China: Late to the AI sprint, strong in the marathon

While slower off the mark, China is not necessarily late to AI as a structural transformation, but its path is unlikely to mirror the US model.

Rather than being evaluated solely in the race for the most powerful chip or the highest-scoring model, China’s opportunity may lie in “full-stack” competition across the entire AI value chain:

• Capable domestic models
• Alternative AI accelerators
• Microchip design
• Foundry capacity
• Memory manufacturing
• High-speed connectivity
• Power infrastructure
• Software optimisation
• Large-scale deployment

The distinction matters because the next phase of AI may be less about benchmark leadership and more about usable, affordable and scalable deployment. The winners may not simply be those with the best single chip or model, but those who can transform computing into productivity at the lowest practical cost.

In short, China may not be copying the US playbook. Rather it may be building a different one with a broader and more diversified set of investable winners.

Perhaps an imperfect but useful analogy: In the 4G mobile network cycle, China was not viewed as the original technology leader. Yet once adoption in China accelerated, the country’s infrastructure rollout, local applications, mobile payments, e-commerce and hardware ecosystems created one of the world’s most dynamic digital economies. China did not merely catch up; it helped redefine how mobile internet could be commercialised at scale.

AI is more capital-intensive, power-constrained and geopolitically sensitive than 4G. But the lesson still holds: China’s strength often emerges less at the point of invention than at subsequent deployment.

This is why the attitude of financial markets towards China AI-related equities appears cautious, in our view. Taiwan and South Korea’s AI beneficiaries have been rewarded for their clearer links to advanced chip manufacturing and memory demand.

By contrast, global investors own less of many China AI and semiconductor-linked companies and they often trade at a discount, despite exposure to domestic substitution, data-centre build-out, cloud capex, optical connectivity, power management and AI infrastructure demand.

The discount is not proof of mispricing, but it does create asymmetry if China’s full-stack AI strategy gains credibility.

Full-stack matters more than the fastest chip

AI remains embryonic. The world is still in the infrastructure build-out phase. Enterprises are only beginning to integrate generative AI into workflows, while industrial AI, robotics, autonomous driving, healthcare, financial automation and smart manufacturing are still far from mass adoption.

The key question is not whether AI has already ‘happened’. It is about which ecosystems can convert AI from a technology theme into productivity, profitability and strategic capability over the next decade.

China’s semiconductor ecosystem has three demand drivers. The first is global AI capex, where Chinese suppliers can participate through components, power, connectivity, equipment, materials, testing, assembly and broader hardware supply chains.

The second is domestic AI capex, as Chinese cloud platforms, telecoms operators, internet companies, model developers and industrial champions build AI infrastructure.

The third – and most strategic – is self-reliance. US technology restrictions have not removed China’s desire to develop advanced semiconductors; they have intensified it.

The constraints imposed by the US are real. China still lacks full access both to Extreme Ultraviolet (EUV) – a photolithography technology used in the world's most advanced microchips – and to leading-edge semiconductor manufacturing tools. China cannot simply follow the same path of shrinking transistors to close the performance gap.

But the AI infrastructure race is increasingly moving from a single-chip to a full-system contest. Performance now depends on how efficiently chips, memory, networking, power, cooling and software work together at rack and cluster scale.

The bottleneck is not always the chip. It can be memory bandwidth, interconnect, communication latency, power efficiency, cooling, software stack, data availability or deployment cost.

Chinese standout firms point to a different path

One of China’s leading tech firms has introduced the clearest example of this alternative path, which is critical in the context of China’s EUV constraints making the traditional route of simply shrinking chips harder.

The company’s Tau Scaling Law shifts the focus from transistor geometry to system delay using LogicFolding, software-hardware co-design and faster interconnects to make chips and systems work more efficiently together.

The same logic applies to AI infrastructure. While perhaps not matching the best chips in the US one-for-one, the strength of a Chinese company in connectivity could allow it to link more chips together, improve rack-level performance and narrow the usable performance gap.

The point is not only that China can produce capable models, but that model efficiency can reduce dependence on brute-force computing. Inference performance can improve substantially after launch through software tuning, kernel optimisation, memory handling and serving-framework improvements. In plain terms, AI performance is not fixed on day one. It can exponentially improve through engineering.

This matters for China. If domestic models can lower memory intensity and inference cost, and if domestic hardware platforms can improve through better networking and software optimisation, China does not need to win the AI race on the US’ terms. It can compete by building good-enough, cost-effective AI systems at scale. Cheaper, capable models are not second-tier products if they unlock broader adoption. AI does not scale if it is too expensive to use.

The China AI trade is no longer single-dimensional

The investment implication is therefore not a simple catch-up trade. China’s AI opportunity is broadening from headline models and chips into a wider ecosystem of domestic computing, semiconductors, connectivity, power, data centres, cloud platforms and industrial applications.

A more useful framework is selective exposure to a multi-year broadening-out theme, where re-rating depends less on the narrative and more on earnings delivery, capex visibility, domestic substitution and evidence of adoption. China’s AI progress should be viewed as a marathon, not a sprint.

That is a different race — and one that investors may be underestimating.