金融时报对话李开复：中国AI为何能够领跑全球C端市场？

近日，在接受《金融时报》（Financial Times）专访时，我与记者探讨了当前AI产业发展的两个核心命题：中美AI领域的路线之争，以及企业如何在 AI 时代真正实现“突围” 。

美国巨头们选择以闭源模式押注“赢者通吃”，相比之下，中国AI领域更像是一个极具韧性的“共学小组”：在资源有限的前提下，通过开源模式和极致的工程落地能力，走出了一条高效率、重应用的破局之路。

但技术上的赶超只是第一步，竞争的重心正悄然转向谁能率先让AI走入厂房车间、走进企业核心生产场景。我认为，企业AI落地本质上是“一把手工程”。企业要敢于重用重塑组织的CAIO（首席 AI 官），与CEO并肩，携手如零一万物般懂AI的企业，以AI重塑核心业务逻辑。

以下为专访文章正文：

在过去的四十年里，李开复博士亲眼见证了中国人工智能产业的飞速发展。在产业萌芽阶段，他将当时世界前沿的科研创新机制引入中国，培养出了大批高科技人才，如今，这些人才已成长为活跃在各大科技巨头的中坚力量，也为AI产业提供了起飞的土壤。

这位国际AI专家曾主导创建微软亚洲研究院，使其成为中国顶尖AI人才的“黄埔军校”；随后，他又筹组领导了谷歌中国。2009年，李开复博士创办了著名科技创投机构创新工场，带领团队投资孵化了10多家AI独角兽企业；2023年，他创办了零一万物。零一万物是一家总部位于北京的大模型独角兽企业，致力于打造性能领先的产业大模型和为全球企业打造智能体解决方案。

在与《金融时报》中国科技记者 Eleanor Olcott 的对话中，他深入剖析了中美AI领域之间的行业竞争，并阐述了为何企业必须以更积极的姿态去拥抱这场技术变革。

Eleanor Olcott：能够介绍一下您的初创公司零一万物吗？

李开复：零一万物致力于打造全球领先的AI 2.0大语言模型平台及行业应用，助力企业AI数智化转型，提供构建AI智能体的工具和平台。我们基于顶尖的开源模型，会根据企业的具体业务场景挑选最优模型，以“一把手工程”为核心进行定制化开发。在AI智能体落地企业的早期阶段，提供从战略到落地的一站式服务至关重要，所以零一万物会详细解释技术如何应用，并与客户协同共创。这种深度参与决不仅是为了帮企业降本，更是为了创造实实在在的商业产出。

EO：这些企业对AI的接受程度如何？

李开复：在与银行、保险、矿山和能源等传统行业的合作中我们发现，相比于科技公司，这些行业在数智化转型方面准备不足，有些企业甚至连基本的数字化转型都尚未完成。对于这类客户，我们会审慎评估投入产出比，建议客户先夯实基础，避免后续产生过高的改造成本和时间损耗。另一个问题是，很多企业在提需求时是在“看后视镜开车”。比如他们只想做一个客服机器人，但这早已不是技术的前沿，更不是智能体最具价值的应用领域。

缺乏AI专业能力的企业必须与AI公司合作，共同制定AI数智化转型战略。这种转型必须由CEO主导，而且执行难度极大。目前看来，可能只有百分之一的企业真正做好了这种准备。当这种合作意愿达成时，我们就会深度介入，要求他们设立首席AI官（CAIO），因为传统的首席信息官（CIO）往往由于职业惯性显得过于保守，不能胜任AI数智化转型过程。CAIO必须具备大局观和冒险精神，直接配合CEO重塑组织架构。如果客户配不齐这个岗位，零一万物可以直接派驻前沿部署工程师（FDE）。

我们的商业模式类似于Palantir，都是由顾问协助制定战略，然后由执行人员负责实施。项目收入与最终取得的业务成果挂钩：前期收取的战略咨询费仅用于覆盖成本，核心收益则取决于为企业客户所带来的核心业务增量情况。

