By Lin Jianlan
Inside the gleaming atrium of Nanfang Hospital in Guangzhou, an experiment is underway that could rewrite the economics of medical artificial intelligence. For years, China’s top-tier hospitals have rushed to deploy AI – lung nodule detectors in radiology, cell screeners in pathology, risk models in nephrology. Each arrived with fanfare. Each solved a small problem. And each, in its own way, made the larger one worse.
The industry has reached a paradoxical state, according to dozens of clinicians and technology executives interviewed for this article. Hospitals are willing to invest. Vendors are eager to supply. Doctors genuinely want to use AI. Yet truly sustainable, scalable, continuously improving systems remain rare.
The culprit is what engineers call “stovepipe” architecture – isolated AI models purchased department by department, each with its own data pipeline, compute allocation, and user interface. A hospital might buy five high-accuracy models from five vendors. None speaks to the others. Radiology cannot share findings with pathology. The physical examinationcenter’s ultrasound AI cannot talk to the electronic patient record department.
“You buy many AIs, but you end up with none you can use well, use thoroughly, or use for long,” lamented a senior IT manager at a Beijing tertiary hospital, who declined to be named discussing sensitive procurement matters.
Huawei to the rescue
The waste is staggering. Compute resources sit idle in silos. Data from the same patient – text records, CT images, pathology slides – remain segregated. Doctors toggle between half a dozen systems. And when a model needs updating, the original vendor must return at premium rates to patch the interface.
This is the moment Huawei, the Chinese telecommunications and technology giant, has chosen to pivot. Rather than chasing another single-point diagnostic model, the company spent the second half of 2025 embedded inside Nanfang Hospital, conducting 284 interviews across 24 departments, cataloguing more than 100 distinct AI demands.
Their conclusion: hospitals do not need another application. They need an operating system.
Enter HAIP – the Hospital AI Platform. Launched at Nanfang Hospital and detailed in a white paper published April 10 in partnership with leading institutions, HAIP is not a diagnostic tool. It is the underlying layer that manages compute, data, models, and applications across the entire facility. Think of it as a common railway, allowing any AI train to run on the same tracks.
The early metrics are compelling. By implementing a “day-inference, night-training” scheduling system – prioritising real-time diagnostic support during clinic hours, then automatically switching to model retraining overnight – HAIP lifted overall compute utilisation by 30%.
Unified hospital-wide AI systems
More striking is the impact on pathology. Previously, a pathologist might spend an entire shift annotating a single cervical cytology slide. With HAIP’s unified data platform, which harmonises text records, imaging, pathology screening, and even electrocardiogram strips, the AI can now pre-mark suspicious regions for review. Throughput has jumped from 50 to 300 slides per person per day – a sixfold increase – with annotation accuracy exceeding 85%.
Perhaps the most significant innovation is what Huawei calls “natural language assistants”. A clinician can simply describe a need – “build an assistant that organises follow-up records for lung nodule patients and alerts me when a follow-up examination is due” – and the platform generates and deploys a working AI agent within minutes. The development cycle has shortened by 70%.
The white paper, Hospital General AI Platform Technical White Paper, represents the first systematic attempt to define technical standards for what is effectively a hospital AI operating system. Critically, it requires any vendor’s AI to adhere to a common interface standard to prevent hospitals becoming locked into a single vendor’s eco system as healthcare systems digitize more core functions.
“The foundation, once built, may end the era of vendor lock-in,” one hospital executive involved in the project said. “The platform approach is about openness, not exclusivity.”
There are, of course, caveats. The system has only been validated at one hospital. Scaling across China’s fragmented, unevenly digitised healthcare landscape will require years of testing. The upfront investment is substantial, and the coordination challenges – aligning departments, vendors, and regulators – are formidable.
Yet the direction is clear. As healthcare systems worldwide grapple with how to move from pilots to platforms, the HAIP experiment suggests a path forward that prioritises shared infrastructure over point solutions. It is a longer, harder road. But if it succeeds, it will change not just a single aspect of a hospital’s operations, but the entire way the medical AI industry is built.
Source:
LatePost
