A year on from DeepSeek: US versus China in the AI race

NEXT WEEK marks a year since DeepSeek, a Chinese AI start-up, stunned Silicon Valley by unveiling a high-performing large language model at a fraction of the cost of those developed by the US tech giants.

So, to mark the upcoming anniversary, today’s blog post examines this critical question for the global economy and for patterns or locations of chemicals demand which, of course, is linked to GDP growth: Will the US or China win the race to be the global No1 in AI?

Or will the outcome be more nuanced with the US dominant in one area, frontier research, whereas China is better at the “nuts and bolts” of applying AI to real-world applications?

What do you think? Contact me on john.richardson@icis.com

The AI race is stretching into a long marathon

The global AI race is no longer about who builds the single smartest model. It is shifting towards who can deploy AI at scale in the real world. China’s ability to mobilise resources gives it an advantage in this kind of long game.
Electricity is starting to matter more than chips

The US still leads in top‑end chips and training capability, thanks to Nvidia’s latest designs and massive capital flows. But data centres are hitting the limits of the American power grid. Several regions are already close to their reserve‑capacity thresholds, slowing the pace of construction.

China, by contrast, has been building power generation far faster than the US. Analysts think it could have roughly 400 GW of spare capacity by 2030. That would be enough to support a rapid build‑out of data centres, even if Chinese chips are less efficient than US equivalents.

China’s “good‑enough” models are spreading quickly

DeepSeek’s rise showed that strong models can be built cheaply on less‑powerful chips. This lowered the perceived barrier to competing with US AI. It also pushed more attention toward efficiency rather than sheer size.

Chinese open‑weight models, including Alibaba’s Qwen family, have begun to dominate downloads worldwide. Many developing countries prefer these models because they can be run locally and at low cost. This widens China’s influence far beyond its borders.

Europe’s Anti‑Coercion Instrument complicates US diffusion

The EU now has a tool — the Anti‑Coercion Instrument (ACI) — that it can use if it feels pressured by US trade moves. If Washington intensifies its tariff actions against the EU, Brussels could respond in ways that make European buyers more cautious about depending on US AI systems. Even the threat of the ACI may push the European firms toward local or open‑weight models.

We must consider the dispute over Greenland, which has accelerated over the last few days following the deployment of more troops by European countries to the giant island state and the US threat of 10-25% additional tariffs against eight European countries. The EU is considering whether to put the ACI into effect as a countermeasure.

China’s chip capacity is expanding where it matters

China still struggles to match the US in high‑end training chips. But it is improving quickly in inference chips which run models day‑to‑day. Analysts say China could meet its domestic AI chip needs by around 2028.

Meanwhile, the US has allowed limited exports of Nvidia’s H200 chips to China under strict caps. These chips are not the most advanced, but they offer better performance than earlier restricted versions. This gives Chinese firms extra breathing room as local chip capacity grows.

Embodied AI gives China another structural advantage

China excels at embedding AI into physical systems — from robots to public‑service infrastructure. This is partly because China can scale pilot projects quickly and integrate them into manufacturing. Strong research output reinforces this momentum.

China also leads in industrial robotics installation, which puts it in a strong position to adopt AI rapidly in factories. This can increase productivity and drive wider economic uptake. It also allows China to export AI‑enabled industrial capacity abroad.

Scenario 1: The United States pulls ahead

In this outcome, the US fixes its grid bottlenecks through accelerated transmission upgrades and new generation capacity. Trade tensions ease, avoiding ACI disputes. Frontier leadership combines with a faster build‑out of data‑centre power.

Scenario 2: US keeps the lead in models, China leads in deployment

Here, the US continues to dominate frontier research. But slow US power expansion constrains real‑economy adoption. China keeps building generation capacity and data centres, spreading AI into industry and the Global South.

Scenario 3: Two separate ecosystems

Trade frictions escalate. The EU activates the ACI, and Washington tightens chip controls again. China accelerates local chip adoption and rolls out domestic cloud infrastructure, leading to two parallel AI worlds.

Conclusion

If winning means pushing the frontier of model quality, the United States may be declared the winner. But if winning means getting AI into the physical economy fastest, China now has the clearer path.

The swing factors in determining overall, definitive victory include electricity and geopolitics,

The US can come out on top if it fixes its power bottlenecks quickly and if it can avoid geopolitical disputes that lead to trade barriers preventing monetisation of its technology that beats China. China’s risks are whether domestic chips and software efficiency improve fast enough to maintain momentum.

The post A year on from DeepSeek: US versus China in the AI race appeared first on Asian Chemical Connections.