EO：那么对于另外99家不想这样做的公司，原因是什么？

李开复：有时是因为人们仅仅将AI视为另一种软件，有时是CEO对AI的本质缺乏基本的了解。有时人们把AI误认为是另一种ERP（企业资源规划）系统。或者，他们将任务委派给了错误的人选。如果你想制定AI战略，CIO通常不是合适的人选，因为他们的职责是确保系统平稳运行，而不是思考如何进行转型。

EO：目前普遍认为中国模型落后美国顶尖水平6到12个月。您认为这种差距会持续吗？

李开复：目前，全球 AI 研究领域的突破性成果大多数源自美国。他们拥有世界上顶尖研究人员和庞大的算力资源，在模型研发上确实步步领先。但中国团队的优势在于极其出色的工程落地能力，基于这些突破性成果，中国团队能够迅速掌握同类技术，并且往往能实现更高的运行效率。这促成了我们自己的“觉醒时刻”（中国大模型公司DeepSeek2025年发布的推理模型DeepSeek-R1，以更低的训练成本达到了OpenAI突破性模型的性能水平）。

正如“登月”最难的是第一次。一旦有人证明了路可以走通，即便不知道具体细节，成功的难度也会大幅降低。中国拥有非常强大的工程和研究底蕴，因此即便美国企业不开源也不发表论文，但中国公司通过研究这些模型的运作逻辑，已经实现了多项自主创新。可能实验结果本身就是一种启发，可能是通过巧妙的逆向工程、模型蒸馏，又或许是从技术的第一性原理出发，甚至可能是探寻出另一条不同的底层路径……但最终殊途同归，都是得到了相同的结果。

因此，中国模型往往能迅速追赶。DeepSeek发布时，中美大模型之间相差的研发周期缩短到了三个月；现在看来谷歌的Gemini已经领先，差距可能拉大到了12个月。这种差距呈现动态起伏规律，均值在6个月左右。每个人都会从已发布的优秀理念和模型中学习，因为人工智能领域吸引了众多顶尖人才，无论在中国还是美国，情况都是如此，而且他们都渴望着彼此学习。这种学习不是单向的，比如在DeepSeek问世时，所有美国公司同样在对它进行研究。

EO：2025年初DeepSeek发布R1模型时，OpenAI曾指责其通过“模型蒸馏”走捷径，随后OpenAI就表示已采取措施防止这种情况发生。我们暂且忽略窃取技术的指控，因为这似乎无法证实。但中国公司是否因为更严格的商业机密保护措施，而更难从美国公司那里学习呢？

李开复：OpenAI 对闭源的坚持不难理解：在投入巨资实现技术突破后，一旦开源，他们的核心成果很容易就会被他人低成本地获取。更深层的逻辑在于，他们深信 AGI 将带来一种质的飞跃。在那样的未来，率先攻克 AGI 技术的公司将对全球竞争者形成降维打击，无论是美国公司还是中国公司。因此，如果你认定 AGI 的终局是“赢家通吃”，那么对实现路径绝对保密，就是一种必然的战略选择。

这些美国公司在持续筹集千亿美元的巨量资金。为了支撑这种规模的估值，他们必须向投资者描绘这样一个愿景：一旦率先建成 AGI，他们将引领世界，因此即便今日投资500亿美元也依然“物超所值”，因为公司的市值终有一天会迈入50万亿美元。正是这套逻辑，让OpenAI的故事在商业上自洽，听起来不仅合乎情理，甚至颇具可信度。

但我认为，这个故事目前来看还有另一个版本。这场竞赛并不是只有一两个“天才少年”。在美国，OpenAI、Anthropic、Google 和 xAI 都在同台竞技，每一家都自认为是那个能解开 AGI 终极命题的“天才少年”，渴望以此实现赢家通吃，“赢得诺贝尔奖”。

而中国路径更像是一个“共学小组”。一家公司发布模型，另一家公司就去研究和尝试；甚至可能会去请教对方是如何训练模型的。学习小组的所有成员都在构建开源模型并进行分享。

