In a reading room in Nanning, a radiologist opens a study that was taken somewhere else.
The patient had been scanned at another hospital, and those images were not in front of the doctor deciding now.
Then the earlier scan appears on the screen, pulled from a provincial store rather than a film bag.
The software is the least interesting part. What matters is who paid for the store, who made the hospital fill it, and what it costs not to.
Key takeaways
- China's medical-insurance fund held nearly 440 million imaging index records covering 31 provincial-level regions and the Xinjiang Production and Construction Corps at end-June 2026.
- The lever is the price list: a cloud image is priced at RMB 5 (≈US$0.70) per examination, and hospitals lose RMB 5 per body part for images not uploaded.
- Five provinces have tendered provincial image storage; six more plus the XPCC are starting, with the fund buying storage as a service.
- Guangxi's platform covered 1,780 institutions across 14 cities, holding 17.76 million studies and 3.47 million cross-hospital retrievals, by 5 November 2026 (Guangxi Healthcare Security Administration).
- A national imaging-AI contest drew 1,332 teams; the argument now is that retrieval rates, not model scores, decide whether the build worked.
A price, not a purchase order
The medical-insurance fund (医保基金) is not a hospital operator and cannot tell a hospital which software to install. It can define what an examination includes — and therefore what a hospital is paid for.
In December 2024 the National Healthcare Security Administration folded image processing, upload and cloud storage into the price of a radiology examination. Provincial bureaus then bought storage through tender at a settled RMB 5 per examination.
The fund does not build the storage itself. It creates a recurring revenue line for whoever stores the images and lets the tender pick the operator. It calls the five yuan a way to draw outside capital into storage and form "a closed commercial loop" (NHSA, 9 July 2026).
Then comes enforcement. By September 2026 the administration had completed 40 batches of pricing guidelines, spanning radiology, pathology, auxiliary procedures, general treatment and laboratory work. In those batches, uploading results is written into the price, and a hospital that does not upload is paid less: for CT and MRI, RMB 5 per body part (NHSA, September 2026).
Revenue from a scan now depends on the scan leaving the building in standard form.
What was built underneath
In February 2025 the administration issued an index for cloud images: a unique identifier per study, uploaded to the national platform, so a record can be found even when the image sits in a provincial store. In October 2024, 20 provinces had deployed the software with about 70 million records; by end-June 2026 it was 31 provincial-level regions plus the XPCC and nearly 440 million (NHSA, 9 July 2026).
Retrieval came next. On 20 November 2025 the administration switched on cross-province retrieval; at Peking University Third Hospital, clinicians called up images of five insured patients from Sanming, Xiong'an, Yan'an, Zhengzhou and Guyuan.
Storage is bought province by province. Guizhou, Xinjiang, Guangxi, Ningxia and Hubei have completed tenders and moved to provincial centralised storage; Tianjin, Inner Mongolia, Hainan, Sichuan, Tibet, Gansu and the XPCC are starting (NHSA, July 2026). The centre sets the price and the identifier; provincial bureaus run the procurement.
The province used as the demonstration
Guangxi is the demonstration. Its platform covers all 14 prefecture-level cities, 1,780 institutions, 17.76 million uploaded studies and 3.47 million cross-hospital retrievals, at three-second latency, with tertiary-hospital coverage above 90% (Guangxi Healthcare Security Administration, 5 November 2026).
A three-year conversion plan commits to at least 100,000 standardised imaging cases for high-incidence cancers, at least five datasets built to national standards, at least ten pilot institutions, and up to RMB 3 million (≈US$420,000) per awarded team.
The decisive line sits elsewhere: Guangxi intends to push mature AI diagnostic items into the provincial price catalogue, with a charge and an insurance payment rule.
That is the gate. A model can win a national contest and still earn nothing: a hospital cannot bill for a service with no price code. Approval, evidence and a tariff are three separate permissions, and the third one is provincial.
The metric nobody advertises
The fund ran the competition that produced the candidate models. The 2026 national imaging-AI competition drew 1,332 teams and 4,447 participants from 30 provincial-level regions, Hong Kong, Macao and several ASEAN countries, across eight disease tracks; 66 teams reached the final and 193 awards were given (NHSA Big Data Centre, 5 November 2026).
What the fund wants next is more modest than a leaderboard: assisted diagnosis, health management, claims administration.
Liu Qingfeng, chairman of iFlytek, which built the Guangxi platform, argues that the measure should be the retrieval rate — how many unnecessary repeat examinations were avoided — and that the binding problem is workflow, not accuracy. A query tool that needs a separate page will not be opened during a consultation; a prompt when a doctor orders a scan will be (People's Daily, 6 November 2026). That is an industry argument, not a policy target; no national retrieval rate has been published.
When the payer owns the identifier, buys the storage and runs the competition, a model is a component inside a state-run pipeline rather than a product sold to hospitals.
Honest limitations
The nearly 440 million figure counts index records, not images; no national stored volume or retrieval rate is published. The RMB 5 price is a procurement outcome executed by provincial bureaus, not a national tariff. Guangxi's numbers are self-reported by its own insurance bureau and are not independently audited. The competition dataset is called open, but its access terms are unpublished. The retrieval-rate argument belongs to a vendor executive, not the administration.
What readers can do now
- For analysts, follow provincial storage tenders rather than the contest. Five provinces have awarded contracts and six more plus the XPCC are starting.
- For vendors, ask a provincial insurance bureau whether an AI diagnostic item has a price code yet. Without one, a hospital can adopt the tool but cannot charge for it.
- For newsrooms, do not report nearly 440 million as images stored. It is a catalogue of index records pointing at images held elsewhere.
