There have been some changes involving two individuals recently that I’ve been wanting to discuss together.
One is Li Kaifu. On May 28, an internal letter from 01.AI marking its third anniversary was leaked, providing the outside world with its first clear glimpse that this once high-flying "Chinese OpenAI" star company is proactively adjusting its positioning.
In that letter, Li Kaifu spoke less about grand visions of AGI and Scaling Law, and more about applications, Agents, commercialization, and the pragmatic struggle for survival and revenue.
01.AI's benchmark has also completely shifted. It has quietly changed from being the original OpenAI to the extremely low-key Silicon Valley company Palantir. Palantir started by providing data analysis for governments and large enterprises, burned cash for 17 years before going public, and has a rather un-sexy core business. Yet, today its market cap on the U.S. stock market exceeds $330 billion.
One line in Li Kaifu's letter was particularly poignant: To withdraw is to turn off the light just before dawn breaks.
While giving up right before the dawn is regrettable, he chose to proactively exit the old battlefield before the first light.
The new goal he set for 01.AI is: To become China's first profitable AI 2.0 company by 2026, and to IPO by 2027. As of May, 01.AI has announced cumulative order amounts exceeding 15 billion RMB.
The other is Wang Xiaochuan. In a recent interview, his words were even more blunt than an internal letter. This man, who was once the most staunch believer that large models would reshape search, began repeatedly emphasizing that the gap in foundational models with the U.S. is widening, and that the absolute红利 (dividend/advantage) of pre-training has ended.
He said something worth pondering: When the company was approaching its second anniversary, he suddenly didn't know what he was doing, what value he was creating.
He even added nonchalantly in the interview: We weren't short of money back then either.
It sounds like a clarification, but it's actually a heavier confession—it wasn't about money, it was that he himself couldn't see the point anymore.
Baichuan's approach was more radical. In May, it officially launched its new-generation medical large model M4 and the AI Family Doctor product. Previously, their model scored high on OpenAI's medical evaluation benchmark HealthBench, and the team began shifting its core resources to fully All-in on the vertical battlefield of healthcare.
It's worth mentioning that Baichuan still has nearly 30 billion RMB in cash on its books. In other words, this All-in move isn't a desperate escape backed into a corner, but a proactive retreat after seeing the situation clearly.
If we rewind to the peak frenzy of the early large model era, these two individuals would never have spoken like this.
Back in 2023, during the hottest period of China's "War of a Thousand Models," they were almost among the most excited, most idealistic, and most ambitious leaders trying to redefine the era. At that time, the entire Chinese tech industry believed firmly that what Silicon Valley could achieve, China could also achieve through its engineering红利 (dividend/advantage).
It was a轰轰烈烈的 (grand and spectacular) alchemy movement.
I. That Year, Was the Most Idealistic Year for Chinese AI
In 2023, almost the entire Chinese tech industry was restarting from scratch.
Li Kaifu started 01.AI, Wang Xiaochuan started Baichuan Intelligence, Yang Zhilin started Moonshot AI, Zhang Peng started Zhipu AI, Jiang Daxin started StepFun. Everyone was telling the same story: The Chinese Version of OpenAI.
VCs poured money in疯狂ly (frantically), screens were filled with talk of the "Six Little Tigers," the "next ChatGPT," the "new AI era"...
The core industry sentiment at that stage wasn't about making money, products, or users; it was about who resembled OpenAI the most. So all companies began疯狂ly (frantically) competing on parameters, Benchmarks, multimodality, and Scaling Law.
Looking back today, that phase was reminiscent of the全民门户 (nationwide portal) era of the early internet, or comparable to the全民造车 (nationwide car-making) frenzy when new energy first exploded. Everyone felt the era was just beginning, and there was still time for everything. But problems soon emerged.
Starting in 2024, China's large model industry suddenly realized this wasn't mobile internet; it was more like a heavy industry war.
