The China-U.S. AI race: Who is leading?

By Jasmine Duan

Investors may still remember the “DeepSeek Moment” in Jan. 2025, when a small Chinese AI startup launched an open-source reasoning model that matched the performance of leading U.S. AI models. Since then, China has rapidly narrowed the gap in AI development with the U.S., intensifying debate about who will lead this technology race.

Chinese AI models reach tier-1 performance

Despite facing intensifying hardware constraints, Chinese AI models are demonstrating rapid improvements in coding and reasoning capabilities. A recent example is Chinese startup Moonshot’s Kimi K3 model, whose overall performance nearly matches that of advanced U.S. models at a fraction of the cost.

Chinese companies now account for the largest share of generative model downloads on Hugging Face, the primary hosting platform for AI models. Rather than isolated breakthroughs, Chinese developers are releasing high-performing models at a weekly cadence. New releases consistently rank among the platform’s most downloaded each week, which is a sign of both quality and developer productivity, in our view.

Not all Chinese AI models adopt a low-price strategy

Chinese models have gained popularity partly because they are open source, allowing developers to customize and train them on proprietary data, and partly because they are significantly less expensive than closed-source alternatives. This has led many investors to assume Chinese firms could only compete on price. However, we argue that this assumption is being tested.

Chinese AI companies are now pursuing differentiated strategies. Moonshot’s Kimi focuses on offering higher-intelligence models at premium prices. General language models from Qwen and Z.AI adopt two-tier approaches, pairing advanced, expensive models with efficient, cost-effective options. DeepSeek, meanwhile, sticks to a low-cost strategy.

At the same time, U.S. providers are responding with their own price cuts, particularly for lower and mid-tier models. We view this as a healthy market strategy for AI companies: lower-cost models expand accessibility, while premium options serve enterprises’ most complex needs.

China’s biggest constraint: semiconductors

As AI demand surges globally, Chinese companies face mounting supply constraints on semiconductors. While previously reliant on stockpiles, legacy chips and gray-market purchases to train models, these workarounds are no longer sufficient as compute needs have accelerated.

We believe the divergence in investment capacity reveals the severity of the constraint. According to Bloomberg, the eight largest U.S. hyperscalers are projected to increase capital expenditure by 76 percent in 2026, reaching over US$780 billion. By contrast, China’s three major internet companies and three telecom carriers combined plan only roughly $85 billion in data centre capex this year, which is roughly one-tenth of the U.S. total. Even this amount represents a stretch, as memory chip prices have quintupled over the past half-year, further eroding purchasing power.

AI spending in China is a fraction of the U.S. total

More concerning for Chinese AI firms, in our view, is that U.S. export controls prevent China from accessing advanced manufacturing equipment, particularly extreme ultraviolet (EUV) lithography machines needed for cutting-edge chip production. This creates a potential structural disadvantage. While reports periodically surface of Chinese breakthroughs in developing EUV lithography, we doubt these will have meaningful near-term impact given their prototype status and the challenge of reaching mass production.

Chinese chipmakers are responding by stacking multiple chips in smaller spaces, improving interconnects and pursuing system-level optimization to narrow the performance gap. We believe these efforts may not fully offset supply constraints, but they represent the most viable path forward in the near term.

Where will AI competition go next?

For many people, what matters most is how AI will affect their livelihoods, not which country will win or lead the AI competition. Public perception of AI diverges sharply between China and the United States. According to Stanford University’s AI Index Report in 2025, 84 percent of Chinese respondents reported excitement about AI, compared to just 38 percent of Americans. Meanwhile, American concerns centre on job displacement, while Chinese concerns are more varied.

We think this divergence partly reflects structural economic differences. The U.S. Bureau of Labor Statistics reports that more than 70 percent of American workers are in service sectors—which are more vulnerable to AI disruption—compared with approximately 50 percent in China, according to Chinese government data sources. AI’s immediate workforce impact therefore seems smaller in China, explaining why Chinese people view it as a practical tool to increase productivity rather than an existential threat. Furthermore, a recent Chinese court ruling protecting a worker in an AI-related labour dispute signalled the country’s regulatory commitment to labour-rights protections.

Beyond cultural attitudes, the divergence also reflects differing levels of trust in government policymakers and in big technology companies, and belief in personal benefit from the technology—with Chinese attitudes ranking higher than Americans’ views. We believe these perceptual differences could help shape each country’s AI development trajectory.

The China-U.S. AI competition’s outcome remains uncertain, with the competitive landscape shifting rapidly as new models emerge. We think a few factors will be critical differentiators for China’s long-term AI development relative to the U.S., including abundant engineering talent, reliable and affordable energy supplies and capacity for large-scale industry applications.

With the contribution of Leo Shao.

---

---