China is on observe to dominate shopper synthetic intelligence purposes and robotics manufacturing inside years, however the US will preserve its substantial lead in enterprise AI adoption and cutting-edge analysis, in accordance with Kai-Fu Lee, one of many world's most outstanding AI scientists and buyers.
In a uncommon, unvarnished evaluation delivered by way of video hyperlink from Beijing to the TED AI conference in San Francisco Tuesday, Lee — a former government at Apple, Microsoft, and Google who now runs each a significant enterprise capital agency and his personal AI firm — laid out a know-how panorama splitting alongside geographic and financial traces, with profound implications for each industrial competitors and nationwide safety.
"China's robotics has the benefit of getting built-in AI into a lot decrease prices, higher provide chain and quick turnaround, so firms like Unitree are literally the farthest forward on the planet when it comes to constructing reasonably priced, embodied humanoid AI," Lee mentioned, referring to a Chinese language robotics producer that has undercut Western opponents on value whereas advancing capabilities.
The feedback, made to a room stuffed with Silicon Valley executives, buyers, and researchers, represented one of the vital detailed public assessments from Lee concerning the comparative strengths and weaknesses of the world's two AI superpowers — and recommended that the race for synthetic intelligence management is turning into much less a single contest than a sequence of parallel competitions with totally different winners.
Why enterprise capital is flowing in reverse instructions within the U.S. and China
On the coronary heart of Lee's evaluation lies a basic distinction in how capital flows within the two international locations' innovation ecosystems. American enterprise capitalists, Lee mentioned, are pouring cash into generative AI firms constructing giant language fashions and enterprise software program, whereas Chinese language buyers are betting closely on robotics and {hardware}.
"The VCs within the US don't fund robotics the way in which the VCs do in China," Lee mentioned. "Identical to the VCs in China don't fund generative AI the way in which the VCs do within the US."
This funding divergence displays totally different financial incentives and market buildings. In the US, the place firms have grown accustomed to paying for software program subscriptions and the place labor prices are excessive, enterprise AI instruments that enhance white-collar productiveness command premium costs. In China, the place software program subscription fashions have traditionally struggled to achieve traction however manufacturing dominates the financial system, robotics provides a clearer path to commercialization.
The end result, Lee recommended, is that every nation is pulling forward in several domains — and will proceed to take action.
"China's bought some challenges to beat in getting an organization funded in addition to OpenAI or Anthropic," Lee acknowledged, referring to the main American AI labs. "However I believe U.S., on the flip facet, could have bother creating the funding curiosity and worth creation within the robotics" sector.
Why American firms dominate enterprise AI whereas Chinese language corporations battle with subscriptions
Lee was specific about one space the place the US maintains what seems to be a sturdy benefit: getting companies to truly undertake and pay for AI software program.
"The enterprise adoption will clearly be led by the US," Lee mentioned. "The Chinese language firms haven’t but developed a behavior of paying for software program on a subscription."
This seemingly mundane distinction in enterprise tradition — whether or not firms pays month-to-month charges for software program — has change into a vital issue within the AI race. The explosion of spending on instruments like GitHub Copilot, ChatGPT Enterprise, and different AI-powered productiveness software program has fueled American firms' potential to take a position billions in additional analysis and improvement.
Lee famous that China has traditionally overcome comparable challenges in shopper know-how by creating different enterprise fashions. "Within the early days of web software program, China was additionally effectively behind as a result of folks weren't keen to pay for software program," he mentioned. "However then promoting fashions, e-commerce fashions actually propelled China ahead."
Nonetheless, he recommended, somebody might want to "discover a new enterprise mannequin that isn't simply pay per software program per use or per 30 days foundation. That's going to not occur in China anytime quickly."
The implication: American firms constructing enterprise AI instruments have a window — maybe a considerable one — the place they’ll generate income and reinvest in R&D with out dealing with critical Chinese language competitors of their core market.
How ByteDance, Alibaba and Tencent will outpace Meta and Google in shopper AI
The place Lee sees China pulling forward decisively is in consumer-facing AI purposes — the type embedded in social media, e-commerce, and leisure platforms that billions of individuals use day by day.
"By way of shopper utilization, that's more likely to occur," Lee mentioned, referring to China matching or surpassing the US in AI deployment. "The Chinese language giants, like ByteDance and Alibaba and Tencent, will certainly transfer rather a lot sooner than their equal in the US, firms like Meta, YouTube and so forth."
Lee pointed to a cultural benefit: Chinese language know-how firms have spent the previous decade obsessively optimizing for person engagement and product-market slot in brutally aggressive markets. "The Chinese language giants actually work tenaciously, and so they have mastered the artwork of determining product market match," he mentioned. "Now they’ve so as to add know-how to it. So that’s inevitably going to occur."
This evaluation aligns with current business observations. ByteDance's TikTok grew to become the world's most downloaded app by means of refined AI-driven content material advice, and Chinese language firms have pioneered AI-powered options in areas like live-streaming commerce and short-form video that Western firms later copied.
Lee additionally famous that China has already deployed AI extra broadly in sure domains. "There are a variety of areas the place China has additionally finished an incredible job, equivalent to utilizing pc imaginative and prescient, speech recognition, and translation extra broadly," he mentioned.
The shocking open-source shift that has Chinese language fashions beating Meta's Llama
Maybe Lee's most hanging information level involved open-source AI development — an space the place China seems to have seized management from American firms in a remarkably quick time.
"The ten highest rated open supply [models] are from China," Lee mentioned. "These firms have now eclipsed Meta's Llama, which was once primary."
