AI slow down: US-China AI rivalry puts global AI governance at a crossroads

As the US and China race to lead AI, safety and governance face a difficult test: can the rest of the world have a say if the two powers set the pace of development?

Anthropic chief executive Dario Amodei has argued that AI development should be paced so that safety measures can keep up with increasingly capable models. But he also recognises the risk of slowing down too much, as it could give other countries, particularly China, an opportunity to close the gap. This captures the larger tension around AI today: how fast should the technology develop, and how much safety should come before speed?

 

Much of this debate is now centred on the US and China, which are competing to stay ahead in AI. If Washington and Beijing begin to see AI mainly as a race for technological advantage, countries outside that contest could have less influence over the rules that shape the technology. And if the two remain dominant in frontier AI, their decisions could eventually influence not only how AI is developed, but also how it is governed across the rest of the world.

 

 

AI is seen as a potential threat as increasingly capable models can act autonomously and behave in unexpected ways. Incidents involving AI systems in tests, including the OpenAI-Hugging Face episode and Anthropic’s Mythos, have highlighted why stronger testing, monitoring and safeguards are needed as AI capabilities grow.

 

AI leadership comes first

 

US President Donald Trump has opposed calls to slow AI development, arguing that the US cannot afford to lose its lead over China. As per Reuters, Trump said the US was ahead of China in AI and stressed the importance of maintaining that advantage, saying that “whoever wins AI wins”. His position reflects a broader concern that slowing development in the US while China continues to advance could change the balance of technological power.

 

Anthropic chief executive Dario Amodei takes a more cautious position. He does not call for stopping AI development, but argues that its pace should be managed so that safety measures and evaluation systems can keep up with increasingly capable models. At the same time, he says the US and other democratic countries need to maintain their lead over China, because falling behind could create its own risks.

 

This leaves Amodei between two competing concerns: moving too fast could increase AI safety risks, while slowing down too much could give China an advantage. That tension is becoming an important part of the wider debate over how AI should be developed and governed.

 

Industry is divided over how much restraint is possible

 

The debate within the AI industry is not only between people who want AI to move quickly and people who want it stopped. Amodei has explicitly said that pacing does not mean halting model training or technical progress. His proposal is to ensure that companies have enough time to improve safeguards and allow independent evaluators to examine whether those safeguards are working. His three-part approach involves embedded third-party evaluators, coordination among AI companies in democratic countries and ultimately, global coordination.

 

OpenAI chief executive Sam Altman has also warned about the consequences of AI advancing without adequate safety measures. Altman identified two major risks: humans losing control of AI systems and excessive concentration of power in the hands of one person or company. He argued that safety and alignment techniques need to stay ahead of model capabilities.

 

But the China question makes restraint much harder. Amodei himself acknowledges that pacing within democratic countries is constrained by the lead that US companies have over China. He argues that if US companies slow down beyond that lead, Chinese projects could move ahead, creating what he describes as a significant national security risk.

 

That means the safety debate cannot be separated from the geopolitical competition. The more important AI becomes to national power, the greater the pressure on governments and companies to keep advancing. 

 

China is closing the gap in ways beyond model scores

 

The US still has important advantages, particularly in computing infrastructure. Fortune recently reported that the US accounts for 74 per cent of the world's compute, according to a White House report, while American hyperscalers continue to invest heavily in data centres. China, meanwhile, has faced restrictions on access to Nvidia's most advanced chips and has had to rely more heavily on domestic alternatives.

 

But measuring the AI race only through access to compute or the performance of individual models risks missing another part of the competition, which is efficiency.

 

Fortune reports that Chinese AI companies have developed methods to extract more value from limited computing resources. Analysts cited by Fortune say Chinese labs have found ways to make parts of the attention mechanism more efficient, reducing the computational burden of AI models.

 

That matters because China does not necessarily have to match the US chip-for-chip if its models can deliver useful results with fewer resources.

 

Fortune also notes data from AI measurement platform Larridin showing that Chinese models such as GLM 5.2 and Kimi 2.6 and 2.7 can handle about 75 per cent of engineering tasks tracked by the platform at around one-fifth of the cost of US models. The report also mentioned that US frontier models retain an advantage on the most complex tasks.

 

The gap is therefore more than who has the best model. It is also about compute, cost, efficiency and how easily models can be deployed.

 

Open models could widen China’s reach

 

Another part of the competition is the availability of open models. Fortune reports that Chinese open-source models accounted for 41 per cent of total downloads on Hugging Face last year, a larger share than US models. The report also points to growing experimentation with Chinese models by companies including DoorDash, Airbnb and Siemens.

 

This gives the Chinese AI ecosystem another route into the global market.

