Chinese artificial intelligence startups are overhauling their business models as they fight to stay competitive following the widespread adoption of technology from their national rival DeepSeek.
Zhipu was once considered China’s most prominent large-scale language model startup, but according to two people familiar with the issue, it focuses on building an enterprise sales business, and therefore has locked its hopes into an initial public offering to maintain cash-intensive growth.
Among other major generation AI startups in China, 01.AI has stopped large-scale language models “pre-training” and focused on selling tailored AI business solutions using DeepSeek’s model. Baichuan has chosen to focus on the healthcare market. Moonshot has significantly reduced the marketing budget for Kimi Chatbot, which focuses on Model Training.
People close to these companies said they all declined or did not respond to requests for comment, but the change suggests that Deepseek has dramatically changed the shape of China’s burgeoning AI industry.
Since the breakthrough R1 model was launched in late January, the Hangzhou-based startup has quickly been the country’s AI champion by Beijing, seeing the adoption of lightning bolts of deliberate technology from hospitals to local governments.
To replicate Deepseek’s success, we have left some of the country’s top AI startups, which have gained a great deal of support from domestic investors as part of the AI boom, to reevaluate existing strategies.
“China’s LLM market is rapidly consolidating around a small number of leaders,” said Wang Tiezhen, engineer at AI Research Hub Hugging Face. “DeepSeek has encouraged many companies to redirect resources to their applications rather than basic model development.”
Based in Beijing, 01.AI was founded by a venture capitalist and former head of Google China Kai-Fu Lee, where he pivoted his business in what he called “The Deepseek Age.”
The group, which launched a series of open source models called YI, stopped pretraining in late 2024, where developers used large datasets to train their models, as their rivals trained larger, more powerful models. In the contract with Alibaba, its basic model team transferred to the Internet giant, according to people familiar with the issue.
Last week, 01.AI announced that it would sell customized AI solutions to companies that want to deploy Deepseek models. 01. AI is selling so-called “combined expertise” as a competitive advantage. This is also used in the way DeepSeek uses to train the model.
Rather than training one “dense model” on a vast database that reduces data from the Internet and other sources, this approach combines many small models trained with industry-specific data. The MOE structure allows Chip poor companies to train larger models to reduce computing power, but it can be more difficult for third-party developers to deploy.
Deepseek decided to focus on research rather than trying to maximize revenue by selling applications to businesses, but left a gap embedded by intermediaries like 01.AI. Internet giant Baidu has pivoted to provide the same service in recent weeks.
Moonshot attracted attention last year about the virus’s AI chatbot, but its popularity has been halting frequently and suffered after rivals launched competitive products.
In recent weeks, two people familiar with the issue said the startup has focused on model training to reduce Kimi’s marketing spending, replicate the success of Deepseek’s breakouts, and improve chatbot performance.
But Kimi is overtaken by other apps, so Moonshot portrays an uncertain future as he burns out with cash that trains models without steady revenue. The startup aims to make money by inviting users to send virtual gifts to the AI character “Kimi” behind the chatbot.
According to people familiar with the transaction, it raised more than $1.3 billion in funding through two investment rounds last year, as it has mixed computing credits from Chinese tech giant Alibaba and cash from venture capital companies.
In early 2024, Alibaba considered Moonshot as a potential acquisition target, and as part of its $800 million investment, people said it secured the first right to buy the startup with future sales. In recent months, Alibaba has curbed start-up investments after founder Jack MA directed CEO Eddie Wu to instead focus on internal AI efforts. People added that this shift would make Alibaba unlikely to try and acquire you in the future.
Beijing-based startup Baihian has doubled its healthcare business after working on consumers facing AI chatbots and enterprise business pitches before doubled into education, finance and healthcare companies.
In February, Baichuan closed its business line by rejecting a sales team focused on selling customized financial AI applications to banks and investment funds, two people familiar with the issue said.
At the time, the company’s leadership announced to employees that it was focusing on developing technology for hospitals, including AI doctors to help with diagnosis.
Some content could not be loaded. Please check your internet connection or browser settings.
In contrast, Zhipu, founded by Tang Jie, a well-known computer scientist at Tsinghua University, still pursues multiple business lines. It launched an enterprise business selling several consumer applications and personalized AI applications to local governments and businesses, China’s notoriously competitive, low margin business.
Startups are burning through cash to build enterprise sales businesses. In 2024, Zhipu created RMB300MN ($410 million), an RMB2BN with loss, according to three investors who were explained to the figures.
Ballooning costs have sparked concern among some investors after Deepseek demonstrated its path to building cutting-edge models with a smaller budget. In contrast to Deepseek’s small workforce of around 160 employees, Zhipu employs around 800 people, making it the largest LLM startup by personnel.
According to two people familiar with the issue, Zhipu wants a cash boost after receiving one of Beijing’s coveted IPO recommendation letters. The company requires regulatory approval before pursuing a list of technology-centric Star Innovation Boards.
The startup received a nod from Beijing before Deepseek changed the landscape of the competition for Chinese AI players. Zhipu previously told investors it aims to list the list by the end of the year, saying that two people know the problem. However, they added that if Deepseek’s development moves forward in an IPO, it could have an impact.
The company’s investors have also expressed concern that the government’s Deepseek hug could threaten Zhipu’s business model of selling customized AI solutions to local governments, according to two people familiar with the issue.
However, Deepseek has decided to shake up the AI race in China, guide some rivals and challenge the group directly or adopt an open source model to focus on smaller potential markets.
“By adopting a first-class model, companies eliminate the need to invest tens of millions of dollars a year in training inferior internal replacements,” said Wang of Face.