The further models progressed, the clearer it became that what truly determined victory wasn't ideas, but GPUs, computing power, data, electricity, capital, and inference costs.
This industry was becoming increasingly unsuitable for startups.
OpenAI has Microsoft behind it, Anthropic has Google and Amazon, xAI has Musk and Nvidia.
What was truly suffocating was a set of capital expenditure figures disclosed by these American giants for 2026. Microsoft planned $190 billion for the year, Google $180-190 billion, Amazon $200 billion, Meta raising its上限 (ceiling) to $145 billion.
The combined money these four companies are throwing down this year exceeds $7250 billion. This figure is接近 (close to) Japan's annual fiscal expenditure, even more than Germany's annual national budget.
A year ago, their combined total was only $4100 billion. In one year, they increased spending by over $3000 billion.
China's tech giants are already catching up. ByteDance initially set its 2026 AI capital expenditure at 160 billion RMB, but上调 (raised) it to 200 billion a few months ago. Alibaba announced last year it would invest 380 billion RMB over the next three years in AI and cloud. At the earnings call in May this year, Alibaba's CEO Wu Yongming said that figure was set too low and needs further追加 (additional investment). Tencent's pace is slightly slower, but its Q1 2026 single-quarter capital expenditure had already reached 31.9 billion RMB.
Putting these numbers together, the处境 (situation) of Chinese AI startups is very clear: they raise at most a few hundred million dollars a year.
This is not a war on the same scale.
Even more残酷 (cruel) is that OpenAI resets the industry every few months. GPT-4o arrived, Sora arrived, Agent arrived, reasoning models arrived. Each upgrade meant the entire Chinese industry had to catch up again from scratch.
Many companies only then看清 (saw clearly) their true position: they weren't competing with domestic peers; they were sandwiched between two layers of giants. Above them is the American AI industrial system, and beside them are Chinese internet giants.
This is the real turning point for China's AI industry after 2025.
II. The Six Little Tigers Are Long Gone as Six Tigers
In the early fervor of the large model era, the industry still talked about the "AI Six Little Tigers." The six names并列 (listed side by side), sounding like a整齐 (neat and uniform) troop.
Looking now, these six have already split onto three completely different paths.
The fastest runners are those who have already "reached shore."
Zhipu AI went public on the Hong Kong Stock Exchange in April 2026, with a current market cap as high as 700 billion HKD. MiniMax followed closely, its市值 (market cap) once exceeding 300 billion HKD. Moonshot AI secured a post-investment valuation of $20 billion. StepFun just completed a $2.5 billion financing round, aiming to be the third listed domestic large model stock. They are still walking the old path, continuing to burn money on the通用大模型 (general-purpose large model) road, cashing out valuations through the secondary market. (Extended Reading: The Eve of Large Model Market Clearing)
Walking in the opposite direction are Baichuan and 01.AI. One fully转向 (turned to) healthcare, the other对标 (benchmarked against) Palantir, both abandoning the identity of a通用大模型公司 (general-purpose large model company).
The remaining one is DeepSeek. Valuation rumors have reached $51.5 billion, walking an independent path of开源 (open-source) plus底层突破 (breakthrough at the foundational layer), not playing in the same game as the others.
There's a微妙 (subtle) logic worth mentioning here. Logically, Zhipu and MiniMax going public successively and their市值屡创新高 (market caps repeatedly hitting new highs) should have made all six tigers eager and excited, but what actually happened was the opposite effect.
There's a very委婉 (tactful) line in Wang Xiaochuan's interview. He said the first two listed companies踩在了 (stepped onto/rode on) the技术红利 (technological红利/dividend) of general models and the窗口 (window) of national科技强国 (tech powerhouse) support policies.
Then he paused and added, the成熟 (maturation) of AI healthcare will come a bit later.
Translated, it means: the tickets for the通用大模型 (general-purpose large model) ship have already been sold out.
