彭博社專訪李開復：相較美國AI巨頭“贏家通吃”打法，中國AI企業以開源生態走出自主創新之路

推薦語：

分享我這周在彭博社的采訪：不同于美國“贏家通吃”的寡頭模式，中國企業正以開源生態實現“換道超車”，有望在AI的黃金十年引領應用層創新。

以下內容轉載自：

近日，零一萬物CEO李開復博士接受彭博社專訪，系統性闡釋了他對中美大模型發展路徑差異、商業模式演進及未來技術趨勢的深度觀察。

他深度剖析了中美兩國在技術路線上的根本分歧——以OpenAI為代表的美國路徑，正以豪賭式投入押注通用人工智能（AGI）的終極勝利；而中國模型公司則憑借開源模式走出了一條迥異的“協同進化”之路：依托開源生態、強調平臺共建與應用創新，在理性投入中尋求可持續增長。

零一萬物CEO李開復博士

在李開復博士看來，美國模型公司所選擇的路徑本質上是一種“贏家通吃”的雄心壯志：一家公司率先研發出通用人工智能（AGI），然后碾壓其他公司，贏家通吃，最終獨攬數萬億美元的利潤。

科技巨頭豪賭式的投入也使得“泡沫論”甚囂塵上，李開復博士明確指出：當前市場議論的所謂“泡沫”，并非否定AI技術本身的顛覆性潛力，而是部分公司估值的上漲階段性地超前于商業化兌現的節奏。

對比之下，中國企業則更愿意通過開源生態，進行某種程度上的“合作”，集力構建起強大的平臺，使得每個人都能從平臺的知識共享中受益。同時，他也清晰描繪了大模型初創公司在全球AI競賽中的獨特優勢與可持續發展的路徑。

在李開復博士看來，大模型初創企業更務實的生態卡位是，依托科技巨頭的開源基座模型，聚焦應用層，避免陷入資源劣勢。以零一萬物為例，李開復博士表示，零一萬物選擇基于全球頂尖開源模型進行繼續訓練、模型微調與工程化落地，既保障企業客戶的數據主權與私有化部署需求，也能夠大幅壓縮研發成本。

但從基座模型到企業場景，并非“開箱即用”。具備復雜任務規劃與執行能力的AI Agent已經成為企業最直接地從AI技術上獲得增長價值的方式。具備復雜任務規劃與執行能力的AI Agent，正成為企業實現AI價值轉化、驅動業務增長的關鍵路徑。

零一萬物萬智企業大模型一站式平臺上已推出企業級Agent“超級員工”，并搭配有開發工具與配置平臺，支持企業快速定制專屬智能體。目前，零一萬物萬智平臺已匯聚政務、金融、工業（電力和制造業）、辦公、銷售等5大行業的30類“超級員工”，企業可實現AI Agent能力的快速接入和快速應用。

萬智企業大模型一站式平臺：b.01.ai

李開復博士表示，AI Agent以大語言模型為“大腦”，搭配記憶能力與任務執行能力，將帶來“基于結果”的商業模式再造，成為推動企業智能化變革的核心力量。AI Agent的重要性不僅在于提升效率，更在于重構組織形態乃至社會協作方式。未來的企業競爭力，或將不再取決于人力規模，而在于其構建、調度與進化AI Agent網絡的能力。

以下為對話全文：

大模型領域存在“泡沫”？

開源改變競爭形態

彭博社：

過去三年間，您如何看待中國人工智能與美國人工智能的整體發展演進？中國在人工智能領域具備哪些優勢，又面臨哪些挑戰？

李開復：首先，我對“美國AI領域存在嚴重泡沫”這一觀點持不同看法。以OpenAI為例：其年收入約120億美元，年投入高達400億美元——表面看，財務損益表極不合理；但關鍵的一點在于，這400億美元中絕大部分屬于戰略性前置投入，用來支撐未來的收入增長預期。如果你相信未來三年OpenAI的營收能實現兩三倍甚至五倍的增長，目前OpenAI的超高估值是否合理還沒法拍板定論，但這估值泡沫下肯定有某些實質性的價值存在。

當前市場議論的所謂“泡沫”，實際上并不是否定AI未來的增長潛力，而是估值已經階段性地超出了業績兌現的節奏。換句話來說，我并不是說當前估值是完全合理的，但必須承認：OpenAI的內在價值是不容忽視的。

再來看中國人工智能領域，和美國人工智能發展至今的本質區別在于閉源還是開源。美國模型所選擇的路徑本質上是一種“贏家通吃”的雄心壯志：一家公司率先研發出通用人工智能（AGI），然后碾壓其他公司，贏家通吃，最終獨攬數萬億美元的利潤。以OpenAI、Anthropic、谷歌等為代表的頭部企業，都汲汲營營大手筆下注。這種邏輯很像一位年輕的天才篤信自己終將獲得諾貝爾獎，孤注一擲地全力往前沖。