值得注意的是，尽管这些学习小组是由一群非常聪明的孩子组成的，但资助他们的公司却希望每个季度都能看到利润。这与美国的情况非常不同，因为美国公司并不在意回报，但是在中国，公司的支出是受到限制的。举个例子，阿里巴巴不能在下个季度亏损100亿美元，但OpenAI可以。因此，种种原因使得中国公司在资源有限的情况下，需要像学习小组一样协作的方式运作，这与美国“赢者通吃”的策略截然不同。

EO：目前有一种主流观点认为美国在AI上的领先源于地缘战略优势。但我认为未来也有一种可能，我们会将中国曾经的落后视为一种优势。因为存在时间差，中国可以观察西方如何演进，看到AI带来的经济和社会动荡，并根据所看到的错误和陷阱选择不同的路径。您对此怎么看？

李开复：几乎可以预见，未来AI产生的负面影响将率先出自美国公司。无论是被不法分子滥用，还是因程序错误失控，这种无意间留下的隐患其实根源是在于美国公司的运作模式。在“赢家通吃、快鱼吃慢鱼”的心态下，公司自然而然地会减少安全防范方面的意识。同时，由于他们的模型和技术更加先进，他们也掌握着杀伤力更强的武器。

在中国，人们普遍不认为AGI的走向会是一家公司对全行业的降维打击。行业更倾向于相信，这将是一个领跑者不断易主的线性发展过程。

EO：难道中国公司不想成为赢家吗？

李开复：当然想，但大部分的企业不愿付出那种倾家荡产的代价。中国公司更关注商业产出和盈利能力，以及构建能从模型中赚钱的产品。腾讯有微信、阿里巴巴有淘宝，字节跳动有抖音，这些巨头都希望构建一个与其产品相匹配、能盈利且有竞争力的模型。

EO：您认为今年中国AI行业会发生什么？

李开复：企业级应用（B端）方面，我认为中国会稍稍落后于美国，因为中国企业普遍还没有养成支付订阅制服务费用的消费习惯。但在消费级应用（C端）领域，中国将领先美国。两国都有大量创业公司在深耕 C 端应用，且目前时机已经成熟、模型能力也已足够，但由于中国科技巨头在这方面始终都有展现出坚韧的态度、也渴望追求市场支配地位，所以我认为，他们在打造爆款应用方面将远超美国大厂。对中国大厂而言，应用开发本身就是他们研发技术的初衷，因此他们也会更专注。无论是用 AI 赋能现有产品，还是开发原生的 AI 应用，这些工作都已经成果初显了。

在我看来，中国互联网公司将成为 AI 应用创新的主要源头，其动力远超美国同行。反观美国的标杆性应用，无论是Instagram、YouTube还是Snapchat，它们正变得非常乏味。我不认为美国的互联网公司具备中国 C 端厂商那种拼搏精神，以及那种自我革命的果决。相比之下，像字节跳动、腾讯、阿里、美团、拼多多、小红书这些公司，拥有极强的韧性和求胜欲。他们中的许多企业正在重金投入研发顶尖的 AI 技术、Agents（智能体）和模型，其投入力度远超传统的美国模式。

其次，2026 年将开启“AI 原生设备”的元年。我们将在今年首次看到、亲手触摸、并购买到以 AI 为核心设计的原生硬件。它未必是最终胜出的终极形态，但它可能是“诺基亚时刻”、“黑莓时刻”，或者是“iPhone 时刻”。虽然尚不确定AI原生设备处于哪个阶段，但这三个节点在移动通讯史上都至关重要。人类一直渴望通过语音和自然语言向设备“委派任务”，因此 AI 原生设备是大势所趋。这意味着你只需告诉设备你想要的结果，而非完成工作的步骤。剩下的，交给智能助手去办就好。

这一趋势在智能体技术上已初现端倪。但它需要一个由语音驱动的交互界面，而目前来看，这种界面绝不是智能手机。手机并不是理想的载体，因为它无法做到“始终在线”和“实时倾听”。因此，你需要一种能够全天候运行、实时收音并捕捉全天信息的设备。它会存储你所见、所闻的一切，并以此为基础进行逻辑推理。

