2025 年 3 月 20 日，零一万物 CEO、创新工场董事长李开复博士接受了彭博社的专访。
在访谈中他表示，DeepSeek 热潮之后，中国经历了自己的“DeepSeek”时刻。零一万物看到了DeepSeek在基座模型领域的技术突破，也洞察到在基座模型与市场需求之间存在着缺位——由于缺少连接企业知识库、搭建应用的中间件，DeepSeek 优质模型难以转变为企业生产力。
李开复以 Windows 内核类比：DeepSeek 是 Windows 内核，而零一万物提供的平台则是 Windows 操作系统、应用程序、UI 界面。没有 Windows 操作系统、应用程序、UI 界面，Windows 内核就无法发挥真正的价值。
李开复强调，AI 需要市场，市场也需要 AI。要真正发挥大模型的价值所在，AI 领域内仍然没有出现自己的“Windows 操作系统”，零一万物想要做的正是填补这一空缺。
因此，零一万物选择全面拥抱DeepSeek，推出万智企业大模型一站式平台，不仅提供安全模型部署方案、模型微调方案，还封装了联网搜索、知识库RAG、Agent搭建等工具，助力DeepSeek等优质基座模型赋能企业场景。
李开复博士（右一）在访谈现场
谈及大模型领域未来的竞争格局，李开复认为，在中美，超大模型的预训练正在逐渐寡头化，并且寡头化的程度在不断加大。他预测中国大模型领域未来将会收敛至 DeepSeek、阿里巴巴、字节跳动三家，其中以 DeepSeek 势头最猛；美国则以 xAI 的发展势头最为迅速，但OpenAI、Anthropic、Google也都会做出令人敬佩的研发工作。
至于中美之间的 AI 竞赛，李开复认为 GPU 资源稀缺的情况反而倒逼着 DeepSeek 做出了惊艳业界的技术创新，实现训推成本与模型性能的突破。在印证了“需求是创新之母”的同时，DeepSeek 的突破也使得开源模式呈现出压倒性的优势，坚持闭源策略的OpenAI、Anthropic等公司，都会面临更大的压力。
“OpenAI 在 2024 年的运营成本为 70 亿美元，而 DeepSeek 的运营成本可能只有OpenAI 的 2%。目前各家的模型都很优秀，问题不在于哪家模型性能高出 1%，而是 OpenAI 成本高昂的模型路线是否具有可持续性。”李开复强调，“OpenAI 每年花费 70 亿美元，面临着巨额亏损。现在出现了一个竞争对手，将成本低廉数倍的开源模型免费开放给市场，且这个竞争对手资源充沛，并已经有效地将计算成本降低了五到十倍。
“有了这样一个强大的竞争对手，我想 OpenAI 的萨姆·奥尔特曼可能夜里会辗转难眠。”李开复开玩笑说。
以下为对话全文：
彭博社：今年可谓是人工智能真正渗透各行业的元年，此次我们特邀零一万物CEO、创新工场董事长李开复博士来深入探讨这一话题。今天也是一个绝佳的时机来和您一起讨论，因为零一万物刚刚对外发布了“万智企业大模型一站式平台”，首发能帮助企业更广泛地部署、应用 DeepSeek 模型的解决方案。能不能请您详细介绍一下这项业务？
李开复：当 DeepSeek 问世的时候，中国迎来了自己的“ChatGPT 时刻”，我们可以称之为“DeepSeek 时刻”。在中国春节假期期间，几乎每个人都在谈论 DeepSeek，许多企业的 CEO 们节后返工马上要求在自己的公司里用上 DeepSeek。他们发现，DeepSeek 是一个非常出色的模型， AI 的表现令人惊叹，但它不能直接赋能公司里面的人力资源、财务和客服等场景，因为开源模型还缺少连接企业数据库、应用程序等必要的中间件。
零一万物认同 DeepSeek 所取得的巨大进步，我们决定全力支持 DeepSeek ，让它更好用。我们补充了那些缺失的中间件，开发了易用的使用界面，使 DeepSeek 在企业中变得可用。这就是零一万物于本周一发布的产品——万智企业大模型一站式平台，它在中国大陆和香港地区的初期反响相当不错。
彭博社：这款“万智”新平台背后的考量是怎样的？
李开复：许多人现在都下载了 DeepSeek，也很喜欢使用 DeepSeek。我有一个 CEO 朋友问他的员工：你们都用 DeepSeek 做什么，得到的答案是算命。这是很有意思的用法，你们或许都可以试一试，但目前的 DeepSeek 还无法直接深入到企业级应用里去。
