While media attention is focused on OpenAI, Anthropic or Google DeepMind, another battle is being played out in the artificial intelligence industry. It does not concern the biggest fundraising rounds or record valuations, but takes place in the deepest layers of the ecosystem: those which determine which models will actually be used by developers, companies and future generations of AI agents.
For two years, the dominant narrative has pitted American laboratories against each other in a race for power. OpenAI, Anthropic, Google DeepMind or xAI seek to build the most efficient models on the market, capable of pushing the boundaries of reasoning, programming or scientific research.
Yet as these models become more expensive to train and operate, a parallel dynamic is accelerating, and a new generation of open models, largely developed in China, are gaining traction in AI businesses and infrastructure.
Another read on the AI market
The AI market is often analyzed through a reading framework inherited from the software industry where the best product ends up winning. This logic has long favored American players, OpenAI, Anthropic or Google have privileged access to capital, talents and the most advanced computing infrastructures in the world, but AI is not only an innovation industry and is gradually becoming an infrastructure industry.
As companies expand usage, it’s no longer just a question of which model performs best, but it’s a question of which model produces a good enough result at the lowest cost.
In many companies, the goal is not to solve the international mathematics olympiads or beat the best benchmarks on the market, but to handle support tickets, generate summaries, analyze documents, automate workflows or assist developers.
DeepSeek is no longer an industrial accident
When DeepSeek published its R1 model, part of the industry considered the event as an incongruous anomaly, but the startup then demonstrated that it was possible to obtain performances close to the American leaders with budgets much lower than those mobilized by OpenAI or Anthropic.
The dominant interpretation was to see DeepSeek as an isolated case, but a year later, behind the tree is the forest. DeepSeek is now part of a much larger group including Qwen from Alibaba, Kimi from Moonshot AI, GLM from Z.ai, Doubao from ByteDance and MiniMax.
These companies do not constitute a simple group of competitors, but participate in the emergence of a true industrial ecosystem. Like what happened in batteries, solar panels or electric vehicles, several players are advancing simultaneously on different technological layers while benefiting from considerable volume effects. The striking power is there today and this is only the beginning.
The real Chinese weapon: the performance-price ratio
Current competition is often presented as a confrontation between technological performances. Apart from this, the competitive advantage of Chinese models lies in their ability to offer similar performances at a much lower cost.
As companies seek to reduce their token and inference spending, the economic trade-off becomes more important than the marginal performance gap. Which explains why companies like Baseten are now building infrastructure specifically designed to run and optimize open models.
The market is starting to consider that all uses do not need the most powerful model but the least expensive model.
A battle to control the open layer of AI
While American laboratories dominate closed models, Chinese players are making progress on open models.
A strategic distinction, if closed models make it possible to capture more direct revenue, open models make it possible to win over developers.
However, the history of tech shows that these two dynamics do not always produce the same winners. Microsoft dominated proprietary software when Linux established itself as the benchmark operating system for the digital economy. Android has become the dominant mobile operating system while Apple maintains the highest margins in the industry.
AI could follow a comparable trajectory; if American laboratories could maintain the technological frontier, Chinese models could become the software building blocks used by millions of developers to build agents, co-pilots and specialized applications.
A new kind of technological dependence
This dynamic raises a rarely addressed geopolitical question.
For a decade, Western concerns regarding China have mainly concerned physical supply chains (batteries, rare earths, photovoltaics, electronic components).
AI introduces a dependency that is more difficult to identify; tomorrow, a European company could run its applications on American GPUs, in American data centers, while relying on a fundamental model developed in China. The dependence would no longer relate to the hardware, but would concern the software layer which structures uses.
This development is all the more significant as open models can be downloaded, modified and deployed locally, and thus become much more difficult to control than services accessible exclusively via API.
The paradox of sovereignty
The recent affair surrounding American restrictions targeting certain advanced models has reinforced a concern among many governments and large companies. A dependence on a proprietary model implies a dependence on its supplier.
The slightest regulatory change, political decision or modification of access conditions can have immediate consequences.
Open models offer a partial response to this problem; once deployed locally, they largely avoid this risk. But this solution creates a new paradox, because by seeking to reduce their dependence on American laboratories, certain organizations could increase their dependence on technologies developed in China.
The question of sovereignty simply changes form.
