In The Innovator’s Dilemma, Clayton Christensen tells the story of the American steel industry in the 1970s and 1980s1.. The large integrated mills, US Steel and Bethlehem and the others, made every grade of steel from iron ore in enormous plants, while a new kind of competitor, the mini-mill, melted scrap in an electric arc furnace at a substantially lower cost and could only produce rebar, the lowest-margin, least demanding product in the market. The integrated mills were happy to give rebar away, because losing their worst business improved their average margin, and they moved upmarket.
However, the mini-mills soon learned to produce angle iron, structural beams, and even sheet steel, the only commodity remaining for the integrated mills. With each advance, the incumbents retreated to products in which they were dominant, but the market for these products shrank until there was nowhere left to retreat.
I have lived and worked in Shanghai for seven years, helping large companies figure out how to use AI in China and Europe. I keep coming back to this story because it describes China’s approach to AI better than any geopolitical commentary. When I present the Chinese ecosystem to a European audience, the discussion usually jumps straight to chips, sanctions, and politics. It never gets to what matters for businesses: how AI is being adopted here, by whom, and why it’s happening faster than elsewhere in the world.
The commodity bet
China has made a late but strategic choice to treat AI as a commodity rather than as a product, with strong incentives to release their models as open weights. Xi Jinping made this explicit in his keynote at the World AI Conference in Shanghai this July, the first time he attended in person, when he called on countries to “encourage open source, openness, collaboration and sharing“2..
When the model itself is free, the value moves to everything that sits on top of it: deployment, integration into existing systems, caching of tokens to make models more efficient, expertise to make it work in a real factory or bank, and armies of engineers who provide that expertise at scale. This is essentially theweb2 playbook applied to AI, where the technology itself isn’t sold, but everything around it is. It’s not so different from what Mistral is doing in Europe. However, Chinese labs have a much deeper bench of AI scientists to sell with it. A Georgetown study projected that by 2025, Chinese universities would graduate close to twice as many STEM PhDs per year as American ones. Also, close to half of the researchers publishing at the top AI conference did their undergraduate studies in China3. .
China can afford this bet because it has a structural advantage on energy and talents. Inference is a combination of an electricity and hardware business, and China continues to build generation capacity at a pace that Europe has more or less stopped contemplating. This means the marginal cost of a token can continue to fall in a way that it cannot elsewhere and be subsidized for as long as necessary to secure technological autonomy.
Chip restrictions paradoxically strengthened the mini-mill dynamic. The stated intention was to slow China down. However, the practical effect has been to push Chinese labs into an obsession with efficiency. This means more intelligence per token, per watt, and per RMB because they could not simply throw more hardware at the problem. This is very much the Chinese industrial playbook: the cheaper, constrained producer relentlessly optimizes on cost, starts with work that the incumbent does not want, and works its way up. When you examine the rankings based on cost-efficiency and question why Chinese models are at the top, you’ll realize that sanctions and constant creativity play a significant role.
Adopt now, ask questions later
Chinese business is extremely pragmatic and driven by individual profit. I say this without any judgment as someone who experiences it every day. At the individual level, money is the score, and the government’s role is to preserve social peace and punish those who abuse the system. Within that framework, Chinese businesses ask fewer questions and take more action.
The consumer side tells the same story, which is striking for a country of 1.4 billion people where everyone has to make a living. In the 2026 Ipsos AI Monitor, a 32-country survey, 85% of Chinese respondents say that products and services using AI have profoundly changed their daily life in the past three to five years, against 36% in the United States, 33% in Germany and 32% in France4.. Ipsos also found that about half of Chinese users say they trust AI tools enough to not check their work, although they should definitely do so, as with any model. On some issues, such as avoiding bias, they trust AI more than people. While Europeans and Americans are generally nervous about AI, the Chinese are generally excited about it.
From an industrial perspective, the push towards automation and robotization is not economically rational in the short term. However, due to its aging population, China does not have the option of waiting. Between 2024 and 2050, the population aged 15 to 64 will fall by about a quarter, from 984 million to 745 million people6.. Therefore, China is investing now in what it will need over the next 20 or 30 years, including robots and agents.
Being the second mover has clearly helped China achieve outstanding performance on AI models. The “China for China” approach works because the domestic market is large enough to reach scale without ever needing to leave the country. These factors provide Chinese AI labs with the demand necessary to grow and compete globally.
Not frontier models: For most uses it doesn’t matter
Chinese models are not on par with the best American models. However, this is the mini-mill “rebar” stage. For most business applications in China, frontier performance isn’t necessary. What matters when running agents at scale is token efficiency: how much useful work you get out of every RMB of inference. On that metric, Chinese models have been winning for about a year: as of September 9, 2026, eight of the ten most-used models on OpenRouter are Chinese. DeepSeek alone handles more requests than OpenAI or Google combined. This ranking measures agentic work in terms of tokens consumed rather than dollars spent. OpenRouter is a marketplace where developers route traffic for their coding and autonomous agents. Its top applications by volume are agents, not chatbots7.. Stripe just agreed to pay over seven billion dollars for the marketplace, showing that investors see the value not in any particular model, but in the ability to switch between them for a 5% markup8..
OpenAI’s GPT-6, released on September 3, 2026, is the first significant breakthrough in performance versus tokens from a Western lab in a long time, at least in agentic coding benchmarks. This means that, for once, the incumbents are not retreating upmarket, but rather, are meaningfully competing on cost9.. An interesting question is how quickly Chinese labs will catch up. Another caveat is that the business model of Chinese frontier labs is still nascent. They are developing the expertise to sell to business users on a large scale, but that is an unproven strategy.
In the meantime, AI-native companies already exist in China. They are not operating at scale yet, but one-person companies that run on agents are a reality and are actively promoted by local governments 10. . This is changing the economic structure of a country that cares more about what works than what used to work.
The trust paradox
Many Western companies do not trust Chinese models, and this is a legitimate position. This has forced Western companies operating in China to build infrastructures that give them complete control over their AI capabilities. They implement LLMs on-premises with strong agent access monitoring, which leads to greater transparency and accountability than most Western AI infrastructures provide. Ultimately, this results in AI platforms that allow for flexibility in switching LLMs while keeping data and knowledge safe. The model you distrust is actually the one you can control. I believe that, as a result, these companies have built the most sustainable enterprise AI infrastructures in the world.
What this means for a Western business?
Christensen’s lesson was never that the mini-mills were better. Rather, it was that the integrated mills lost by making a series of reasonable decisions to cede the low-margin business and concentrate on the premium product. Eventually, the cheap competitor learned to produce the premium product as well. I would draw three conclusions for a Western company, and I think they hold regardless of what happens to the models next year.
- First, keep control of your AI choices. Things are always changing. The lab that is leading in September may not be leading in March. You should build your systems so that you can switch between models according to performance, cost, and jurisdiction without rebuilding the whole system.
- Second, keep control of the harness, meaning the scaffolding around the model, including the orchestration, tools, memory, and guardrails. In our experience, this has a greater impact on agent performance than the next generation of models. It is where your competitive advantage lies, so it’s the one thing you should never outsource.
- Third, keep control of your business knowledge, context, data, and processes, structured in a way that any model can consume. The model is becoming a commodity, but what your company knows is not.
For the rest, let’s watch the mini-mill story unfold.
