The AI race is becoming a race to adapt

When the leaders of the AI corporate world gathered at the White House September 29, the immediate story was about information safety. The companies—OpenAI, Anthropic, Meta, Google, Nvidia and xAI—signed a voluntary agreement covering internal monitoring, external evaluation and stronger oversight of frontier AI systems.

It was not legislation (the Financial Times described the agreement as a new form of AI self-regulation), and questions remain about how much practical constraint a voluntary accord will impose. But the meeting revealed something more consequential: the institutions surrounding artificial intelligence are now having to evolve almost as quickly as the technology itself.

For Asia, and particularly China, this matters because the global AI competition is entering a different phase. The first phase was dominated by a relatively simple question: who could build the most powerful models? Increasingly, that is no longer enough.

The technological frontier is fragmenting across models, autonomous agents, semiconductors, energy, data centres and applications. At the same time, governments are confronting questions of safety, cybersecurity, liability and public trust. Competitive advantage is therefore becoming a function not simply of technological capability, but of how quickly firms and countries can adjust to the wider challenges surrounding it and maintain their social licence.

China presents perhaps the most interesting contrast with the United States. American AI leadership has largely been associated with extraordinarily well-capitalised frontier firms such as OpenAI, Anthropic, Google and Meta. China has faced greater constraints, particularly regarding access to the most advanced semiconductors. Yet those constraints have encouraged a different pathway: cost-efficient models, wider deployment and an increasingly important open-weight ecosystem. Recent Chinese models from companies including DeepSeek, Alibaba and Moonshot have reinforced the importance of this open-model pathway.

The distinction between open and closed models therefore has geopolitical and commercial consequences. Closed frontier models allow firms to retain greater control over technology, monetisation and deployment. Open-weight models sacrifice some of that control but can spread much faster. They can be downloaded, adapted, locally deployed and incorporated into applications across different industries and countries. This does not mean giving up commercial value. Rather, it can shift where that value is captured, from charging primarily for access to the model itself towards cloud and computing services, enterprise integration, specialised applications, proprietary data and the wider hardware and software ecosystems built around the model.

For Chinese firms, widespread adoption can therefore expand the market for complementary products and services even when access to the underlying model is relatively open. This creates an alternative route to technological influence. China does not necessarily have to displace American firms at the absolute frontier. If Chinese models become widely embedded in applications, developer communities and enterprise systems across Asia and other emerging markets, China can exercise influence through diffusion rather than domination.

This is particularly relevant because the economics of AI may be shifting. Frontier-model leadership will remain strategically important, particularly in areas such as defence, intelligence and scientific discovery, where superior capabilities can confer disproportionate advantages. But technological supremacy and economy-wide value capture are not necessarily the same thing.

As capable models proliferate and prices fall, possessing the best model at a single point in time may become less decisive. Value can migrate towards the complementary capabilities around models: computing infrastructure, energy, proprietary data, distribution, industry applications and the ability to integrate successive generations of technology.

That was one implication of my argument that Asia does not have one route to AI advantage. Singapore, Korea, China and India occupy different positions across the AI stack. But there is now an additional dimension: it is not enough to occupy an advantageous position in the stack today. Countries must be able to deepen existing strengths, develop complementary capabilities and move towards activities where strategic and commercial value is emerging as the technology evolves.

This is what I mean by capability advantage: being better positioned after one transformation to undertake the next. Adaptability and improvisation in the face of adversity are part of how such an advantage can develop. China’s open-weight strategy can be understood in those terms. Restrictions on advanced chips created a constraint. Chinese firms responded by putting greater emphasis on efficiency, alternative hardware, domestic supply chains and model diffusion.

Whether this ultimately closes the frontier-model gap is uncertain. What matters strategically is that adaptation to one constraint can create capabilities useful for confronting the next.

The United States faces a different adaptation problem. Its firms remain at the technological frontier, but increasingly powerful AI systems create new demands around safety, legitimacy and governance. The White House accord represents one attempt to address those complementary requirements without substantially slowing technological development. Critics question whether voluntary commitments are sufficient, but the underlying problem is real: maintaining technological leadership increasingly requires institutions capable of moving with the technology.

The contrast between the two systems should not be reduced to a simple choice between American closed models and Chinese open models. Both ecosystems are becoming more hybrid. Meta, for example, spans open-weight and more controlled approaches, while Chinese firms also retain proprietary capabilities. The important strategic question is how openness, control, scale and diffusion are combined. In Asia, countries adopting foreign AI systems will increasingly have choices between powerful closed models, capable open-weight alternatives and combinations of the two.

For governments across the region, this changes the policy problem. The question is not simply which model or supplier to choose, but what domestic capability is being built around those choices. Skills, computing infrastructure, energy, industry applications, regulatory capacity and trusted international partnerships will determine whether adoption creates dependence or becomes a platform for further capability development.

Australia has a particular interest in this shift. It is unlikely to lead the frontier-model race, but it can still become more valuable within the regional AI system by developing selected capabilities and by remaining connected to both technological and institutional change across Asia. That means understanding not only what the United States is building, but also how China is diffusing AI capability through open models, industrial ecosystems and increasingly capable domestic firms.

The next phase of AI competition may therefore look quite different from the first. The countries that matter most may not simply be those possessing the strongest model at a particular moment. They may be those that become best at absorbing one technological wave, learning from it and repositioning themselves for the next. That is where the deeper contest for AI capability advantage is beginning.

Vikas Kumar is Professor of International Business in the University of Sydney Business School.