相較之下，中國更傾向于“協同進化”：大家攜手合作，共享成果。所有人在競爭中開放共享，A借鑒B的思路，B學習C的長處。這更像一群聰明而非“天才型”的學生組成了專題學習小組。

彭博社：

開源是否是導致許多中國人工智能創業公司無法盈利的原因？

李開復：當前中美業界對人工智能有一項高度共識：人工智能將帶來人類歷史上最具顛覆性的技術革命。所以人們也愿意相信人工智能的商業化進程終將到來。只是節奏有先后，美國率先啟動 AI巨大的商業化機會，中國將緊隨其后。因此，中國企業愿意投入通過開源生態，進行某種程度上的“合作”，集力構建一個強大的平臺，那么每個人都能夠從平臺的知識共享中受益。或者，如果其中有一家公司脫穎而出，那么它就有機會成為下一個偉大的公司。所以，人們樂觀地相信人工智能正在改變世界。在當前的AI模型競爭范式中，閉源模型的信念是，一家公司最終會碾壓其他所有公司。而開源模式則像是，“先把路修起來，再看誰跑得最快。”

彭博社：

開源模式是否在某種程度上改變了企業的競爭形態？——就像在課堂上，學生不需要從頭解題，而是參考最優解法，在此基礎上進一步優化與拓展。

李開復：這恰恰是我所在的零一萬物的商業模式。我們已全面轉向“基于開源做創新”：誰的模型最優，我們就用誰的；無論是深度求索（DeepSeek）還是通義千問（Qwen），我們一視同仁、完全中立。某種程度上，我們的處境與英偉達（NVIDIA）很相似。當然，英偉達的處境要好得多，但核心邏輯一致：我們不押注任何一家勝出，只關注開源生態能否持續向前推進。

這種模式的優勢顯而易見：開源模型近乎零成本，我們在此基礎上進行應用開發的成本會低很多。另外，我們可以自由地獲取、改造這些開源模型。對企業用戶而言，價值更為直接：他們可以將模型私有化部署，全程本地化部署運行；如果存在不適合上傳云端的保密數據，企業也可以遵從自定義的規范，把模型自主權掌握在手中。

相較于美國“贏家通吃”

更看好中國協同創新

彭博社：

但創新終究需要有人買單——那么，誰來承擔？不妨以OpenAI為例，資本開支就高達萬億元量級，各方仍期待OpenAI能夠探索出如何解決這么高的成本負擔。長期來看，OpenAI需要解決如何收回前期巨額投入的難題。在中國的情況完全不同。由于模型是開源的，無需付費，那么技術創新要如何募集資金？又該如何向投資者證明他們的投資會有合理的回報？

李開復：實際上，我更看好中國模式。雖然阿里巴巴、字節跳動、百度等企業投入巨大，但它們愿意持續注資，因為AI對其核心業務具有戰略意義，無論是電商、搜索還是社交方面。

這些企業的管理者都是聰明的商人，懂得平衡之道。目前它們的投入規模與OpenAI相比只是九牛一毛。OpenAI的燒錢策略是一場高風險的豪賭，賭的是贏家通吃的野心。而中國巨頭們則采取更審慎的投入策略：它們會通盤考量整體財務狀況，確保交出穩健的業績報表。我們目前還沒有見到這些企業報告出現巨額虧損。

因此，對財務投資者而言，中國模式顯然風險是更可控的。美國市場追捧贏家通吃法則，投資者必須押注個別贏家，這意味著要么賺得盆滿缽滿，要么輸得一敗涂地。

彭博社：

您如何看待中國AI領域的發展格局？資金雄厚的阿里、字節等科技巨頭中，誰有足夠的財力繼續進行這場人工智能競賽？它們與MiniMax、月之暗面（Moonshot）等新興大模型初創公司相比，競爭優勢究竟如何？

李開復：我欣賞那些選擇參與模型競爭的初創企業的勇氣。我認為對初創企業來說更務實的選擇是，依托科技巨頭的基座模型，專注于開發應用，面向企業端或是面向消費者端都會有機會。畢竟現階段，中國初創企業很難像OpenAI那樣募集巨額資金，甚至連其十分之一的規模都難以企及。在資金受限的情況下，企業發展必然面臨挑戰。雖然中國初創企業和科技巨頭都擁有優秀人才，但兩者之間存在著資金投入的差距、人才規模與密度的差距，初創企業若想自主攻堅基礎模型，將會面臨一場艱辛的征程。