Zhipu relies on over 8,000 B端机构客户 (B-side institutional clients), including 95 major私有化部署 (private deployment) clients like governments and state-owned enterprises/Central SOEs;政企订单 (government and enterprise orders) are its core support. MiniMax primarily targets海外市场 (overseas markets), with全球累计用户 (global cumulative users) of about 300 million to date,海外用户占比 (overseas user proportion) exceeding 70%, achieving rapid growth through its庞大的海外 C 端用户 (massive overseas C-side user base).
These two have essentially written the playbook for how a通用大模型公司 (general-purpose large model company) goes public. For others wanting to use the same script later, the估值溢价 (valuation premium) is gone, and market attention and investor patience have weakened significantly.
Li Kaifu and Wang Xiaochuan stepping back isn't entirely because they can't compete with OpenAI. It's also because they看清 (see clearly) one thing: the first two ships have sailed, and the资本窗口 (capital window) for通用大模型 (general-purpose large models) has essentially closed.
III. The First Batch of Entrepreneurs Begins to Step Back
Note:这里的后退 (this stepping back) is not failure.
It's them finally starting to let go of that original idealistic script.
Li Kaifu began emphasizing no longer chasing the strongest model, no longer执着于 (obsessing over) AGI, but focusing on applications, Agents, and commercialization.
In early January 2025, 01.AI merged most of its预训练与 AI Infra (pre-training and AI Infra) teams into Alibaba Cloud, which was initially误解 (misinterpreted) by outsiders as a "sellout." Looking back, this was Li Kaifu making the earliest proactive降速换挡 (deceleration and gear shift) among the six tigers,放弃 (abandoning) the money-burning race for超大规模模型 (ultra-large models),转向 (turning to) low-cost, implementable To B commercialization, also避开 (avoiding) the subsequent collective pressure risks of the industry.
Wang Xiaochuan押 (bet) all resources on specialized scenarios like healthcare and search.
Many companies started laying off基础模型团队 (foundational model teams),转向应用 (turning to applications), and talking about收入 (revenue).
This isn't regression; it's反而 (on the contrary) the real beginning of产业成熟 (industry maturation).
IV. Chinese AI Begins to Accept One Thing
The deepest change across the industry today is that Chinese AI is finally starting to accept that China's true advantage也许从来不在 (perhaps never lay in) foundational models.
What the U.S. is truly strong at is original model creation, Scaling Law, the GPU ecosystem,底层基建 (underlying infrastructure), and顶级科研 (top-tier scientific research). What China is truly strong at is场景 (scenarios), applications,制造业 (manufacturing),用户规模 (user scale),工程能力 (engineering capabilities), and商业化速度 (commercialization speed).
Put bluntly, the U.S. is more like the inventor of AI, while China is more like the one who turns AI into an industry.
So you see ByteDance疯狂做 (frantically making) AI products,腾讯开始强调 (Tencent beginning to emphasize) Agents,阿里开始强调 (Alibaba beginning to emphasize) Industrial AI. What DeepSeek is most successful at isn't AGI, but低成本推理 (low-cost inference).
The entire industry is shifting from "whose model is strongest" to "who can best embed AI into production scenarios."
V. Li Kaifu and Wang Xiaochuan Are People from Two Different Eras
I particularly want to elaborate on this part.
If you place the founders of the six little tigers together, you'll notice a clear断代 (generational gap).
Wang Xiaochuan is a veteran who grew up in the Sogou era. In 2003, at age 25, he personally led Sogou Search from scratch, later pushing Sogou all the way to a Nasdaq IPO; in 2021, Sogou was acquired by Tencent, he stepped down and achieved financial freedom.
Li Kaifu is the教父 (godfather) of the Innovation Works era, moving from Carnegie Mellon to Apple, SGI, Microsoft, Google, having witnessed more tech bubbles than one can count on two hands.
They are not AI-native entrepreneurs; they are individuals who, using the business intuition of the previous internet generation, saw the时代红利 (era红利/dividend) and jumped into the trenches.