每个企业都有 ERP 和 CRM 数据库，有员工数据等各种分门别类的企业内部信息，企业需要的是具备行业纵深的知识型模型。比如彭博社会希望有一个金融知识渊博的模型，平安保险集团会希望有一个保险专业知识的模型。零一万物发布的平台，就是为了满足这类各行各业的企业级需求。
这就像是如果我交给你 Windows 内核，你可能不知道如何把它用起来，你会需要 Windows 软件操作系统和应用程序界面，还有配套的 Windows 系列应用，Windows 内核才能变得有用。零一万物推出新的万智企业大模型一站式平台，就像是在为 Windows 内核构建可视化界面和操作中台，让 DeepSeek 的基座模型高效便捷地转化为有效的生产力工具。
彭博社：选择 DeepSeek 作为底层架构而非美国模型，是否折射出中美AI竞赛格局变化？
李开复：在中美，超大模型的预训练正在逐渐寡头化，并且寡头化的程度在不断加大。虽然未必所有人都认同，但我们的确同时看到开源圈展现出压倒性的优势，OpenAI 和 Anthropic 都相信自己还能训练出远超其他玩家的闭源模型。当他们看到一个性能相近的开源模型时，我想他们可能受到了沉重的打击。OpenAI 在 2024 年的运营成本为 70 亿美元，而 DeepSeek 的运营成本可能只有OpenAI 的 2%。
各家的模型都很优秀，问题不在于哪家模型性能高出 1%，而是 OpenAI 成本高昂的模型路线是否具有可持续性。OpenAI 每年花费 70 亿美元，面临着巨额亏损。但现在出现了一个竞争对手，将成本低廉数倍的开源模型免费开放给市场，且这个竞争对手资源充沛，目前看来 DeepSeek 有足够的资金储备持续投入模型研发，并已经有效地将计算成本降低了五到十倍。有了这样一个强大的竞争对手，我认为 OpenAI 的萨姆·奥尔特曼可能夜里辗转难眠。
彭博社：那么，你认为DeepSeek真的挑战了现有 AI 的商业逻辑吗？另外，无论是在美国还是在中国，AI 的商业落地已经相当规模化。在开源模式占据优势的当下来看，你认为中国市场未来会跑出多少个 AI 大模型？
李开复：我认为很大概率会收拢到三家，DeepSeek、阿里巴巴和字节跳动。他们的模型会随着时间不断迭代进化。
彭博社：这份名单上少了一些公司的名字。
李开复：当然，还会有许多其他公司希望构建他们自己的模型，我们也会拭目以待看是否会有第四、第五个玩家。但即使在这三个中，我看好 DeepSeek 目前最具势头，这也是我们基于它推出新产品的原因。
但我也想指出，关于你问到 AI 模型商品化的问题，我不认为今天的 AI 已经商品化了。我认为，吸纳大量资金训练的基座模型开始趋近同质化，它们成本高昂、越来越难和开源技术竞争，但人工智能技术并没有商品化。
就像 Windows 内核、Linux 内核这些已经比较商品化，但 Windows 软件系统并没有商品化，微软多年还靠此赚了很多钱。零一万物想要做的是填补 AI 圈还不存在的 Windows，并使用 DeepSeek 作为我们类 Windows 的内核之一。
彭博社：那么，你认为现在的 AI 增加了什么价值？对于投资人来说，在思考这个行业的价值增长点时，具体有哪些主要的经济因素？而为什么你认为只有三个玩家的空间？
李开复：的确可能还有第四、第五个，但在多数市场中，大部分的终局最后将以寡头局面收场。第一名赚了很多钱，第二名收支平衡，后头几名多半不能算是成功。在大模型的赛道中，由于底层模型很难直接转化为商业模式，所以情况更加复杂。但一旦你是最终赢家，就有许多方法可以变现。近期我们可以看到的变化是，中美两地的投资人几乎不再下注更多更贵的底层模型公司。他们更愿意投资 AI 应用、消费者应用、AI 基础设施类型的创新企业，还有企业应用的广大市场需求，未来每个应用、千行百业都会被 AI 重塑，这些应用能为企业提供真正的价值。所以，AI 相关的资本正转往这些方向，而不再投入基座模型。
彭博社：半导体技术封锁是否会成为制约中国 AI 发展的一个限制因素？