彭博社：

DeepSeek的崛起，我們或許可以總結為“兩個關鍵時刻”，今年1月它真正進入了全球視野，而實際上早在去年此時，業界就已開始討論DeepSeek。作為業內深度參與者，大家現在更關心的是：下一個突破點將會是什么？當前正在孕育哪些重要創新？面對即將到來的2026年，我們應該做好哪些準備？

李開復：我認為在基礎大模型領域，美國確實處于領先地位。但中國正在迅速借鑒美國的先進成果，同時融入自身優勢，目前差距大約保持在三到六個月。這個判斷我堅持了一年半，至今依然成立。

在我看來，中國要孕育出真正重塑全球技術范式的重大突破，仍需時間積累。目前更可能誕生的是讓人眼睛一亮的創新消費級應用，最具潛力的當屬硬件領域：比如一家名為Plaud的企業，其開發的AI裝置能持續感知，自動生成會議紀要、規劃待辦事項并發送個性化提醒。這僅僅是個開始，在中國智造的基礎上，這類智能硬件的性價比優勢顯著，迭代速度更快，產品形態也日益精巧。

正如我們之前討論過的，中國風投對大語言模型投資確實持審慎態度，但正全力押注硬件方向，包括我所創立的創新工場也在積極布局。當人們談論硬件時，往往首先想到機器人，特別是人形機器人，但這些領域目前存在估值過高的問題。相比之下，投資輕巧的智能硬件顯然更為明智。

當這波人工智能顛覆邁進下一個階段，我們會發現手機不再是AI的最佳載體。無論是智能眼鏡、手表、手環、微型項鏈乃至近乎隱形的微小設備，都可能成為新終端。因為AI的首要特性是持續在線、隨時感知、實時思考，它將成為你的智能外接硬盤，完整記錄所見所為。當然，您可能接著會提到隱私議題，那得單獨探討了。

彭博社：

這有點“細思極恐”啊。

李開復：也許我的設備已經在聽你的每句話了。（笑）

彭博社：

我的天哈哈哈哈。

彭博社：

最近AI智能體確實成了熱議焦點。您能否詳細談談這股熱潮背后的原因？據您判斷，這類技術距離真正成為主流還需要多久，特別是針對企業應用？

李開復：智能體其實已經真實存在，只是尚未普及。它的底層依托于大語言模型，大語言模型就像是“大腦”，并且甚至可能是多個各司其職的專用模型協同工作。但智能體最核心的是兩大支撐要素：首先是記憶能力。它能記住你的職業背景、行為軌跡、企業信息和規章制度，這使得它能針對行業特性提供服務，而不是像通用聊天機器人那樣泛泛而談。但更關鍵的是任務執行能力，它不再僅僅為你生成旅行計劃，而是直接完成機票酒店的預訂執行；不再停留于戰略建議層面，而是直接落地實施；它不僅僅是提供廣告策略建議，AI就能直接完成廣告投放的工作流。

本文摘編翻譯自彭博社

Bloomberg China Show：

Q:

In the last three years though, how have you looked at this whole evolution of China AI versus the US? And what are the advantages and challenges for China AI?

A:So firstly, I disagree on the prospects of the U.S. AI because you look at OpenAI, it's making 12 billion, spending 40 billion. Looks like a bad balance sheet, but most of that 40 billion is spent for future revenues. And if you believe there are 2 x, 3 x, 5 x growth for the next three years, it's going to justify that valuation at some point. The bubble is merely that it's gotten ahead of itself, not the likelihood of growth in the future. Not saying it's worth its price. But there's absolute substance is under the foam.

Now back to the China AI, the fundamental difference is actually closed source versus open source. Yes, the US model is necessarily propelled by this ambition and dream of one company reaching AGI first and squashing everyone. Winner take all, worth trillions of dollars. That's what OpenAI, Anthropic, Google and others are betting on. And it's like a genius kid who thinks he or she will win the Nobel Prize. The China approach is -- hey, let's all work together and share. We compete, but let's open source our models. So A learns from B, B learns from C. It's more like a study group of smart kids, maybe not geniuses.

Q:

Is the open source a reason why a lot of these Chinese AI startups can not monetize?

A:Well, I think there's a common belief in US and China that AI is the most important technological revolution ever. There is a belief that this monetization will happen. It's happening first in the US, and it will happen later in China. So there is a bet that if all of China kind of pseudo collaborates through open source, then builds a great platform, then everyone can benefit because the knowledge of the platform. Or if one company comes out ahead, there's a chance to leap ahead and be the next great company as well. So there's tremendous optimism that AI is changing the world. With these models, the closed-source approach is a belief that one company will squash everyone else. And the open source is like, 'Hey, let’s share for now, and we’ll see what happens.'

Q:

Has the open source model in some ways discouraged competition? Because I simply have to look across the classroom and see who has the best answer and simply build on that classmate.