 

DeepSeek’s R1 reasoning model, for example, is available for download through platforms such as Hugging Face, allowing organisations to run and adapt it instead of relying entirely on a closed system. Fortune reports that this flexibility has encouraged some US businesses to experiment with Chinese models. This highlights another dimension of the AI race: influence does not necessarily depend on controlling every major AI application. Relatively inexpensive, adaptable and widely available models can also shape how businesses adopt and use AI.

 

At the same time, Fortune notes that US AI companies remain months ahead in performance. So the picture is not one of China simply replacing US technology. Instead, the two countries are developing different strengths, with the US retaining an advantage at the frontier while Chinese companies narrow the gap through efficiency, cost and open models. 

 

AI race turns geopolitical

 

The competition is also visible in the infrastructure surrounding AI. Amodei argues that the US and its allies should restrict the sale of powerful AI chips and semiconductor manufacturing equipment to China, prevent chip smuggling and remote access to data centres, strengthen security at AI companies and prevent the theft of model weights. He believes these measures could widen the US lead over the next three to five years.

 

This turns AI governance into something broader than rules for companies developing models.

 

It brings semiconductors, computing infrastructure, model access and national security into the same discussion. The consequence is that decisions made by Washington and Beijing can affect the technological choices available elsewhere. If access to advanced chips, computing capacity and frontier models becomes increasingly tied to geopolitical alliances, countries outside the two leading powers could find themselves having to operate within systems largely shaped by decisions they did not make.

 

That is where the question of global governance becomes more complicated.

 

So who gets to write the rules?

 

Amodei's proposal recognises that global coordination will be much harder than coordination among companies or democratic governments.

 

His proposed third stage is global coordination involving the US, other democratic governments and authoritarian governments. But he acknowledges that geopolitical stakes make such cooperation difficult. He argues that any agreement would need strong verification or would have to be limited enough that one side breaking the agreement would not create an existential military advantage.

 

The levels of cooperation he outlines range from relatively narrow agreements, such as prohibiting certain dangerous uses of AI, to testing models for acute risks, establishing a global standards body and potentially setting limits on the speed of recursive self-improvement.

 

A broad agreement to substantially limit the overall rate of AI development, however, is described as much harder to achieve because the incentive to defect could be enormous.

 

This leaves a basic governance problem. If the two leading AI powers cannot agree on how quickly AI should advance, or on how much access should be allowed to the technology, other countries could have limited influence over the direction of the debate.

 

The issue is not that other countries would have no role in shaping AI. Rather, the countries with control over the most advanced models, computing capacity and chips are likely to have the greatest influence over how the technology develops and the rules that govern it

 

Can the rest of the world create a third path?

 

For countries outside the US-China race, the challenge is to have a role in setting AI standards rather than simply following rules decided by others. But the US-China competition could make that difficult. If US companies remain ahead in frontier AI while Chinese models gain ground through lower costs and wider availability, other countries could come under pressure to lean towards one side. This makes AI governance a larger question of who controls access to technology, who sets safety standards and who gets a voice in deciding how AI should be used.

 

The issue of fragmentation could be seen in how Europe seeks to define its own AI path.

 

The European Union is trying to shape its own approach to AI rather than simply follow the US-China race. Its AI Act uses a risk-based framework, with stricter rules for AI systems that pose greater risks to safety and fundamental rights. It also bans certain uses of AI and sets requirements for high-risk systems covering areas such as risk assessment, data quality, transparency, human oversight, cybersecurity and accuracy.

 

The EU has also introduced rules for general-purpose AI models, including requirements related to transparency and copyright. Models considered to pose systemic risks face additional obligations to identify and reduce those risks. The AI Act became applicable on August 2, 2026, although some provisions have later implementation dates. The European Commission says the framework is aimed at promoting safe and trustworthy AI while also supporting innovation and investment.

 

This gives Europe a way to influence how AI is developed and used, even if it is not leading the frontier race in the same way as the US and China. The challenge, however, is whether regulation alone can give the EU enough influence when the most advanced models, chips and computing infrastructure remain concentrated elsewhere. The bloc’s ability to shape standards will therefore depend not only on its rules, but also on its ability to build technological capacity and participate in the wider AI ecosystem.

 

For India, the US-China AI race creates a different challenge.

 

The country is building its AI capabilities and expanding access to computing, but it is not yet in the same position as the two leading powers when it comes to advanced chips, frontier models and large-scale compute. Through the IndiaAI Mission, India is trying to build a stronger domestic AI ecosystem, including wider access to GPUs for startups, researchers, academia and industry. But much of the underlying hardware still comes from global chipmakers, highlighting the gap between using advanced AI infrastructure and controlling it.

 

India’s efforts to develop indigenous AI models and expand domestic compute are therefore important not just for adoption, but for having a stronger voice in how AI develops. If the US and China continue to dominate frontier models, chips and computing infrastructure, countries relying on them could find themselves largely responding to standards and systems shaped elsewhere.