李开复：我们可以回顾过去来理解这个问题。过去，中国总是能够在底层技术方面迎头赶上，并构建出伟大的工程产品，而美国更擅长取得前沿突破。我曾经在《AI·未来》这本书中深入论述这点，这种情况目前来看仍在继续。
如今我们看到 DeepSeek 展现出了极强的适应性和竞争力，再次证明了“需求是创新之母”这一观点。因为他们拥有比美国大厂少很多的 GPU，所以不得不设法把已有 GPU 的效率提高 10 倍。虽然现在很难去预测未来的走势，但我会说，正因为 DeepSeek 拥有计算资源比美国大厂少了几个数量级的，反而被迫更加精益求精，倒逼出的技术创新使得他们的模型训练和推理成本比 OpenAI 和其他公司低五到十倍，让全球刮目相看。
彭博社：大约几个月前，在 2024 年的 10 月份我们曾经有过一次交流，当时零一万物也推出了与 DeepSeek 相似的模型，并且早于 DeepSeek 发布V3 之前。您如何看待模型迭代的频率？这种频率是被什么因素主导？DeepSeek 的成功是一种趋势使然，还是产品本身的魅力？今年，大模型行业又会给我们什么样的惊喜？
李开复：中国大模型行业的竞争仍会热火朝天好一段时间，最终可能有三个赢家，美国可能有四个赢家，但这并不意味着其他几十家公司会放弃竞争，所以我判断那些正在快速发展壮大的公司会继续高频发布新的模型和产品。在我看来，xAI、DeepSeek 是发展速度最快的公司。同时，OpenAI、阿里巴巴、谷歌和 Anthropic 都会做出令人敬佩的研究工作。
尽管字节跳动可能不在大多数人的视野里，但我会把它列在名单上。字节在大模型领域已经投入了海量资源。在前沿模型探索上敢于投入大量资金，其中一个最合理的逻辑是掌握了大量的用户，当所投入的巨额成本分摊到能获利的大体量用户平台时，商业模式就能够成立。基于字节跳动拥有最多的用户、最多元化的商业变现模式，我做了这个预测。
彭博社：百度是个让我印象深刻的公司，但不在你的名单上。此外，中国政府对AI的支持，有哪些变化？
李开复：我们都非常尊重并感谢百度在中国真正开启了Transformer革命。他们试图聘请杰夫·辛顿，后来延揽了吴恩达。我认为这证明了李彦宏先生具有远见。
另一个关于政府支持的问题。中国政府的一个重点工作方向被称为“新质生产力”。这意味着利用高科技创造生产力、降本提效，并通过高新技术的引入，将所有传统行业转变为更具盈利能力和竞争力的行业，而 AI 正是其中的核心引擎。
彭博社：几周前的民企座谈会上，中国科技巨头的 CEO 们与国家领导人同处一室，这种情形很少见。从这个角度来看，你认为政府会有哪些具体的措施来支持这一目标？
李开复：我认为目前中国地方政府都在非常积极地拥抱 AI，各地都在探索如何将 AI 应用到当地的传统优势产业中，用技术打造“新质生产力”，进而促进城市 GDP 增长，这是一个从上到下都在贯彻落实的方向。对于地方政府来说，“新质生产力”是令人振奋的、非常有效的新增长点，零一万物和其他友商公司都开放和地方政府及产业展开紧密的合作探索。
彭博社：开复博士，非常感谢您加入今天的对话，与你交谈真是令人愉快，也祝贺零一万物万智企业大模型一站式平台的发布。
本文摘编翻译自彭博社 Bloomberg China Show：
Q: This is the year where we start to see really AI being applied across a range of different industries. Let's get more on that now because we've got Kai-Fu Lee. He's the CEO and Founder at 01.AI and also Chairman of Sinovation Ventures. It's actually perfect timing because you've just launched an enterprise large language model that's also gonna be deploying DeepSeek more broadly. But talk us through exactly what it entails.