这是一个很复杂的命题，但我认为核心在于这种“环境 AI”（Ambient AI）。它始终在线、实时倾听、拥有无限记忆，而且让你几乎感觉不到它的存在。

EO：回顾您在中国AI行业的职业生涯，与开始时相比，今天行业的哪些方面会让您感到惊讶，哪些方面又基本保持不变？

李开复：我一直乐观地相信，“AI将改变世界”。让我感到惊讶的是过去三年AI进化的速度。我原本以为这会是一个跨越十到二十年的漫长过程，但它来得太快了，成熟得也非常迅速。当然，前路依然漫长。

回想起 1980 年代我刚进入AI行业的时候，AI 就像一堆派不上用场的“破铜烂铁”。偶尔有成效的时候，它也会被立刻包装成某种产品，从此不再被称为 AI。那时候人们嘲笑我们，觉得我们这群人疯了，居然认为机器能像人一样思考。可现在，万物皆可 AI。每一家 IPO 的公司都标榜自己是 AI 企业。我们见证了 AI 从“空想家的美梦”，变成了如今每个人都想参与的舞台中心。

本文翻译自《金融时报》报道，原文如下：

Kai-Fu Lee has had a front-row seat to the rapid growth of China’s AI industry over the past four decades, playing a central role first in building institutions that have spawned much of the talent now powering the country’s leading companies.

The Taiwanese-American computer scientist helped establish Microsoft Research Asia, which became a vital training camp for China’s leading AI talent, before later heading up Google’s operations in the country. Today, Lee heads Sinovation Ventures, a venture capital firm that invests in AI start-ups and is the founder of 01.ai, a Beijing-based AI start-up building agentic tools for companies worldwide.

In conversation with the Financial Times’ China technology correspondent Eleanor Olcott, he talks about the competition between AI’s two superpowers — China and the US — and why companies need to be more proactive to adopt the changing technology.

Eleanor Olcott: Can you introduce your start-up 01.ai?

Kai-Fu Lee: 0.1.ai makes tools to develop AI agents for companies. We build on open-source models, picking the right model for the company’s application and customising it for each customer. We believe that at an early stage of AI agent adoption, it’s essential to provide a white-glove service where we explain how technology can be applied. Together with the company, we co-create the most valuable applications that generate not just cost savings, but also business outcomes.

EO: How prepared are these companies to adopt AI?

KFL: We work with companies in traditional industries, including banking, insurance, mining and energy, which, compared to technology companies, are unprepared to adopt AI. Some of them haven’t done the digital transformation necessary for AI. In these cases, we won’t work with them because it will take too long and cost too much. The other problem is that some companies are looking in the rear-view mirror in terms of what they want. They might request to build a customer service agent, but that really isn’t where the technology or the best application areas are.

Companies that lack AI expertise must partner with an AI company to co-create their AI strategy. This kind of transformation is CEO-led, and it’s very difficult. Maybe one out of a hundred companies is prepared to do this.

When we partner with a company, we go in deep. They commit. We want them to hire a chief AI officer [CAIO], because the CIO [chief information officer] won’t do. CIOs tend to be very conservative. The CAIOs need to be bold and think big about strategy and the company organisation. They work directly with the CEO to reshape that. When our customers can’t provide a chief AI officer, we provide one for them.

Our business model is Palantir-like in the sense that we have consultants who help shape the strategy and then implementers who build it. We’re paid in accordance with the business outcome we create. We charge a set amount for the strategy development to recover our costs, but if there’s no business outcome, then we don’t get paid any more.

EO: And for the other 99 companies that don’t want to do this, why is that? KFL: Sometimes it’s because people think of AI as just another piece of software. Sometimes CEOs don’t have a natural understanding of what AI is. Sometimes people think of AI as just another kind of ERP [enterprise resource planning] software. Sometimes they delegate it to the wrong person. And the CIO is often the wrong person if you want to delegate AI strategy because . . . their job is to keep the company’s computers and software running smoothly, not to think about its transformation.

