China AI, Honestly.
No fear-mongering. No dismissal. What is actually happening and what it means for your business.´
Western coverage of China’s AI development tends toward two extremes: ‘China is about to surpass us on everything’ and ‘Chinese AI is derivative and overhyped.’ Both are wrong. Here is what is actually happening, and why it matters for decisions you will make this year.
What China Is Actually Building
The framing of Chinese AI as ‘trying to catch up to OpenAI’ has been obsolete for at least 18 months. Chinese labs are not primarily trying to win Western benchmark leaderboards. They are pursuing a different strategy with a different definition of success.
The strategy: build deployable, scalable AI that runs on infrastructure China controls, serves Chinese market needs at scale, and can be exported to markets where US AI infrastructure is not the default.
DeepSeek is the clearest example. DeepSeek’s models have repeatedly demonstrated competitive performance on reasoning tasks at a fraction of the training cost of Western frontier models. More importantly, DeepSeek V4 is being built to run on Huawei chips — not because Huawei chips are better than Nvidia, but because Huawei chips are not subject to US export controls.
This is not a technical story. It is a geopolitical one.
The Domestic Market Nobody Talks About
Alibaba, Tencent, Baidu, ByteDance, and a dozen smaller players are engaged in an AI deployment race in the Chinese domestic market that dwarfs anything happening in the West in terms of user scale.
Baidu’s Ernie Bot has more active users than any Western AI assistant. ByteDance’s AI features are embedded in products used by hundreds of millions of people daily. This is not frontier model research. It is mass deployment at a scale Western companies have not achieved.
The strategic implication: China’s AI advantage is not necessarily at the frontier of capability. It is in deployment experience at scale, in AI user behavior data that nobody else has, and in the practical infrastructure of making AI work in real products for real people at hundreds of millions of simultaneous interactions.
What This Means for the Infrastructure Question
The global AI infrastructure is fracturing. This is not a prediction — it is a description of what is currently happening.
There are now meaningfully different AI ecosystems: the US ecosystem (OpenAI, Anthropic, Google), the Chinese ecosystem (Baidu, Alibaba, DeepSeek), and an emerging European ecosystem trying to establish AI sovereignty.
These ecosystems are not just different products. They are increasingly subject to different regulatory frameworks, different data governance rules, and different geopolitical pressures.
For businesses operating across these regions, the practical implication is that the AI tools available to your US operations may not be available to your operations in other markets, and vice versa. This is not a hypothetical. European companies are already dealing with GDPR complications from running on US AI infrastructure. Companies operating in China are already navigating regulations about which AI models can be used for which purposes.
The Benchmark Question
When DeepSeek releases a model that claims competitive performance with Western frontier models, Western media tends to either uncritically accept the claim or dismiss it as propaganda. The correct response is neither.
Benchmarks measure specific capabilities on specific tasks. A model that scores well on reasoning benchmarks may perform differently on tasks relevant to your specific use case. The right question when any new model is released — Western or Chinese — is not ‘what are the benchmark scores?’ but ‘how does it perform on my specific tasks with my specific context?’
Test it yourself. Form your own view. Do not rely on either the press releases or the dismissals.
What Businesses Should Actually Do
Three practical implications for business decisions:
Map your AI dependency geography. Know which AI providers your critical workflows depend on, and understand what regulatory or geopolitical pressures those providers are subject to. If you operate in multiple markets, understand what is available in each. Do this now, before a supply shock forces the analysis.
Do not build single-provider dependency without a fallback. Regardless of which ecosystem you primarily operate in, design your AI architecture so that you can switch models if necessary. This means abstraction layers between your workflows and specific model APIs. The cost of building this abstraction is low. The cost of not having it when you need it is high.
Watch the deployment layer, not just the frontier. The most important AI developments from China in the next 24 months will not be new frontier models. They will be advances in how AI is deployed in specific industrial and consumer applications at scale. These applications will reach non-Chinese markets faster than most Western observers expect.
The AI race is not one race. It is several races happening simultaneously with different metrics and different winners.
The honest assessment: neither panic nor dismissal is warranted.
China’s AI development is real, significant, and pursuing a different strategy than Western development. The geopolitical dimension is real and is affecting business decisions now, not in the future. The specific capabilities of Chinese models on specific tasks are worth evaluating directly rather than through the filter of either boosterism or skepticism.
The most useful thing you can do: stay specific. Which Chinese AI capabilities are relevant to your business? Which markets do you operate in where the geopolitical dimension creates real constraints? What would you do if your primary AI provider became unavailable or significantly more expensive? These are answerable questions. Answer them.
— Malte | Signal & Noise | thoughtbymalte.substack.com