A: Yeah, well, I think China had it's ChatGPT moment when DeepSeek came out. We can call it "DeepSeek moment." Everyone iss aware of it over the Chinese holidays. Everyone is talking about it. And the CEOs came back to work saying "put DeepSeek to work at my company." And what they found out was DeepSeek is a fantastic model, amazing AI, but it doesn't have the middleware and the user interface that it takes to connect to corporate databases, to build applications, to make it useful for HR, Finance and Customer Service. So what 01.AI did was we saw DeepSeek has been making great momentum. And we decided to really bet on DeepSeek and build that missing middleware and UI so that DeepSeek can be made useful for corporations. That's the product we announced this Monday and we're getting fantastic reception in China and also in Hong Kong.
Q: Tell us more about that launch.
A: Yeah, so we launched on Monday. And what we talked about was many of you have DeepSeek now. You love to use it. In fact, the one CEO friend of mine asked his employees, what do you use it for? And the answer was fortune telling. By the way, that's a great thing to try for you. But it's not very deep into the industry for the companies. Every company has ERP and the CRM databases. They have employee records, they have their internal information. And they want the model to be more a generalist. They want it to be knowledgeable. Bloomberg would want a finance knowledgeable model, right? Ping An would want an insurance knowledgeable model. So our job is to really build that layer for that purpose.
It's sort of like if I gave you a Windows kernel that is the core operating system, you wouldn't know what to do with it. You need all the Windows layers and the application interfaces so that the Windows kernel can be useful. And we like to think that 01.AI provides that layer for DeepSeek, which is the underlying model and technology.
Q: And using DeepSeek as the underlying model. Obviously, you could have chosen other models instead, including from the U.S. What does that tell you about where the competition or the race for supremacy stands between China and the United States right now.
A: Well, clearly in both U.S. and China, the pre-training of a giant model has consolidated and is consolidating. And it's become clear that open-source will be the winner. There's so many that will not concede, right? OpenAI, Anthropic who build their businesses believing they can build a better closed model than everyone else. I think they got shocked when they saw a model as good, they would argue nearly as good. But either way, they're getting it for much much lower cost.
If you think about OpenAI's cost——$7 billion of operating cost in 2024. DeepSeek probably operated with 2% of the operating expense. So the issue really isn't whose model is 1% better. I think they're all very good, but the issue is whether OpenAI's model is even sustainable. I mean, you're spending 7 billion dollars a year making a massive loss. And here you have a competitor coming in with an open-source model that's for free. And that company also is infinitely lasting because this founder has enough money to fund it at a current level and has reduced the cost of computing by a factor of 5 to 10. So with that kind of formidable competitor, I think Sam Altman is probably not sleeping well.
Q: Well, you do think, I mean, obviously, DeepSeek would have really challenged some of the monetization strategies around AI. But at the same time, given the AI is so commoditized now and in China as well, with the open-source approach, how many AI models do you think that the Chinese market can actually sustain?
A: Well, I think probably, I'm guessing three: DeepSeek, Alibaba, and ByteDance. And they'll evolve over time.
Q: OK. There are a few many names missing on that list.
A: Well, there will be many other who aspire to build one. And we'll see if there's a fourth player. But even within these three, I think DeepSeek currently has the momentum and that's why we're embracing that. But I also want to point out that I wouldn't say AI is commoditized. I would say the underlying pioneer model, foundation model that is trained with a huge amount of money, is commoditized. It costs a lot and it's hard to really compete with open-source. But AI is not commoditized. It's like, you know, Windows kernel, Linux kernel, that's kind of commoditized. But Windows is not commoditized. Microsoft makes a ton of money. And what 01.AI wants to do is build Windows and use DeepSeek as our Windows kernel.
