14 certificates, accounting for 7%.

Figure: Distribution of domestic AI-related medical device certificates, early 2020-September 2026. Source: compiled from public information

03 The three commercialization gates of medical AI diagnostics

In the previous chapter we saw that, from the perspective of case classification, medical AI has seven tracks, lively and scattered; but from the perspective of "certificates obtained," 177 of the 212 registration certificates fall under imaging diagnostics, taking 84% alone. This 84% is not an accidental technical preference, but the result of policy guidance in the healthcare industry.

To understand it, we must first understand the logic chain of market promotion for medical products.

The healthcare track is an industry with strong policy intervention, and the commercialization path of any medical product cannot escape the standardized path long framed by the system—productization (companies and regulators) → product certificate (NMPA) → price listing (NHSA) (price listing does not necessarily mean simultaneous entry into the medical insurance list; the step of entering medical insurance to expand commercialization scale is not elaborated in this article, nor does it affect this article's analysis of the industry's underlying logic).

Productization is the prerequisite for all commercial promotion. Medical services that have not completed standardized packaging cannot enter the subsequent product registration and price listing processes (ICL is outside the scope of this article);

Only when a product is included in the NHSA's price listing guidelines can hospitals legally charge patients and reasonably procure enterprise products.

I. Productization: Auxiliary product design following the logic of vertical-domain medical guidelines

In June 2026, the Center for Medical Device Evaluation of the National Medical Products Administration once again clarified in the "Guiding Principles for Clinical Evaluation and Registration Review of Artificial Intelligence-Assisted Diagnostic Medical Devices (Draft for Comments)": "(Artificial intelligence) software may 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 blurred responsibility boundaries and unclear functional connotations to become a product in the healthcare system—AI capabilities must be encapsulated as standardized products with clear capability boundaries, controllable liability risks, and repeatable delivery, rather than a service, a model interface, or a piece of prompt text.

Figure: Original text ① of the "Guiding Principles for Clinical Evaluation and Registration Review of Artificial Intelligence-Assisted Diagnostic Medical Devices (Draft for Comments)" (June 2026). Source: Center for Medical Device Evaluation, National Medical Products Administration

Figure: Original text ② of the "Guiding Principles for Clinical Evaluation and Registration Review of Artificial Intelligence-Assisted Diagnostic Medical Devices (Draft for Comments)" (June 2026). Source: Center for Medical Device Evaluation, National Medical Products Administration

Regarding registration certificates, the NMPA has always limited product registration to only two forms—standalone software, or a software component embedded in a medical device (the registration subject remains the medical device).

As early as March 2022, the NMPA's "Guiding Principles for Registration Review of Artificial Intelligence Medical Devices" clarified the scope of application for AI medical devices—"including Class II and Class III artificial intelligence standalone software and medical devices containing artificial intelligence software components (including in vitro diagnostic medical devices)." The certification forms for AI products were thus locked into two: application as standalone software, or application embedded in a medical device with the medical device as the main subject.

Until 2025, when the NMPA issued the "Announcement on Measures to Optimize Full-Lifecycle Regulation and Support the Innovative Development of High-End Medical Devices," the regulatory stance loosened somewhat—"carry out research on the management attributes and categories of new-function, new-technology, and new-modality products such as medical large models, artificial intelligence medical devices, medical image preprocessing, and process optimization software, and dynamically adjust product management categories in a timely manner based on technological maturity."

But a loosening of attitude does not equal a loosening of principles—according to the latest "Guiding Principles for Registration Review of Artificial Intelligence Medical Devices (2026 Revision) (Draft for Comments)" released in September 2026, the registration forms for AI products remain locked to standalone software and embedded components.

Figure: Original text of the "Guiding Principles for Registration Review of Artificial Intelligence Medical Devices (2026 Revision) (Draft for Comments)" (September 2026). Source: Center for Medical Device Evaluation, National Medical Products Administration

III. Price Listing: The status of an extension item without independent charging eligibility

In November 2024, AI was written into price listing guidelines for the first time, but the document explicitly did not grant it independent charging or additional markup status.

At that time, the National Healthcare Security Administration simultaneously issued three medical service price item listing guidelines (trial)—"Radiology Examination," "Ultrasound Examination," and "Rehabilitation"—writing AI content into listing guidelines for the first time. All three guidelines uniformly added an "artificial intelligence-assisted diagnosis" extension item under the main item, and clarified that "extension items refer to a type of sub-item under the same item that is provided in different ways or applied in different scenarios, which only extends the scope of application of the price item without additional markup; the price of the sub-item shall be implemented according to the main item."

Figure: Original text of the "Guiding Guidelines for Medical Service Price Items in Radiology Examination (Trial)" (November 2024). Source: National Healthcare Security Administration

During the same period, the National Healthcare Security Administration published an official interpretation article giving the reason: "In the absence of independent medical service output and when the quality and utility of auxiliary diagnosis are difficult to determine, after the corresponding examination fees for diagnosis have already been charged, it is inappropriate to charge patients additionally solely for artificial intelligence-assisted diagnosis."

Figure: Official interpretation article by the NHSA on the radiology examination listing guidelines (November 2024). Source: National Healthcare Security Administration

In December 2025, in the "Guiding Guidelines for Medical Service Price Items in Pathology (Trial)" issued by the National Healthcare Security Administration, "artificial intelligence-assisted diagnosis" was likewise listed as an extension item for pathological diagnosis.