EO: The consensus today is that the Chinese models lag the leading American models by six to 12 months. Why? And do you think this will persist?

KFL: Currently, the US accounts for the great majority of breakthroughs in AI research. The US has most of the world’s top researchers and vast quantities of computing power to come up with advances in large language models. But based on these breakthroughs, talented and engineering-focused Chinese teams will quickly figure out how to build similar technologies, and often make them much faster which led to their own ‘aha moment’ [referring to China’s DeepSeek’s reasoning model released last year, which matched OpenAI’s breakthrough model at a much lower training cost].

It’s difficult for the first country to put a man on the moon. But once that has been done, and even though you don’t know the secret of how it was done, the fact that it was will make it so much easier for the second company or country to do it.

China has very strong engineering and research skills. So Chinese companies have made some inventions themselves, but they’ve also been able to figure out how these American models work, even though the American companies don’t do open source or publish papers. Perhaps the empirical result itself is enough of an inspiration. Perhaps it’s through clever reverse engineering. Perhaps it is through distilling the model. Or perhaps it is figuring out the first principles. Or perhaps, it figured out . . . different first principles, but it got to the result anyway.

So the Chinese models tend to catch up. When DeepSeek came out, it was shortened to like three months, and now it looks like [Google’s] Gemini has taken the lead, and lengthened the gap to perhaps 12 months.

That gap will shorten and lengthen, perhaps with six months as a midpoint. Everyone else will learn from every smart idea and model that’s published because the AI field has attracted many of the smartest people. That is true both in China and the US, and they’re all eager to learn. It’s not all one way. When DeepSeek came out, all the American companies studied it as well.

EO: When DeepSeek released its R1 model in January 2025, there were lots of accusations, including from OpenAI, that it cut corners by distilling its reasoning model. OpenAI then said it took steps to stop that happening. Let’s ignore the accusation of tech theft, which seems unprovable. Is it getting harder for the Chinese companies to learn from the US companies because they are taking more proactive measures to protect their secrets?

KFL: OpenAI feels that they have to keep the models closed, because after all this expensive work training breakthrough models, if they open source it, everyone will learn it much more easily. They feel so much money was put into inventing this IP; they don’t want to share it. This is understandable.

Also, they feel that the future of AGI [artificial general intelligence, a term that refers to a hypothetical future when AI has human-level cognitive abilities] will arrive as a giant step function when one company cracks it, and it will squash every other company in the world, be it American or Chinese. So in that sense, if you believe that the future of AGI is one where the winner takes all, then you have to keep how you arrived at that point secret.

The American companies keep raising hundreds of billions of dollars. They have to tell investors that by building AGI, they will dominate the world and that investing $50bn today is cheap because one day the company will be worth $50tn. So that makes the whole story work for OpenAI, and it’s an understandable, somewhat credible story.

But I think the alternate story is instead of having one genius kid or even four genius kids. In America you have OpenAI, Anthropic, Google, and xAI, each of which believes they’re the genius that will beat everyone else and win the Nobel Prize by solving the ultimate problem of AGI.

But the Chinese approach is different. The approach is more like a study group, where one company publishes a model, and the other looks at and plays with it. Maybe even talks to the company about how they trained it. All the members of the study group are building open source and then sharing it. So the study groups are formed of very smart kids who are all funded by companies that still want to show profit every quarter.

This is very different from the situation in the US, where companies do not care about returns. In China, companies are constrained in how much they can spend. Alibaba isn’t going to lose $10bn the next quarter. But OpenAI can. So, all these reasons cause the Chinese companies to behave the way they do with modest resources, learning and improving, working as a study group, as opposed to the American winner-take-all strategy.

