Abandoning large model training? Fully acquired by Alibaba? In the first week of 2025, 01.AI, one of China’s most-watched artificial intelligence startups, sparked a storm of speculation and debate.
On January 7, 36Kr sat down with Kai-Fu Lee, the founder of 01.AI and renowned AI entrepreneur, for an exclusive interview. Lee addressed the company’s recent staffing changes and restructuring efforts.
“What works for a startup in its first year may not work in its second,” Lee told 36Kr. “Adjustments and transitions are a natural part of entrepreneurship. 2025 marks a decisive year for commercialization, and 01.AI must realign its priorities accordingly.”
Just a day earlier, an online rumor claimed that 01.AI had sold its pre-training and infrastructure teams to Alibaba. That evening, Lee personally denied the allegations on social media.employees dedicated to training super-large models had joined the company’s large industrial model laboratory, a joint initiative with Alibaba Cloud.
According to Lee, the restructuring began in mid-December 2024 when both teams were informed of their reassignment. By late December, members of the pre-training team had received offers from Alibaba’s Tongyi division, while the infrastructure team was absorbed into Alibaba Cloud.
The background of this restructuring lies in 01.AI’s realization: pre-training massive models is no longer cost-effective for startups.
“It’s clear that only tech giants can [bear the costs of training super-large models],” Lee said. Since 2024, 01.AI has shifted its focus to developing small-parameter, industry-specific models. He emphasized that super-large models hold significant value as “teacher models,” transferring their capabilities to smaller ones. This makes the partnership with a tech giant capable of handling large-scale model training essential.
For 01.AI, that partner turned out to be its major investor, Alibaba. “Many team members still harbor dreams of artificial general intelligence (AGI), and those individuals are welcome to join the laboratory,” Lee said.
As for rumors about selling GPUs, Lee clarified that 01.AI relies on cloud services to train its models: “We don’t own GPUs, so it’s impossible for us to sell them.”
The most pressing challenge for 01.AI in 2025 is profitability. Alongside revising its model training strategy, the company plans to spin off its gaming, finance, and other AI applications into independent units. This move aims to streamline operations and secure funding tailored to specific verticals.
This decision stems from Lee’s experience with project incubation at Sinovation Ventures. According to him, teams focusing on specific verticals tend to achieve greater depth and impact.
Practical considerations also underpin these spinoffs. “First, we talk to investors to gauge interest. If no one is willing to invest, there’s no point in spinning off the business because it won’t be self-sustainable,” Lee said.
“Startups must carefully allocate every dollar,” he added. “Spending on more GPUs to train large models isn’t the way forward for survival.”
The following interview has been edited and consolidated for brevity and clarity.
36Kr: What happened with 01.AI’s pre-training and infrastructure teams?
Kai-Fu Lee (KL): We’ve been collaborating closely with our major investor, Alibaba, to establish a joint laboratory dedicated to large industrial models. This lab, led by Alibaba, focuses on scaling law research.
Some of our team members who specialize in and are deeply committed to this area have joined the lab and integrated closely with Alibaba’s efforts.
36Kr: Does this mean that members of the pre-training and infrastructure teams were given a choice between joining Alibaba or staying with 01.AI?
KL: I’d rather not delve into specific details here.
What I can say is that there’s immense potential in training super-large models right now, and Alibaba has decided to pursue this space aggressively. We commend its dedication. Over the past year and a half, we’ve had exceptional team members passionate about large-scale model training, so it’s a natural fit for them to align with Alibaba’s initiatives.
36Kr: Does this mean 01.AI has officially moved away from pre-training?
KL: Our approach to pre-training is pragmatic. We focus on developing smaller, faster models that deliver high commercial value.
Previously, OpenAI co-founder Ilya Sutskever noted that scaling laws may have reached their limits. Training ever-larger models is becoming prohibitively expensive, with diminishing returns. It’s clear that only tech giants can afford to pursue this path.
Our Yi-Lightning model, which features billions of parameters, received strong market feedback in 2024. Looking ahead, 01.AI will prioritize faster, smaller, and more cost-effective models. Any future training efforts won’t exceed the scale of Yi-Lightning.
36Kr: With the new lab formed by 01.AI and Alibaba Cloud to train super-large models, what roles do the two parties play?
KL: The value of super-large models lies in their ability to function as “teacher models,” imparting their capabilities to smaller ones through techniques like data distillation and data synthesis. This is a key strategy for training models.
Although large models are expensive and slow to train, they are strategically important in China’s tech ecosystem, particularly given global tech restrictions. Maintaining innovation in this area is critical for securing a strong technological foundation.
We concluded that partnering with a tech giant capable of supporting large-scale model training is the optimal path forward. Alibaba takes on the training of these massive models, while our nimble and efficient teams concentrate on developing smaller, cost-effective models that drive practical applications and market growth.
For those still passionate about building super-large models, the joint lab provides an ideal platform to pursue that vision.
36Kr: There are rumors that 01.AI sold its GPUs to Alibaba. Could you clarify?
KL: I’m not sure where that rumor started. We don’t own GPUs, so it’s impossible for us to sell them. I can’t comment further on something baseless.
36Kr: Is it common for AI startups to rely on cloud services rather than own GPU clusters?
KL: It’s widely understood that small companies are not well-suited for training super-large models. Even organizations with tens of thousands of GPUs should evaluate whether maintaining those resources aligns with their strategic goals if they’re not actively pursuing large-scale model training.