This represents a big shift. Meta's Llama models had been broadly seen because the gold commonplace for open-source giant language fashions as lately as early 2024. However Chinese language firms — together with Lee's personal agency, 01.AI, together with Alibaba, Baidu, and others — have launched a flood of open-source fashions that, in accordance with varied benchmarks, now outperform their American counterparts.
The open-source query has change into a flashpoint in AI improvement. Lee made an intensive case for why open-source fashions will show important to the know-how's future, at the same time as closed fashions from firms like OpenAI command larger costs and, typically, superior efficiency.
"I believe open supply has a lot of main benefits," Lee argued. With open-source fashions, "you may study it, tune it, enhance it. It's yours, and it's free, and it's necessary for constructing if you wish to construct an utility or tune the mannequin to do one thing particular."
He drew an analogy to working methods: "Individuals who work in working methods liked Linux, and that's why its adoption went by means of the roof. And I believe sooner or later, open supply may also enable folks to tune a sovereign mannequin for a rustic, make it work higher for a selected language."
Nonetheless, Lee predicted each approaches will coexist. "I don't assume open supply fashions will win," he mentioned. "I believe similar to now we have Apple, which is closed, however gives a considerably higher expertise than Android… I believe we're going to see extra apps utilizing open-source fashions, extra engineers wanting to construct open-source fashions, however I believe extra money will stay within the closed mannequin."
Why China's manufacturing benefit makes the robotics race 'not over, however' practically determined
On robotics, Lee's message was blunt: the mixture of China's manufacturing prowess, decrease prices, and aggressive funding has created a bonus that will likely be troublesome for American firms to beat.
When requested straight whether or not the robotics race was already over with China victorious, Lee hedged solely barely. "It's not over, however I believe the U.S. continues to be able to developing with one of the best robotic analysis concepts," he mentioned. "However the VCs within the U.S. don't fund robotics the way in which the VCs do in China."
The problem is structural. Constructing robots requires not simply software program and AI, however {hardware} manufacturing at scale — exactly the type of built-in provide chain and low-cost manufacturing that China has spent many years perfecting. Whereas American labs at universities and firms like Boston Dynamics proceed to supply spectacular analysis prototypes, turning these prototypes into reasonably priced industrial merchandise requires the manufacturing ecosystem that China possesses.
Corporations like Unitree have demonstrated this benefit concretely. The corporate's humanoid robots and quadrupedal robots price a fraction of their American-made equivalents whereas providing comparable or superior capabilities — a price-to-performance ratio that might show decisive in industrial markets.
The power infrastructure hole that might decide AI supremacy
Underlying many of those aggressive dynamics is an element Lee raised early in his remarks: power infrastructure. "China is now constructing new power initiatives at 10 instances the speed of the U.S.," he mentioned, "and if this continues, it is going to inevitably result in China having 10 instances the AI functionality of the U.S., whether or not we prefer it or not."
This statement connects to a theme raised by a number of audio system on the TED AI convention: that computing energy — and the power to run it — has change into the elemental constraint on AI improvement. If China can construct energy crops and information facilities at 10 instances the speed of the US, it may merely outspend American opponents in coaching ever-larger fashions and operating them at ever-greater scale.
Lee famous this dynamic carries "very actual nationwide safety implications for the U.S." — although he didn’t elaborate on what these implications could be. The remark appeared to reference rising considerations in Washington about technological competition with China, notably in areas like AI-enabled navy methods, surveillance capabilities, and financial competitiveness.
Regardless of the US presently internet hosting a number of instances extra AI computing energy than China, Lee warned that "this lead is rising" for now however may reverse if power infrastructure investments proceed at present charges.
What worries Lee most: not AGI, however the race itself
Regardless of his typically measured tone about China's AI improvement, Lee expressed concern about one space the place he believes the worldwide AI group faces actual hazard — not the far-future threat of superintelligent AI, however the near-term penalties of shifting too quick.
When requested about AGI risks, Lee reframed the query. "I'm much less afraid of AI turning into self-aware and inflicting hazard for people within the quick time period," he mentioned, "however extra fearful about it being utilized by unhealthy folks to do horrible issues, or by the AI race pushing folks to work so exhausting, so quick and livid and transfer quick and break issues that they construct merchandise which have issues and holes to be exploited."
He continued: "I'm very fearful about that. In reality, I believe some horrible occasion will occur that will likely be a get up name from this form of drawback."
Lee's perspective carries uncommon weight due to his distinctive vantage level spanning each Chinese language and American AI improvement. Over a profession spanning greater than three many years, he has held senior positions at Apple, Microsoft, and Google, whereas additionally founding Sinovation Ventures, which has invested in additional than 400 firms throughout each international locations. His AI firm, 01.AI, based in 2023, has launched a number of open-source models that rank among the many most succesful on the planet.
For American firms and policymakers, Lee's evaluation presents a fancy strategic image. The USA seems to have clear benefits in enterprise AI software program, basic analysis, and computing infrastructure. However China is shifting sooner in shopper purposes, manufacturing robotics at decrease prices, and probably pulling forward in open-source mannequin improvement.
The bifurcation means that reasonably than a single "winner" in AI, the world could also be heading towards a know-how panorama the place totally different international locations excel in several domains — with all of the financial and geopolitical issues that suggests.
Because the TED AI conference continued Wednesday, Lee's evaluation hung over subsequent discussions. His message appeared clear: the AI race is just not one contest, however many — and the US and China are every successful totally different races.
Standing within the convention corridor afterward, one enterprise capitalist, who requested to not be named, summed up the temper within the room: "We're not competing with China anymore. We're competing on parallel tracks." Whether or not these tracks finally converge — or diverge into totally separate know-how ecosystems — will be the defining query of the following decade.