It is worth noting that in the final in vitro diagnostics price guideline released in August 2026—the "Guiding Guidelines for Medical Service Price Items in Laboratory Testing (Trial)"—"artificial intelligence-assisted diagnosis" had been listed as an extension item during the draft-for-comments stage, but was deleted in the official version and instead incorporated into the price composition of the main item, losing its status as a separately listed "extension item"—that is, from the perspective of price item design, AI-assisted diagnosis did not even manage to retain "extension item status" in the broader in vitro diagnostics category, and was instead incorporated into the basic capability composition of the main item.

Figure: Summary of the NHSA's policy stance on "artificial intelligence-assisted diagnosis." Source: Drawn from public information

First, the "extension item" status of AI-assisted diagnosis was first confirmed in price listing guidelines for imaging examinations; second, all 177 imaging AI registration certificates are software products—because policy currently recognizes only two registration forms: standalone software, or a software component embedded in a medical device (the registration subject of the latter remains the medical device, not the AI).

04 The commercialization bottleneck of medical AI diagnosis: the seriousness of policy and the effectiveness of marketization

I. Policy dimension: restrictions on product form and charging capability

Productization gate: the NMPA locks in the auxiliary positioning and vertical product design principles of AI diagnosis. The NMPA requires companies to design vertical-domain AI products in accordance with medical guidelines, even down to specific indications. General-purpose AI capabilities like DeepSeek cannot be directly transferred and applied—what regulation needs is standardized products with clear capability boundaries, controllable liability risks, and repeatable delivery, not a service, a model interface, or a prompt.

Pricing gate: the NHSA prices based on "independent medical service output," yet AI diagnosis precisely lacks independent output and therefore does not qualify for independent charging; medical companies can only embed AI into medical devices and try to "increase revenue" through the device hardware's price barcode.

Understanding the deeper policy intent from a broader perspective: policy is not actually "suppressing" medical AI diagnosis—it merely strictly follows the general principles of "centering on people's health, being oriented toward clinical value, and controlling the burden of medical expenses on the people" (the 2021 eight-department "Pilot Plan for Deepening the Reform of Medical Service Pricing"), while steadily advancing technological progress in medical AI. Especially after the 2023 anti-corruption campaign in healthcare, medical insurance fund revenues no longer maintain the high-speed growth of the past; in the stock era, the allocation of medical insurance resources must be more prudent.

Figure: Historical comparison of medical insurance fund revenue growth and GDP growth. Data source: iFinD

Special note: The following content is only the author's personal impressions from contact with the real hospital market. Limited by the scope of exposure, it inevitably contains obvious biases and strong subjectivity; readers are advised to discern carefully.

First, look at promotion methods. Medical AI diagnosis currently has two commercialization paths: standalone software sales and embedded medical device sales, with the latter absolutely dominant—the former has no charging channel (as discussed in Chapter 3), and the vast majority can only enter hospitals by "riding" on equipment. The certificate is for the software, but the money is for the device.

Next, look at how hospitals view AI diagnosis. In fact, before AI appeared, imaging and laboratory equipment themselves came with analysis software—images taken needed software reconstruction and processing, and sample test results needed software to analyze basic biological information. When hospitals procure equipment, they never buy only hardware; they buy an integrated hardware-software solution. After AI appeared, manufacturers upgraded their long-accumulated software analysis capabilities with intelligence and, on that basis, applied for Class III certificates for AI software. From the supply side, this is undoubtedly an increment in manufacturers' service capabilities; but from the demand side—the hospital's perspective—this is not a from-zero-to-one change: buying equipment in the past already included analysis software, and now manufacturers have upgraded that analysis software to AI. Hospitals mostly regard it as an enhancement of manufacturers' service capabilities rather than a newly added service that should be paid for separately.

Combined with the competitive market environment, this hospital-side attitude is further reinforced. For mature medical devices, once the industry enters fierce competition, price cuts are inevitable, and even vicious price cuts occur—tactics such as giving away equipment or exchanging reagent consumption are common. In such a market, hospitals naturally tend to understand the AI and automation value-added capabilities of manufacturers' equipment as service value-adds that manufacturers provide to cope with fierce competition and preserve market share and prices. Buyers treat it as a bargaining chip (if you don't include AI, I can give the market to a manufacturer that does), while sellers treat it as a competitive barrier (I followed the industry rhythm and added AI value; I can negotiate again to maintain the original equipment price and hold on to my original share)—neither side can treat it as an independently paid commodity.

Therefore, from the actual hospital-side experience, even without policy intervention restrictions, it is very difficult at this stage to push current AI diagnostic capabilities into a revenue-generating stage in market-based negotiations; it can only become an important tool for enhancing manufacturers' competitiveness, used to maintain the price levels of existing equipment and reagents and avoid excessive price cuts.

05 Immature thoughts on breaking the deadlock

The long-term breakthrough for medical AI diagnosis may still need to wait for a DeepSeek moment after the healthcare industry accumulates more, but in the short term there are also some ideas worth considering: turn "insignificant efficiency gains" into "quantifiable productivity."

Either convert efficiency into revenue—significantly shorten report turnaround time and increase examination throughput, so that AI's contribution directly enters the department's revenue statement; or go to scenarios where doctors are scarce and become "productivity itself"—primary care, health checkups, large-scale screening—where AI is not an assistant but affordable productivity for "moving the medical gateway forward at scale"; or step outside the medical insurance payer and look to the budgets of health checkup institutions, pharmaceutical companies, and commercial insurers for customers who will pay for data and services.

This article is written based on public information and is for information exchange only; it does not constitute any investment advice.

This article comes from the WeChat public account "Jinduan" (ID: jinduan006), author: Albert_L, published by 36Kr with authorization.