This is a common industry consideration. Our decision to step away from this path doesn’t stem from doubt about scaling laws. Instead, it reflects a pragmatic focus on partnering with a tech giant like Alibaba to manage resource-intensive aspects. It’s the most sustainable approach for survival and growth.
36Kr: In your social media post addressing rumors, you mentioned that 01.AI confirmed revenue exceeding RMB 100 million (USD 14 million) in 2024. What were the primary revenue sources?
KL: Initially, we concentrated on consumer-facing products, particularly for international markets. Early on, we recognized that commercializing B2C products in China in 2024 would be challenging, so we refrained from spending on customer acquisition for domestic B2C offerings.
For our consumer business, about 20–30% of last year’s revenue came from paid overseas productivity tools like PopAi.
36Kr: Does this mean 70% of revenue came from enterprise clients?
KL: That’s correct. In the second half of 2024, we shifted our focus to enterprise clients. After refining our large model technology, we turned to identifying specific applications and use cases.
We explored a range of industries, seeing significant growth in gaming and achieving notable results in finance and energy sectors.
36Kr: Transitioning from B2C in the first half to B2B in the second half seems like a major shift in business focus.
KL: While external observers may view the changes beginning in early January 2025, we had actually been planning this transition for months. Systematic adjustments to our organizational structure, resource allocation, and project priorities had been underway, with milestones becoming apparent by year-end.
For example, discussions with Alibaba regarding the joint lab were ongoing for some time before the recent announcement.
36Kr: How has 01.AI’s organization and resource allocation evolved over the past year?
KL: We’ve restructured to strengthen our B2B capabilities and build industry partnerships for specialized large models. This involved bolstering teams in key areas such as presales, customer demand analysis, product development, and commercial strategy.
Additionally, we reinforced our engineering and development teams to focus on standardization and platformization, steering clear of one-off projects that often result in losses.
36Kr: With 01.AI focusing on training smaller models, what is your commercialization strategy?
KL: We see fast, high-value models as particularly well-suited for industry-specific applications. This doesn’t mean creating extremely small models with just three or four billion parameters, as those are often inadequate for many use cases.
Instead, our goal is to develop models comparable in scale to Yi-Lightning or slightly smaller, ensuring they remain versatile and applicable across various industries.
36Kr: 01.AI entered the enterprise market relatively late. How do you plan to overcome this latecomer disadvantage?
KL: We’re focusing on a few core verticals. Unlike some AI companies that can hire 300 salespeople, we’re leveraging my personal network to break into specific high-value domains and find industries where we can deliver standard solutions.
We’re also collaborating with Sinovation Ventures, which has incubated many complementary companies. These partnerships are highly synergistic and allow us to work together effectively.
Moreover, we believe industry-specific models need further segmentation. For instance, we don’t view finance as a singular sector—it can be divided into numerous subfields. I can’t share further details here, as revealing our plans could give competitors an edge.
36Kr: During an internal meeting earlier this month, you emphasized that 01.AI must fully align with applications. Can you elaborate?
KL: Initially, our focus was on pursuing AGI, which allowed us to attract highly skilled technical talent and develop world-class technology.
But to some extent, 2025 will be a decisive year for the commercial success of large model companies. If we fail to demonstrate that our technology delivers real-world applications, we risk being left behind.
RELATED ARTICLE
36Kr: What are the benchmarks for proving real-world applications? User numbers? Revenue? Profit?
KL: The core criterion is whether an application generates revenue. It’s not about blindly pursuing user growth or increasing traffic—it’s about sustainability and income generation.
When I say “align with applications,” I mean that without a profitable application, you’re essentially running a tech lab.
36Kr: Reports suggest that some of 01.AI’s business lines have been spun off into separate companies for independent operations and external funding. What drove this decision?
KL: It aligns with the incubation model I developed at Sinovation Ventures. For instance, our AI gaming subsidiary was recently spun off. We had been operating this business for several months and concluded that it had the potential to stand on its own.
36Kr: Why spin off certain businesses?
KL: When a team is deeply focused on a single vertical with strong growth potential, it’s better positioned to achieve excellence and make a significant impact.
Compared to being part of a broader team working across multiple industries, a standalone entity can devote its full attention to its domain and access resources tailored to its specific needs.
36Kr: How do you evaluate whether a business should be spun off?
KL: First, we talk to investors to gauge interest. If no one is willing to invest, there’s no point in spinning off the business because it won’t be self-sustainable.
36Kr: It’s rare for a startup to pursue so many business directions simultaneously. Doesn’t this dilute your focus?
KL: Our goal is to create a platform for enterprise clients. Centralized R&D costs are shared across business lines—maybe 20% goes to one area, 30% to another, and 50% toward general-purpose tools and infrastructure.
Sometimes, business success is driven by market trends or emerging concepts. It’s hard to plan everything in advance—you have to adapt as you go. Ideally, we aim to develop a few standout product lines in 2025.
36Kr: After the past year and a half, what new insights do you have about AI entrepreneurship?
KL: Be bold in making proactive adjustments and recognize the right strategic path. What works for a startup in its first year may not work in its second. Clinging to unsustainable practices can be harmful to the company.
We hope to build a more certain future for 01.AI, giving us greater confidence in achieving our goals.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Zhou Xinyu for 36Kr.