How China is bursting the AI bubble
Moonshot and Alibaba claim that their AI models have virtually caught up with US suppliers. With this assertion, these Chinese companies are making both policymakers and the stock market nervous.
Two Chinese companies have unveiled AI models that are putting pressure on Silicon Valley. Kimi K3 from Moonshot and Qwen 3.8 from Alibaba are reportedly as good as the best models from Anthropic and OpenAI – and, unlike those, are publicly available. Kimi and Qwen are also open weight, meaning they can be downloaded in their entirety, run locally and are virtually impossible to monitor.
This moment reminds me of the release of DeepSeek in February 2025. Even back then, a Chinese AI model sent shockwaves through the industry and caused the stock prices of US tech companies to plummet. A few weeks ago, Chinese startup Ziphu AI also released a cybersecurity model (GLM-5.2) designed to rival Anthropic’s Mythos 5. DeepSeek and GLM are also open weight.
Copy, paste, flood
Not only are the models from Moonshot, Alibaba, DeepSeek and others freely available, they’re also a fraction of the cost of the most advanced chatbots from the USA. In short, China is engaging in AI dumping. It works the same way as with many other products, such as solar panels or cars. Instead of developing everything from scratch on their own, Chinese companies copy existing technologies, make them cheaper (and in some cases even better) and flood the market with them.
In the past, when it came to physical products such as cars or smartphones, this usually happened because China encouraged Western companies to set up production there by offering attractive terms. This is how Tesla and Apple came to operate huge factories in China. This saves them money, but at the same time, they’ve trained their direct competitors, because the expertise gained from production ended up at car companies like BYD and smartphone manufacturers like Huawei. And that’s not all – the country now trains millions of its own highly specialised professionals each year.
AI models are even faster and easier to copy – using distillation. Simply put, a large, expensive model acts as the teacher, and a new one is the student. Using millions of queries, the new model systematically extracts training data, becoming nearly as powerful as the original.
This is a well-established technique in the industry, and isn’t exclusive to China either. For example, within the same AI company, smaller models are trained for smartphones. And Elon Musk admitted that xAI distilled data from OpenAI for its chatbot, Grok.
From a moral standpoint, this constitutes theft of intellectual property. On the legal side of things, distillation falls into a grey area. There’s no specific law against it because, technically speaking, it’s not a copyright infringement. That being said, major AI suppliers’ terms of service prohibit the systematic queries required for distillation. In the event of violations, they can suspend accounts and take civil action against the account holders.
But this is a weak instrument that’s virtually useless on an international scale. That’s why AI suppliers try to build technical barriers into their models. But the most these measures can do is slow down the process somewhat, as long as the chatbots are to remain both publicly accessible and functional.
Source: Shutterstock
If US-based frontier labs continue to improve their AI systems, they could retain their lead in the future. But firstly, the gap between the Chinese models and the rest of the field appears to be smaller than previously assumed. And, secondly, from an economic standpoint, it plays only a minor role. A model that offers 90 per cent of the performance at 20 per cent of the cost compared to US models is more attractive for many applications.
AI as a political strategy
Beijing designated AI as a «strategic technology» as early as 2017, and incorporated it into its five-year plans for the first time in 2021. The political leadership there is pursuing several goals this way – productivity gains, control of information flows, military applications and geopolitical influence.
Chinese chatbots are gaining ground, particularly in emerging markets (linked page in German). And whoever provides the most widely used models influences what content is visible to millions of people, and in what form.
To slow China’s advance, the USA is restricting access to high-performance GPUs from Nvidia. At the same time, the government intends to keep its own models under stricter control. For instance, Anthropic was ordered by the Department of Commerce to temporarily take its cybersecurity models, Mythos 5 and Fable 5, offline. They are now available again, although the more powerful Mythos program is reserved for selected companies.
China is responding with technological decoupling (linked page in German). Beijing is pushing ahead with the construction of data centres in sparsely populated regions in the west of the country. Huge AI server farms powered by wind and solar energy are being built there for suppliers such as Huawei, Tencent, Alibaba and DeepSeek. The government reportedly covers up to 50 per cent of electricity costs, provided that domestically produced chips are used. The goal is to break away from Nvidia and other US suppliers and create an independent AI industry.
The political dimension doesn’t stop at industrial policy. AI is also viewed as a military factor in China. The People’s Liberation Army relies on «military-civil fusion», meaning the use of civilian technologies for military purposes. China’s powerful AI models therefore not only pose an economic risk to the USA, but are also a matter of strategic vulnerability. From cyberattacks and disinformation campaigns, to autonomous weapons systems.
Meanwhile, Europe’s watching from the sidelines. The continent is lacking its own high-performance AI models and the infrastructure needed for them. This makes it dependent on either the USA or China. And both options come with risks. If it came down to it, neither of them can be relied on to consider European interests. In the USA, the Mythos and Fable case was proof of this. All models also raise privacy concerns. But running open-weight models locally can mitigate these.
China is shaking up the house of cards
From an economic standpoint, the rise of Chinese AI models threatens the business models of Anthropic and OpenAI – and with it, a massive house of cards. Cloud providers such as Microsoft, Meta, Alphabet, Amazon and Oracle are building massive data centres to meet future demand from the two major AI suppliers. They are pouring vast sums of money into hardware manufacturers such as Nvidia, Samsung, SK Hynix and Micron.
The bottom line is that the majority of current US economic growth is tied to AI. Circular deals between big tech companies and neocloud suppliers such as CoreWeave and Nebius are also inflating balance sheets and stock prices.
The narrative of long-term profitability is becoming increasingly dubious. It’s questionable whether expensive frontier models actually increase productivity so much that it’s financially worthwhile for customers. At least to the extent promised. So far, suppliers have been skewing the bottom line with heavily subsidised plans and burning through vast amounts of venture capital in the process. This is evident, for example, from the leaked figures from OpenAI. But how many subscribers will ever pay the actual cost if Chinese suppliers flood the market with comparable models at rock-bottom prices?
China, too, is taking a big risk with its strategy. But there, it’s society as a whole that bears the burden. Beijing is investing billions (linked page in German) in public funds to expand AI. At the same time, a large part of the Chinese economy remains mired in a deep structural crisis, for instance in the real estate sector. If the global craze for AI were to fade, China, too, would be left with an expensive mess of redundant infrastructure to clean up.
My fingerprint often changes so drastically that my MacBook doesn't recognise it anymore. The reason? If I'm not clinging to a monitor or camera, I'm probably clinging to a rockface by the tips of my fingers.
Interesting facts about products, behind-the-scenes looks at manufacturers and deep-dives on interesting people.
Show all
