数字主权：新科技秩序

The New Tech Order Executive Conversations KAI-FU LEEFounder & CEO 01.AI OPEN VS. CLOSEDAI MODELS author of the New York Times and Wall StreetJournal bestseller AI Superpowers, which examinesthe role of the US and China in the future of artificialintelligence, as well as its impact on society. Dr.Lee trained as a computer scientist in the UnitedStates and has spent his career at the intersection ofAmerican and Chinese technology development. Dr. Kai-Fu Lee is the founder and CEO of 01.AI,a Beijing-based artificial intelligence companyfocused on large language model developmentand enterprise AI transformation, venture-built byhis tech venture capital firm Sinovation Ventures.He previously led Google's operations in Chinaand held senior roles at Microsoft. Dr. Lee is the UNDERSTANDING AI SOVEREIGNTY You have navigated the intersectionof technology, culture, and geopoliticsacross the US and China for decades. Howdo you define AI sovereignty today, andwhy does it matter? Kai-Fu Lee:Sovereignty means a numberof things. First, it means having controlover the technologies themselves. Second,it means ensuring that data, which is acritical asset to every company, does notleak. That concern is widespread acrossmost countries, but perhaps less so in theUS, given its more advanced acceptanceof private cloud. But throughout theworld, companies are deeply concernedabout confidential data becoming an inputthat enriches a general model owned byanother country. Kai-Fu LeeFounder & CEO, 01.AIChairman,Sinovation Ventures There is also the question of suitability.A model trained in one country tendsto embed the values of that country,and those values may not be applicableelsewhere. Having control over the modelallows a company to fine-tune it, to makeit fit the organization. Hosting the modelin someone else's cloud limits that abilitysignificantly. Throughout theworld, companies aredeeply concernedabout confidentialdata becoming aninput that enriches ageneral model ownedby another country." When Beijing, Brussels, and Washington each talk about AI sovereignty,are they talking about the same thing? Kai-Fu Lee:Not entirely. There are layers. The first is suitability: every countrywants a model that reflects its language, culture, and legal norms. In someIslamic countries, questions around same-sex marriage, alcohol, or religiousguidance are handled very differently than they would be in the US. Ensuringcompliance with local law, religion, and culture is a legitimate and universalconcern. The second layer is competitive ambition. Some countries feel they need tobuild their own model in order to have a chance at a meaningful position inthis technology race. China clearly has that ambition. Japan, Singapore, SaudiArabia, and India have all declared theirs. That is admirable. But it is sometimesoverly ambitious. Training a frontier general-purpose model from scratch isnot feasible for every country. The US has enormous resources. China hassubstantial resources. Most others do not. The practical alternative, and I think the right one for almost all countriesoutside the US and China, is to take a leading open-source model and continuetraining it for the country's specific language, values, and regulations.Continuing training is different from fine-tuning. It treats an open-sourcemodel as a half-baked product. Think of it as buying a frozen pizza and thenadding your own ingredients and baking it further per your preference. Theresult is something adapted to your context, whether an Indian pizza or aJapanese pizza. And, critically, the cost is only a few percent of training fromscratch: a few million dollars rather than a few hundred million. "The practical alternative, and I thinkthe right one for almost all countriesoutside the US and China, is to take aleading open-source model and continuetraining it for the country's specificlanguage, values, and regulations." OPEN-SOURCE MODELS AND THE BUSINESS OF AI DOMINANCE What has been the actual effect of US export controls on China's AIdevelopment? Kai-Fu Lee:It is a double-edged sword, and the jury is still out. The GPUexport restrictions are only one side of the equation. The other side is thatChinese companies are extremely practical. If you survey them, the greatmajority are not trying to beat Anthropic or OpenAI. The Chinese companiesknow they cannot spend fifty billion dollars on supersized model training.So, with the GPUs available to them, some domestically produced and not asscalable, they ask: what can we do excellently within these constraints? The answer has been rigorous engineeringefficiency. DeepSeek is the most well-knownexample. When people are constrainedbut motivated, they innovate around theconstraint. They optimize architectures, moveaway from CUDA dependencies, and targetperformance that is very good rather thandefinitively the best. Chinese companies arespending less than ten percent of what thetop American companies spend on training,yet are prod