A:Well, that is the business of my company 01.AI. We've pivoted ourselves to build on open source. Whoever has the best model, we will use it. If it's DeepSeek or Qwen, we are agnostic. So we're in a position not so dissimilar from Nvidia. Obviously, Nvidia is much better, but they don't care who wins. We don't care either. We only care that open source moves forward. And it's benefits are that open source is free. It's much much lower cost when it's basically free and we can build on it. We take it, we do whatever we want. And also enterprises, they can take the model in-house, run everything in-house. If they have secret data they don't want to send it to the cloud, well, they don't have to.

Q:

But someone needs to pay for the innovation though. Who? I'll just take the OpenAI example, but you know, there's a complete difference in the Chinese context. Well, you have capex of 1 trillion and people are trying to figure out how are they gonna pay for that. And then people, to your point, longer term, will then have to worry about how they're gonna pay for that. In China's case, when you don't have to pay for it, where it's open source. How is the innovation getting funded? And how do you justify to those asset allocators that they will get a return for their investment?

A:Well, I actually like the Chinese model better. While the likes of Alibaba, ByteDance, Baidu they're paying a lot of money, they're willing to subsidize because it's fundamental to their core business, whether it's e-commerce or search or a social.

And I think they're smart business people. They'll do the balancing. They'll start spending a tiny fraction of the OpenAI spending. Now, the OpenAI's spending is a big bet with big return. It's betting that it will win. Whereas I think Alibaba, Baidu, Tencent, ByteDance, if they're spending more prudently, which they are, they will be able to look at the whole balance sheet and deliver a reasonable result. We don't see any of these people reporting huge losses.

So I think the Chinese model is more prudent for a financial investor. The US is a winner take all, so you pick your winner. If you win, you win big. If you lose, you lose big.

Q:

How do you see it though play out in China? Whether you have the big AI giants like Alibaba, ByteDance, who have the deep pockets to continue on with this sort of AI war in China? How do they stack up against some of these startups like MiniMax or Moonshot?

A:Well, I admire the courage of the smaller companies who choose to compete. I feel that it's more prudent for a smaller company to build either applications or either enterprise or consumer and betting on the fundamental foundation models of the big companies. Because I don't think it's very likely that any startup can raise money the way OpenAI does or even one tenth of OpenAI. And with that, with less money, it is going to be difficult. The talents are obviously very good in the Chinese startups as well as the large companies. But the lack of spending and the lack of huge number of talent and talent density make it an uphill struggle for the startups that try to do that.

Q:

The DeepSeek moment, well, let's call it two moments, right? Because the January moment this year was where it became public to the world. Although it was this time last year where people had started to already talk about DeepSeek, and I want, people are now looking for what's the next one you're connected to and where connected into the industry. What innovations are taking place right now? What should we be prepared for going into next year?

A:Well, I think in terms of the fundamental foundation models, the US is head. But China is very rapidly figuring out the great things US is doing, plus a few of its own good things to basically trail by three to six months. I've been saying this for a year and a half, and it still remains the case.

So I don't think there's gonna be a huge break-out change the world technology in China. That's not likely. What is more likely is that there may be great consumer apps. And what's even much much more likely is Chinese hardware. There's a company called Plaud. It is basically an ambient AI that listens all the time and gives you summaries of all the meetings, tells you what to do, and gives you reminders. And that's just the beginning. That is so much cheaper to build in China, and it's iterated much faster and it's getting smaller.

And I think the Chinese VCs are hesitant to invest in large language models for the reasons we discussed. But they are going all out, including my own Sinovation Ventures, are betting on hardware. When people talk hardware, people think always embodied robotics and humanoid robots. But those are also a bit overpriced. It's much more prudent to invest in small devices.

Because when AI revolution moves forward at the next step, we will see that the best place, the best device on which to use AI is not the phone. Whether it is a glasses or watch or wristband or a little necklace or almost invisible, are all possible. Because AI first means it's always on, always listening, always thinking and it's your external hard drive that’s smart and remembers everything you see and did. I know you're gonna talk about privacy, I know, but that's a separate discussion.

Q:

It gets a bit creepy.

A:I might be listening to you now.

Q:

My goodness, hahaha.

Q:

There's a lot of talks about AI agents. Can you tell us more what's all the buzz around it and how long before it becomes pretty mainstream, when it comes to enterprises?

A:It's real already, but it's not yet mainstream. An agent is built on the larger language model. The large language model is the brain. And there may be multiple large language models, each doing special things. But the two most important things wrap around it. One is memory. It knows who you are, what you did, who's your company, what your policies are. So it's relevant to what you do as opposed to a general chatbot. But the other even bigger thing is that it executes tasks. So instead of giving you a travel itinerary, it gets the travel booked, your airline, your hotel, instead of suggesting what's the strategy, it starts executing. Your ads are placed as opposed to suggesting what ads you might do.