In an office outside Beijing filled with robotic hands of varying sizes, floating on cables and flexing their fingers, the founder of one of China's leading dextrous hands companies offers a comparison that ought to unsettle anyone charting the UK's approach to health technology. Making a functioning robotic hand, Alex Zhou, the founder of Linkerbot says, is a hundred times harder than making the humanoid body that carries it. The dexterity required is ten times greater than for any other part of the machine, yet the available volume is a tenth of the size. It is, in engineering terms, the wrong end of every trade-off at once.

That imbalance is the real story here, and it has little to do with China's ambitions for elder care robots or household helpers, which remain years away by the industry's own admission. The more instructive point is about where genuine technical difficulty sits in AI development, and where it does not. The almost infinite amount of content on the internet is used to train large language models. Robotic hands must be trained on data that barely exists: the pressure of a fingertip on an eggshell, the feel of a shoelace tightening, the thousands of hours of teleoperation needed to teach a machine to pack a bag of groceries without crushing it. Chinese firms are now sending workers out with sensor-laden gloves simply to capture what a human hand already knows. It is slow, physical, unglamorous work, and it cannot be shortcut by scaling a model.

Britain's own AI strategy for the NHS has, almost without exception, avoided this kind of problem. The ambient scribe rollout, the enthusiasm around agentic AI accountability now occupying the MHRA, the continued reliance on platforms such as Palantir's Federated Data Platform, all sit at the software layer, where the underlying architecture is largely bought rather than built and the marginal cost of deployment is low. This is not a criticism of the technology itself, much of which delivers genuine value in reducing administrative burden on clinicians. It is an observation about strategic choice. Software-layer AI is attractive to a health system under fiscal strain precisely because it is fast, comparatively cheap, and politically uncomplicated to procure. Hardware-layer capability, of the kind China is now building at scale through firms such as LinkerBot and Wuji Technology, demands the sort of patient industrial commitment, manufacturing depth, and years of accumulated component expertise that Britain has shown little appetite to fund in life sciences or medical technology since at least the pandemic.

The data problem is where the parallel sharpens further. NHS leaders have spent several years insisting that clinical AI's limiting factor is data quality and availability rather than model sophistication, and they are right. Real-world, consented, well-structured clinical data is scarce in precisely the way that physical manipulation data is scarce for robotics, and for similar reasons: it cannot be scraped, it must be captured, and capturing it properly requires sustained investment in infrastructure that produces no immediate headline. China's response to its equivalent problem has been to treat data capture as core industrial policy, not an afterthought bolted onto a procurement contract. The UK's data strategy, by contrast, has repeatedly been reduced to a governance and consent argument, important in its own right but not a substitute for the harder work of building the systems that generate usable data in the first place.

None of this means Britain should attempt to compete with China's manufacturing base, an ambition that would be neither credible nor sensible given the UK's comparative strengths. But it is worth being honest about what has actually been chosen. A health system can adopt AI quickly by buying the layer that is easiest to procure, or it can build the layer that is hardest to replicate and therefore hardest to lose. Those two paths look similar in a press release. They diverge completely a decade on, and only one of them leaves anything durable behind.