Among China’s other leading generative AI startups, Zero One Wish, founded by Kai-Fu Lee, has stopped “pre-training” large language models and instead focuses on selling customized AI business solutions using DeepSeek’s models; Baichuan Intelligence has chosen to focus on the healthcare market; Dark Side of the Moon has significantly reduced the marketing budget of its Kimi chatbot to focus on model training.
People close to the companies said the shifts show how DeepSeek has dramatically changed the landscape of China's emerging artificial intelligence industry. The companies declined to comment or did not respond to requests for comment.
Since launching its groundbreaking R1 model in late January, the Hangzhou-based startup has quickly been seen as China's AI leader, with its technology quickly being adopted from hospitals to local governments.
This has forced some of China's top artificial intelligence startups to re-evaluate their existing strategies in an attempt to replicate DeepSeek's success. During the artificial intelligence boom of the past two years, these start-ups have received strong support from domestic investors.
Wang Tiezhen, an engineer at HuggingFace, an artificial intelligence research center, said: "China's large language model market is rapidly concentrating on a few leading companies. DeepSeek has prompted many companies to redirect resources to application areas rather than basic model development."
Beijing-based Lingyiwu has adjusted its business direction in what Kai-fu Lee calls the "DeepSeek era." The company had launched a series of open source models called "Yi". As the models trained by competitors became larger and larger, more powerful, and costs continued to rise, Zero One Wagon stopped pre-training at the end of 2024. In a deal with Alibaba, its basic model team was transferred to the internet giant, according to people familiar with the matter.
Last week, Lingyiwu announced that it will sell customized artificial intelligence solutions to enterprises looking to deploy DeepSeek models. Lingyiwang promotes its expertise in so-called "hybrid expert models" as a competitive advantage, and DeepSeek also adopts this approach to train models.
DeepSeek's decision to focus on research rather than maximizing revenue by selling applications to enterprises has left room for intermediaries like Zero One Thing. Internet giant Baidu has also shifted to offering the same service in recent weeks.
Dark Side of the Moon attracted attention last year for its popular artificial intelligence chatbot Kimi, but Kimi's popularity suffered as frequent service outages and competitors launched competing products.
In recent weeks, the startup has cut back on Kimi’s marketing spending to focus more on model training to replicate DeepSeek’s breakout success and improve the performance of its chatbots, according to two people familiar with the matter.
But as Kimi was surpassed by other applications, Dark Side of the Moon invested money in model training without stable income, and the future was full of uncertainty. The startup seeks to make money by inviting users to send virtual gifts to “Kimi,” the artificial intelligence character behind the chatbot.
Last year, Dark Side of the Moon raised more than $1.3 billion in two rounds of financing, including computing resources from Chinese tech giant Alibaba and cash from venture capital firms, according to people familiar with the matter.
In early 2024, Alibaba viewed Dark Side of the Moon as a potential acquisition target, and with its $800 million investment, Alibaba received the right of first refusal in any future sale, these people said. Alibaba has tightened its investment in startups in recent months after founder Jack Ma instructed CEO Wu Yongming to focus on in-house artificial intelligence operations. The shift makes Alibaba less likely to acquire Kimi in the future, these people added.
Beijing-based startup Baichuan Intelligence, which previously worked on consumer-facing AI chatbots and marketed enterprise business to education, finance and healthcare companies, has now made healthcare a focus.
According to two people familiar with the matter, in February this year, Baichuan Intelligence disbanded its sales team focused on selling customized financial artificial intelligence applications to banks and investment funds and terminated this business line.
At the time, company management announced to employees that it would focus on developing technology for hospitals, including artificial intelligence doctors to assist with diagnosis.
In contrast, Zhipu, founded by Tang Jie, a well-known computer scientist at Tsinghua University, is still advancing multiple business lines. It has launched several consumer-facing apps, as well as an enterprise business selling personalized AI apps to local governments and businesses.
The startup has been burning through cash as it expands its enterprise sales business. According to three investors with knowledge of relevant data, in 2024, Zhipu will achieve sales of 300 million yuan and a loss of 2 billion yuan.
The ballooning costs have some investors worried after DeepSeek demonstrated a way to build cutting-edge models on a lower budget. Compared with DeepSeek's small team of about 160 people, Zhipu employs about 800 people and is the largest large language model startup by number of employees.
Zhipu had previously told investors it aimed to go public by the end of this year, according to two people familiar with the matter. But they added that DeepSeek's development could have an impact on it if it proceeds with an IPO.
The Financial Times stated that DeepSeek has shaken China's artificial intelligence competition landscape, causing some competitors to have to decide whether to challenge the company directly or adopt its open source model and focus on smaller potential markets.