US export controls were meant to slow China’s AI stack, but the clearest evidence in healthcare points the other way. China’s National Medical Products Administration approved 9 AI-based medical devices in 2020, 45 in 2024, and 154 by June 2025; the portfolio reached a 49.53% compound annual growth rate, 79.9% of those devices were Class III, and only 7 held both FDA and NMPA clearance [2]. That is a deployment curve, not a press release. An older Harvard Business School study of 2007 sanctions found that directly affected Chinese firms increased R&D spending by 49% and patent output by 41%, which is not the same regime but does fit the broader pattern: pressure can redirect investment into capabilities that matter locally [1].

Semiconductor chip constraints transforming into a medical cross and heartbeat waveform

Approval Is the Signal

The approval trajectory matters because it shows where the pressure landed. A market can have research activity without changing clinical practice; approvals are closer to the point where a hospital, vendor, or regulator has to say yes. Here, the movement is fast enough to matter. The count rose from 9 approved devices in 2020 to 45 in 2024 and then to 154 by June 2025 [2].

MilestoneApproved AI-based medical devicesWhat stands out
20209Baseline before the surge
202445More than a fivefold increase from 2020
June 2025154Trajectory implies 49.53% CAGR
Portfolio mix79.9% Class IIIMost approvals sit in the highest-risk category
Cross-border overlap7 dual FDA-NMPA clearancesLimited dependence on Western regulatory validation

The composition of that pipeline is as important as the count. Nearly four-fifths of approved devices are Class III, so this is not mostly low-stakes workflow software or consumer-facing wellness tooling [2]. It is a higher-risk clinical category moving through a domestic approval path. That reduces the relevance of a simple “cut off the chips and the system stops” story. Even when external supply tightens, local demand and local clearance still create a route to market.

The small number of dual FDA-NMPA clearances tells a similar story. If only 7 devices have both approvals, the practical center of gravity is not a shared transatlantic or global pipeline; it is a Chinese one [2]. For hospital procurement, that means the key question becomes whether a tool can clear NMPA requirements and be deployed under domestic constraints, not whether it looks impressive in a benchmark deck.

The Mechanism Is Not Brute Force

The software side of the story is easiest to see in DeepSeek. CSIS describes the company achieving frontier-level performance on downgraded H800 chips through model distillation and PTX-level optimization [3]. That matters because the lesson is not that restrictions became irrelevant; it is that constraints pushed optimization deeper into the stack. Once compute is expensive, waste becomes a strategic liability.

A larger restricted GPU and a smaller efficient chip connected by a compressed neural network transformation

That same discipline travels well into healthcare AI, where the job is usually inference rather than headline-grabbing training runs. Radiology triage, image prioritization, transcription support, and other clinical workflows reward latency, reliability, and deployment efficiency more than raw model size. A model that can be compressed, tuned, and kept responsive is often more valuable to a hospital than one that looks stronger in a lab benchmark but is harder to run.

The hardware counterpart is Huawei’s Ascend line. CSIS estimates the Ascend 910C at roughly 60% of H100 inference performance [3]. That is not parity, but it is enough to keep a domestic pathway alive when top-end U.S. chips are restricted. CSIS also reports testimony that Huawei produced about 200,000 AI chips in 2025, while China legally imported about 1 million Nvidia H20 chips in 2024 alone [4]. The scale gap is real; so is the fact that a local stack exists and is being built out.

For healthcare buyers, that combination is more important than either chip count or model benchmark in isolation. If software is being optimized to run on less, and domestic inference silicon is good enough to support deployment, the procurement conversation shifts. Hospitals do not need absolute frontier performance to deploy a system that reduces reading backlogs, supports triage, or improves throughput.

Why Hospitals Can Still Buy

Demand is not the missing piece. One review projects China’s AI healthcare market rising from $900 million in 2020 to $18.88 billion by 2030, a 42.5% CAGR, alongside $1.4 trillion in government AI investment [5]. CKGSB also cites a radiologist ratio of 1 per 70,000 people in China versus 1 per 7,000 in the U.S. [6]. Those numbers do not prove that every approved product will succeed, but they do explain why there is persistent demand for systems that can take pressure off clinical staff.

That is the part policy commentary often misses. Restrictions do not operate in a vacuum; they land on a system with shortages, procurement needs, and regulatory pathways of its own. In that environment, an NMPA approval is not paperwork. It is the gate that decides whether a local model becomes a deployable hospital product.

Parallel technology stacks for U.S. and China converging into healthcare AI applications

A Parallel Stack Is Emerging

Taken together, the evidence points to restructuring rather than simple delay. Sanctions and export controls have made the easy path harder, but they have also rewarded three things that travel well in healthcare AI: smaller and more efficient models, domestic inference hardware, and a regulatory system willing to clear local substitutes at scale [2][3][4]. That is a different competitive dynamic from the one Washington often describes.

The likely result is not one clean winner but two partially separate hardware-software ecosystems. One still depends on U.S. chip supply and frontier training advantages. The other is being organized around compression, inference efficiency, domestic accelerators, and NMPA approvals. For healthcare AI, that split matters because the products buyers can actually deploy are shaped by the full chain from model design to chip availability to regulatory clearance.

References

  1. How US Trade Sanctions Fueled China's Innovation Surge — Harvard Business School Working Knowledge
  2. Approval of AI-Based Medical Devices in China From 2020 to 2025: Retrospective Analysis — PMC
  3. DeepSeek, Huawei, Export Controls, and the Future of the U.S.-China AI Race — CSIS
  4. Choking off China's Access to the Future of AI — CSIS
  5. Artificial intelligence in Chinese healthcare: a review of applications and future prospects — PMC
  6. AI Applications in China's Healthcare System — CKGSB