Q: And is that where you think the value-add comes now? In other words, for VC's thinking about where the value added is in this industry, it's no longer, for example, AI still presents that value. And I wanna understand, for example, economics and why you think, say "there's only room for three."
A: There may be room for four, but in most markets you end up with room only for two. The winner who makes a ton of money, the runner up who breaks even and everyone else who dies. And this one is exacerbated because it's hard to monetize the underlying pioneer model. But once you're the winner, there are many ways to monetize it, right? So that's the trick. For the VCs in both US and China are moving away from wanting to fund more pioneer models. They're happy to fund AI infrastructure, AI applications, consumer apps, because every app will be remade. And also enterprise apps that can provide real value add to enterprises. So AI funding is going gangbusters. It's just that nobody wants to fund pioneer models.
Q: Does China's innovation, though, come up against the headwind this year of greater restrictions on China's access to advance semiconductor technology? Is that gonna end up presenting a problem?
A: We can look at it historically. China is able to catch up and build great engineering products, while US is better at making breakthroughs. That was what I argued in AI Superpowers, and that continues to be the case. Today, we actually see DeepSeek not only resilient, competitive, but also basically made true the statement that "necessity is the mother of innovation." Because they had so few GPUs, they had to make it 10 times more effective. I would say that DeepSeek actually benefited from having less resources and able to make training and inference at 5 to 10 times lower the expense incurred by OpenAI and others.
Q: You know, I remember so DeepSeek was, let's call it two months ago. I remember, which we had the chat and the show, I think was back in October, when you guys actually launch something similar months before DeeSeek. You guys came out, DeepSeek came out and then since DeepSeek came out that moment. Just the cadence and the frequency of all these ones coming up with their own iterations. Okay, I guess my question is, what do you make up the frequency? What is behind you think the frequency of these releases? And is DeepSeek the symptom or is it the cause of what's going on? In other words, what can we expect from the industry this year?
A: I think it's still ultra competitive. The fact that I think China will have 3 winners, US might have 4 winners at the end. It doesn't mean dozens of others will not want to compete. So I do expect frequent releases to continue, especially from the fast moving companies. And I think the fastest moving companies probably are xAI's Grok and DeepSeek. I see these are the fastest. And then I think OpenAI, Alibaba, Google, and Anthropic are, of course, all making respectable moves.
I listed ByteDance because they're not on the list by people, but we know how much they're spending. And I think one way you can justify spending a ton of money on pioneer models is because you have so many customers. When amortized over many customers, you can make the business case. And ByteDance has more customers and more ways to monetize than anybody. So I would make that speculation.
Q: Yeah, I think Baidu is one of the names that sort of stands out to me. Not on that list and maybe can share some reaction on that, but also government support from Beijing and exactly how do you see that also shaping up over the coming year?
A: We all respect and thank Baidu for really starting the Transformer revolution in China. They tried to hire Jeff Hinton and ended up hiring Andrew Ng. And I think that shows Robin has that kind of vision.
The Chinese government's new big direction is called new quality productivity. What that means is using high tech to create productivity, efficiency and really turn all the traditional industries into more profitable, more competitive by the use of high tech. And AI is at the very core of that.
Q: So our final question for you, it was rare that we saw the CEOs of China's big tech companies in the same room as the president, for example, weeks before that. And I'm wondering from tangibly, what measures do you think we could see from the government to be able to back up really what really see it as very good optics?
A: Yeah, I think right now the local governments probably are the most active, in trying to see. For instance, how would I apply AI to create this "new quality productivity" in these industries that will end up helping the growth of GDP in the city. And I think, having a single direction from the top, new quality productivity is very effective and invigorating to the cities. We're seeing a lot of interest for our company and our peer companies to work more closely with these local cities.
Q: Right. Kai-Fu, thanks so much for joining us. So it's a real pleasure to speak with you and congratulations on the release of the new platform as well.
*点击【此处】跳转彭博社报道链接
END