EO: There is a dominant narrative that America’s lead in AI is a strategic geopolitical advantage. I think there’s a world in which, in the future, we see the fact that China’s been behind as an advantage. Because there is a time lag, Beijing can watch how this is evolving in the west. They can see the economic and societal disruption brought by AI and choose to take a different approach depending on the mistakes and pitfalls that they see. What do you make of this? KFL: It is almost certain that a future bad outcome from AI will come from an American company; if it’s being abused by some bad actors or by some error, a door has been left open that was unintentional. It’s just the way that they operate in this winner-take-all, run fast and break things mentality. It will cause companies to be naturally less conscious. And also they’re playing with more dangerous weapons, because their models and technologies are more advanced. In China, people in general do not believe that AGI will be one company squashing everyone else. I think people believe it’s going to be a linear trajectory where the winner will change.

EO: Surely the Chinese companies want to be the winners?

KFL: They want to, but they don’t want to pay the price; they don’t want to raise $500bn and have the company go bankrupt if they fail. The Chinese companies are more focused on the business results and on building products that make money from the models. Whether it’s Tencent’s WeChat, Alibaba’s Taobao or ByteDance’s Douyin, these companies want to build a competitive model that aligns with their products and can make money.

EO: What do you see happening this year in China’s AI industry?

KFL: I think China will lag the US in terms of enterprise adoption because of the unwillingness of Chinese companies to pay the kind of subscription fees. By contrast, China will lead the US in consumer applications. Both countries have plenty of start-ups working on consumer apps. The time is ready, and the models are good enough. But I think the Chinese giants will, by far, outrun the American giants in building great applications because the Chinese giants have always been tenacious, hungry, and monopolistic. And they see applications as the reason they’re building technology. So they’re going to be more focused. They’re going extend their existing apps with AI. They’re going to build new apps with the AI. It’s already coming out.

I think Chinese internet companies are going to be a source of this app innovation more than their American peers. If you look at the standard American app, whether it’s Instagram, YouTube, or Snapchat, they’re getting very boring. I don’t think the American internet companies have the same kind of approach to hard work, a willingness to reinvent themselves in the same way that the Chinese consumer app companies do. By contrast, the Chinese consumer app companies like ByteDance, Tencent, Alibaba, Meituan, PDD Group, Xiaohongshu, have tenacity and a desire to win and build new and innovative products. Many of them are building great AI technologies, agents, and models. They are already investing heavily in it, more so than the typical American way.

Secondly, 2026 will be the beginning of AI-first devices. This will be the year that we first see, touch, and buy an AI-first device. It may not be the ultimate thing, the format that ends up winning. It will either be the Nokia moment, the BlackBerry moment, or the iPhone moment. We don’t know which one it is, but all three moments were important in the history of mobile development. An AI-first device is needed because humans have always wanted to have a way to delegate to the device using speech and language. So this means telling the device the desired result rather than the steps to get the job done. The smart agent then . . . gets it done.

This is already happening with agent technology. But it needs to have a speech-driven interface, which isn’t a smartphone. The phone is the wrong device because it’s not always on, and it’s not always listening. So you need a device that’s always on, always listening and capturing information throughout your day. It will store everything you have seen and heard, and reason against this. So, it’s a long answer, but I think the key is this ambient AI that’s always on, always listening, infinitely remembering, and invisible.

EO: Reflecting on your career in the Chinese AI industry, if you look back at the beginning, what would you be surprised at about the industry today and what has remained largely the same?

KFL: I think an optimistic belief that AI would change the world has always remained the same. What I was surprised by is the speed at which it grew in the past three years. I thought it would be slower growth over 10 or 20 years, but it came much more quickly and matured very rapidly. We still have a long way to go.

When I started out working in the industry in the 80s, AI was always a bag of things that didn’t work. Whenever it did work, which was infrequent, it got turned into a product and was no longer called AI. People made fun of us or thought we were just a bunch of crazy people who think AI can think like humans. And nowadays, you know, everything calls itself AI. Every IPO calls itself an AI company. So we’ve gone from only the dreamers and wishful thinkers do AI, to now to everybody wants to be a part of it.

*试用企业级多智能体请访问：

www.lingyiwanwu.com/businesspartnership