Whereas梁文锋 (Liang Wenfeng, DeepSeek),杨植麟 (Yang Zhilin, Moonshot),闫俊杰 (Yan Junjie, MiniMax? Note: Article earlier mentioned Zhang Peng for Zhipu and Jiang Daxin for StepFun) this generation更像是 (seem more like) the native generation of this AI wave. Younger, with purer technical backgrounds, and a more宗教化 (religious-like) belief in AGI.
They all once渴望 (hoped) large models could replay the miracle of mobile internet. But Li Kaifu and Wang Xiaochuan are now slowly realizing that AI isn't that script.
Mobile internet played with流量 (traffic),产品体验 (product experience), and用户规模 (user scale) to leverage small forces for big impact; a single breakthrough could tear open a trillion-dollar赛道 (track/field). AI is different; its护城河 (moat) is physical—the footprint of data centers, the carrying capacity of power grids, the产能 (production capacity) of chip foundries. These things cannot be bypassed by a startup's agility.
Thus, the internet veterans who see through the底牌 (cards on the table) begin choosing to retreat from the虚无缥缈的改变世界 (nebulous 'changing the world') to first earning enough to维持生存 (sustain survival).
Veterans of the internet era have experienced too many waves of大浪淘沙 (the sands of time).
They know all too well when to step on the gas and when to change to a safer stance. The AI-native generation holds another kind of信念 (belief) to persevere against odds; they believe this time is different, AGI will definitely arrive, and they must be the ones to build that god.
This path cannot be judged right or wrong. But what is certain is that the上半场 (first half) of China's large model entrepreneurship was ignited by these internet veterans with their influence and idealism, and the接下来漫长的消耗战 (long war of attrition to follow) is destined to be fought by these pure-blooded技术派 (technologists) native to AI.
Li Kaifu and Wang Xiaochuan's strategic pivots, in a sense, represent the completion of a跨时代的交棒 (handover across eras). But this isn't necessarily a bad thing.
China's tech industry has actually experienced too many similar幻灭时刻 (moments of disillusionment) over the past decade-plus.
When new energy first exploded, everyone talked about颠覆 (overturning) century-old traditional automakers. Later, everyone became老老实实 (honest/realistic), sitting in offices meticulously calculating the碳酸锂成本 (lithium carbonate cost) of every battery cell. When云计算 (cloud computing) was hottest, everyone thought it was the next-generation human operating system. Later, everyone started老老实实讨论 (honestly discussing)私有云交付 (private cloud delivery) and现金流 (cash flow).
The true sign of a tech industry's成熟 (maturity) is when it stops迷信 (blindly believing in) grand idealism and starts sitting down to calculate the ROI of every single penny.
The large model industry has finally熬到了 (endured to reach) this moment.
VI. From Deity-Making Back to Business
What's worth noting recently about Li Kaifu and Wang Xiaochuan isn't what new products they've launched.
It's their姿态 (posture) itself.
They represent the清醒 (clarity/sobriety) of China's first batch of large model idealist entrepreneurs. Even they have begun重新校准 (recalibrating) their distance from reality. This意味着 (means) that the most狂热 (fanatical) phase of China's large model entrepreneurship has passed.
What remains is to be handed over to真正的合同 (real contracts),真正的回款 (real payments collected),真正的现金流 (real cash flow) that can keep a company alive for the next decade.
This might be the best direction for the story to go.
Words from "Beyond the Layout":
Three years ago, everyone thought this time China would跑出 (run out/forge) its own OpenAI.
Three years later, the two who ran earliest灭了自己灯 (turned off their own lights) and换了赛道 (changed tracks).
This isn't a story of failure. It's a group of extremely intelligent people, in the most dignified manner, bidding farewell to a era's过高期待 (excessively high expectations).
As for whether a true Chinese OpenAI will appear—
That's a matter for another group of people.
This article is from the WeChat public account "Beyond the Layout," author: Layout君 (Layout Jun)
