While media attention remains focused on OpenAI, Anthropic and Google DeepMind, another battle is unfolding within the artificial intelligence industry. It is not about the largest funding rounds or record valuations, but is taking place in the deepest layers of the ecosystem: those that will determine which models are actually used by developers, companies and future generations of AI agents.
For the past two years, the dominant narrative has framed the market as a race for power among American laboratories. OpenAI, Anthropic, Google DeepMind and xAI are competing to build the most capable models, pushing the boundaries of reasoning, programming and scientific research.
Yet as these models become increasingly expensive to train and operate, a parallel dynamic is accelerating. A new generation of open models, many of them developed in China, is gaining ground within companies and AI infrastructure.
A different way to read the AI market
The AI market is often analysed through a framework inherited from the software industry, in which the best product ultimately wins. This logic has long favoured American companies. OpenAI, Anthropic and Google enjoy privileged access to the world’s deepest pools of capital, talent and advanced computing infrastructure.
But AI is not merely an innovation industry. It is gradually becoming an infrastructure industry.
As companies multiply their use cases, the question is no longer simply which model performs best, but which one can produce a sufficiently good result at the lowest possible cost.
Most companies are not trying to solve International Mathematical Olympiad problems or outperform the market’s leading benchmarks. They need to process support tickets, generate summaries, analyse documents, automate workflows or assist developers.
DeepSeek is no longer an industrial anomaly
When DeepSeek released its R1 model, part of the industry treated the event as an unexpected anomaly. Yet the startup demonstrated that it was possible to achieve performance close to that of the leading American models with budgets far below those deployed by OpenAI or Anthropic.
The prevailing interpretation was that DeepSeek represented an isolated case. One year later, it is becoming clear that the tree was hiding the forest. DeepSeek is now part of a much broader group that includes Alibaba’s Qwen, Moonshot AI’s Kimi, Z.ai’s GLM, ByteDance’s Doubao and MiniMax.
These companies are not merely a collection of competitors. They are contributing to the emergence of a genuine industrial ecosystem. As happened with batteries, solar panels and electric vehicles, several players are advancing simultaneously across different technological layers while benefiting from considerable economies of scale. The industrial firepower is now in place, and this is only the beginning.
China’s real weapon: price-performance
The current competition is often presented as a contest of technological performance. Yet the competitive advantage of Chinese models lies in their ability to deliver comparable results at a much lower cost.
As companies seek to reduce their spending on tokens and inference, the economic trade-off becomes more important than marginal performance differences. This helps explain why companies such as Baseten are now building infrastructure specifically designed to run and optimise open models.
The market is beginning to recognise that not every use case requires the most powerful model. Many simply need the least expensive model capable of doing the job.
The battle to control AI’s open layer
While American laboratories dominate closed models, Chinese companies are advancing across the open-model ecosystem.
The distinction is strategic. Closed models may capture more direct revenue, but open models can win over developers.
The history of technology shows that these two dynamics do not always produce the same winners. Microsoft dominated proprietary software while Linux became the reference operating system of the digital economy. Android became the dominant mobile operating system, while Apple retained the industry’s highest margins.
AI could follow a similar trajectory. American laboratories may retain the technological frontier, while Chinese models become the software building blocks used by millions of developers to create agents, copilots and specialised applications.
A new form of technological dependence
This dynamic raises a geopolitical question that is rarely discussed.
For the past decade, Western concerns about China have focused primarily on physical supply chains, including batteries, rare earths, solar technology and electronic components.
AI introduces a form of dependence that is more difficult to identify. Tomorrow, a European company could run its applications on American GPUs inside American data centres while relying on a foundation model developed in China. The dependency would no longer concern hardware, but the software layer structuring how those systems operate.
This development is particularly significant because open models can be downloaded, modified and deployed locally. They are therefore far more difficult to control than services available exclusively through an API.
The sovereignty paradox
Recent US restrictions targeting certain advanced models have reinforced concerns among many governments and large companies. Dependence on a proprietary model also means dependence on its provider.
A regulatory change, political decision or modification to access conditions can have immediate consequences.
Open models offer a partial answer to this problem. Once deployed locally, they are largely insulated from such risks. But the solution creates a new paradox: in seeking to reduce their dependence on American laboratories, some organisations could increase their reliance on technologies developed in China.
The sovereignty question is simply changing form.
