Chinese artificial intelligence start-ups are overhauling their business models as they fight to remain competitive following the widespread adoption of rival DeepSeek’s technology across the country.
Zhipu, once considered China’s most prominent large language model start-up, has pinned its hopes on an initial public offering to sustain its cash-intensive growth as it focuses on building up its enterprise sales business, according to two people familiar with the matter.
Among China’s other leading generative AI start-ups, 01.ai has stopped “pre-training” large language models to focus on selling tailored AI business solutions using DeepSeek’s models; Baichuan has opted to concentrate on the healthcare market; and Moonshot has slashed its marketing budget for its Kimi chatbot to focus on model training.
People close to these companies, which all declined or did not respond to requests for comment, said the shifts show how DeepSeek has drastically altered the shape of China’s burgeoning AI industry.
Since the launch of its breakthrough R1 model in late January, the Hangzhou-based start-up was quickly crowned the country’s AI champion by Beijing and has seen lightning adoption of its technology everywhere from hospitals to local governments.
It has left some of the country’s top AI start-ups — which over the past two years have gained significant backing from domestic investors as part of the AI boom — to re-evaluate their existing strategies in an effort to replicate DeepSeek’s success.
“The Chinese LLM market is rapidly consolidating around a handful of leaders,” said Wang Tiezhen, an engineer at AI research hub Hugging Face. “DeepSeek has prompted many companies to redirect resources to applications rather than foundational model development.”
Beijing-based 01.ai, founded by venture capitalist and former head of Google China Kai-Fu Lee, has pivoted its business in what he has called “the DeepSeek age”.
The group, which has launched a series of open-source models called Yi, stopped pre-training — in which developers use massive data sets to train models — in late 2024 because of rising costs as its rivals trained ever larger and more powerful models. In a deal with Alibaba, its foundational model team was transferred to the internet giant, according to people familiar with the matter.
Last week, 01.ai announced it would sell tailored AI solutions to companies wanting to deploy DeepSeek’s models. 01.ai is pitching its expertise in the so-called “mixture of experts” as its competitive advantage, the method also used by DeepSeek to train its models.
Rather than training one “dense model” on a vast database that has scraped data from the internet and other sources, the approach combines many smaller models trained on industry-specific data. The MOE structure allows chip-poor companies to train larger models on less computing power, but can be more challenging for third-party developers to deploy.
DeepSeek, which has decided to focus on research rather than seeking to maximise revenues by selling applications to companies, has left a gap to be filled by intermediaries like 01.ai. Internet giant Baidu has also pivoted to offer the same service in recent weeks.
Moonshot courted attention last year for its viral AI chatbot Kimi but its popularity has suffered following frequent outages and rivals launching competitive products.
In recent weeks, the start-up has cut marketing spending for Kimi as it increases its focus on model training to replicate the breakout success of DeepSeek and improve its chatbot’s performance, according to two people familiar with the matter.
But as Kimi is overtaken by other apps, Moonshot is charting an uncertain future as it burns through cash doing model training without stable revenue. The start-up has sought to make money by inviting users to send virtual gifts to “Kimi”, the AI character behind the chatbot.
It raised more than $1.3bn in financing through two investment rounds last year, with a mixture of computing credits from Chinese tech giant Alibaba and cash from venture capital firms, according to people familiar with the deals.
In early 2024, Alibaba considered Moonshot a potential acquisition target and secured the first right to buy the start-up in any future sale as part of their $800mn investment, the people said. In recent months, Alibaba has reined in start-up investments, after founder Jack Ma directed chief executive Eddie Wu to focus instead on internal AI efforts. The shift makes it less likely that Alibaba will seek to acquire Kimi in the future, the people added.
Beijing-based start-up Baichuan has doubled down on its healthcare business after previously working on consumer facing AI chatbots and enterprise business pitches to educational, financial and healthcare companies.
In February, Baichuan dismissed its sales team focused on selling its tailored financial AI application to banks and investment funds and ended the business line, said two people familiar with the matter.
At the time, the company leadership announced to employees that it was focusing on developing its technology for hospitals, which includes an AI doctor that assists with diagnosis.
In contrast, Zhipu, founded by Tang Jie, a prominent computer scientist from Tsinghua University, is still pursuing multiple business lines. It has launched several consumer applications as well as an enterprise business selling personalised AI applications to local governments and companies, a notoriously competitive and low-margin business in China.
The start-up has been burning through cash as it builds its enterprise sales business. In 2024, Zhipu made Rmb300mn ($41mn) in sales and Rmb2bn in losses, according to three investors briefed on the figures.
The ballooning costs have prompted concern among some investors after DeepSeek demonstrated a pathway to building cutting-edge models on a smaller budget. In contrast to DeepSeek’s small workforce of about 160 employees, Zhipu employs about 800 people, making it the largest LLM start-up by headcount.
Zhipu is hoping for a cash boost after receiving one of Beijing’s coveted recommendation letters for an IPO, according to two people familiar with the matter. The company needs approval from the regulators before it can pursue a listing on the tech-focused Star Innovation Board.
The start-up received Beijing’s nod before DeepSeek altered the competitive landscape of AI players in China. Zhipu previously told investors it was aiming to list before the end of the year, said two people with knowledge of the matter. But they added that the DeepSeek developments could affect it if it pushes ahead with an IPO.
Investors in the company have also expressed concern that the government’s embrace of DeepSeek could threaten Zhipu’s business model of selling tailored AI solutions to local governments, according to two people familiar with the matter.
But DeepSeek has shaken up the AI race in China, leading some rivals to decide whether to challenge the group directly or adopt its open-source models to focus on a smaller potential market.
“By adopting top-tier models, companies can eliminate the need to invest tens of millions of dollars annually in training inferior in-house alternatives,” said Hugging Face’s Wang.
Read the full article here
