How to assess China’s real chance of winning AI race against US

It’s a mistake to treat the AI race as a contest with one finish line.

Photo: Shutterstock In January, a top Chinese AI researcher told an industry summit in Beijing there was less than a 20 per cent chance of any Chinese company surpassing a leading US artificial intelligence firm in the next three to five years.

The remark by Lin Junyang, until recently a technical leader working on Qwen, one of China’s most capable open-source AI models under Alibaba (which owns the South China Morning Post), made headlines.

But much of the commentary missed a more important question Lin posed: “Does innovation happen in the hands of the rich or the poor?” The United States held an estimated 74 per cent of global AI computing power in mid-2025, compared with China’s 14 per cent.

Lin described the gap as “one to two orders of magnitude”.

Because US labs command far more aggregate compute, they can allocate substantial capacity to next-generation research as well as product deployment.

Chinese labs, he admitted, are “stretched”: just delivering products consumes most of their compute.

The luxury of exploration is one they simply cannot afford.

Lin’s 20 per cent remark provoked two broad reactions.

Optimists dismissed it: benchmark gaps have shrunk to near parity, Chinese models last year claimed nine of the top 10 open-weight positions on a major leaderboard and export controls have plainly failed to stop progress.

Pessimists, however, thought a 20 per cent chance was generous: Huawei Technologies’ chip output is projected at 2-5 per cent of Nvidia’s aggregate AI computing power through to 2027, and DeepSeek founder Liang Wenfeng has admitted “money has never been the problem” while “bans on shipments of advanced chips are”, acknowledging a fourfold compute disadvantage.

It is only a matter of time, this camp suggests, before the deficit becomes insurmountable.

Both reactions miss the point.

They treat the AI race as a contest with one finish line.

In reality, compute scarcity and compute abundance have produced two structurally different innovation models – each with distinct strengths, blind spots and implications for governance.

Understanding this divergence is key to knowing what Lin’s 20 per cent really means.

Start with scarcity.

When every training run is prohibitively expensive, the tolerance for failure collapses.

You cannot afford to explore 10 architectural ideas hoping one works; you must optimise the one you have.

You prove your model on es

原文链接: 南华早报