Four Chinese generative artificial intelligence start-ups have been valued at between $1.2bn and $2.5bn in the past three months, leading a pack of more than 260 companies vying to emulate the success of US rivals such as OpenAI and Anthropic.
The newly minted unicorns — Zhipu AI, Moonshot AI, MiniMax and 01.ai — have gained significant backing from a largely domestic pool of investors and are fighting to hire the best talent to develop the most popular AI products.
“There is no winner of foundation models yet in the China market. These are some of the names leading the charge to claim that title,” said Charlie Dai, vice-president and principal analyst at tech-focused consultancy Forrester.
US counterparts outrank Chinese generative AI start-ups in terms of technological development and total fundraising. But with ChatGPT and other breakout AI applications like Character.ai unavailable in China, 262 start-ups are competing to bring out homegrown alternatives, according to a tally by data provider IT Juzi.
Overall, China’s generative AI start-ups have disclosed a fundraising total of Rmb14.3bn ($2bn) in the first four months of the year, according to IT Juzi, defying a chill in investment running through other areas of consumer technology. However, many start-ups have not disclosed all their fundraising rounds, shielding their war chests from competitors in the rush to hire talent and compete for limited computing resources.
Beijing has approved more than 40 large language models and related AI applications for public use, while building a supportive regulatory environment that encourages growth in the sector through tax breaks and subsidies.
Zhipu has become China’s biggest AI start-up by number of employees. Spun out of Beijing’s prestigious Tsinghua University, an important breeding ground for AI talent, the over 800-strong company is worth Rmb18bn ($2.5bn) based on its latest fundraising round in March, according to two of its investors. They did not disclose how much they raised.
Moonshot, set up by Yang Zhilin, a former student of a Zhipu founder, was valued at $2.5bn in a $1bn investment round announced in February. Previously, Yang did an internship at Google Brain AI and Meta AI and founded a start-up called Recurrent AI that analysed calls by salespeople.
Moonshot, Zhipu and 01.ai have developed chatbots that target office workers and students, who use the digital assistants to digest long texts and optimise search results. Neither Moonshot nor Zhipu responded to requests for comment on their funding.
Moonshot’s chatbot Kimi, which gets its moniker from Yang’s English name, has emerged as the closest rival to Chinese internet search giant Baidu’s Ernie Bot. Kimi had 12.6mn visits in March compared to its more established competitor’s 14.9mn visits, according to data provider Aicpb.com, but Kimi is growing much faster.
“Kimi has done a good job on their user interface and contextualising the content through RAG,” said one industry insider, referring to the retrieval-augmented generation technique that enables the models to fetch data from external sources and provide up-to-date answers to users’ queries.
But Kimi has become a victim of its popularity, struggling to cope with demand as users flock to the easy-to-use chatbot that industry experts have praised for its summary tool and for giving clear answers that are highly contextualised. Kimi Bot suffered an outage for two days in March, prompting the young company to issue an apology.
Faced with the same constrained computing resources, many AI start-ups have chosen to launch avatar chatbots, which do not need to be as good at reasoning as productivity chatbots. These chatbots are trained on smaller amounts of data, thus requiring fewer computing resources.
“ChatGPT is hard to copy. The model is the product. It is easier to build an avatar chatbot; you can do it on an open-source model with limited computing resources,” said one AI researcher in China.
Zhipu and MiniMax both have avatar chatbots, targeting the world’s largest gaming market with anime-themed characters to banter and flirt with while generating user feedback to iterate their models. Shanghai-based MiniMax is valued at $2.5bn, based on a $600mn funding round announced in March.
01.ai, founded by AI pioneer Kai-Fu Lee, has launched a series of open-source models called Yi, tailored for the Chinese market and built on Meta’s free-to-adapt Llama architecture. Hugging Face, which tracks open-source models, has ranked several iterations of the Yi models high on its scoreboards for common sense reasoning, maths, coding and reading ability. 01.ai has also launched a productivity chatbot called Wanzhi.
The AI start-up recently carried out a fundraising round at a $1.2bn valuation, according to a person with direct knowledge of the deal, and has the backing of Lee’s own venture capital fund Sinovation Ventures, Shunwei Capital, Xiaomi and Alibaba Cloud.
With no clear leader or “killer app” emerging in China, it is difficult for customers to select the right provider, said Jeffrey Ding, an expert in China’s AI ecosystem and assistant professor at George Washington University. “It’s very hard to differentiate between all these similar applications. How do you know which company is going to be more effective for your particular situation,” he said.
All four leaders have attracted funding from Alibaba, which has emerged as a key backer of AI start-ups as it looks to replicate the success of Microsoft’s big bet on OpenAI.
But the pool of investors funding this crop of start-ups is smaller than previously, with global tech funds not playing the role they did with the last generation of AI surveillance start-ups, including SenseTime and Megvii.
“Foreign investors are largely sitting this out,” one AI investor noted.
In addition to adequate funding, Chinese AI start-ups are finding they have both the engineering talent required to launch competitive products and sufficient computing resources to train existing models — despite US restrictions on advanced chip exports.
“Chinese cloud computing companies have sufficient stocks of Nvidia GPUs purchased before the ban for start-ups to do this round of model training,” said one employee at an AI start-up who negotiates the terms of cloud computing contracts.
However, two other employees at AI start-ups said they were targeting products that required less compute power, as the industry adapts to the US restrictions. They added that limited computing resources were a key reason why many Chinese AI start-ups were relying on open source models like Meta’s Llama to build their own models and applications, instead of having to use precious computing resources and talent to build a proprietary one from scratch.
“Chinese AI start-ups are doing a lot of research and development on top of the open source architecture, then fine-tuning the model and contributing to the overall ecosystem,” said Forrester’s Dai. “These start-ups don’t need to do model design and validate the architecture, which is very resource intensive,” he added.
Low salaries for AI engineers, compared with the US and Europe, are also helping to keep costs down. A recent computer science PhD graduate from a Chinese university can typically expect to command between $80,000 and $240,000 year at a large start-up, a range that is around four times higher in Silicon Valley, according to several people with knowledge of the matter.
“We’re growing so quickly,” said an employee at one such start-up, noting that the rapidly expanding product and sales team allowed the company to ramp up commercialisation of its technology and court new customers.
As one deep-tech investor in China said: “Chinese companies are not good at fundamental tech itself. But they are very good at looking at industry trends, following whatever good comes out, and leveraging their engineering dividend to follow whatever has emerged in the US.”
Additional reporting by Andy Lin in Hong Kong
Read the full article here