ARTIFICIAL intelligence (AI) ventures in China have typically been started by academics or experts working for foreign companies. But the furor this year around OpenAI’s ChatGPT has inspired a new breed of AI entrepreneur — the elite of China’s tech giants.

After spending years testing and developing AI at Chinese companies, heavyweights such as Jia Yangqing, a former vice-president at Alibaba, are branching out on their own.

Jia previously worked on computer vision and deep learning at Google Brain, the US tech giant’s artificial intelligence and machine-learning research unit. He also led a research team at Facebook, now Meta Platforms, building a large-scale AI platform for the company’s applications.

Jia left Alibaba in March to pursue “the next challenge” of his career, reportedly to join a new start-up with a focus on AI infrastructure.

Other high-profile departures include Wang Changhu, former head of vision technology at ByteDance, and Li Yan, who led Kuaishou Technology’s multimedia understanding unit in its short video division.

Many of the big-hitters at Chinese tech firms were in fact poached from universities or overseas companies, and since they have been extensively researching and finding end-uses for AI models, making the decision to try their hand at an AI startup is a no-brainer, according to one investor in Chinese hardware.

GET BT IN YOUR INBOX DAILY

Start and end each day with the latest news stories and analyses delivered straight to your inbox.

Sign UpVIEW ALL

The success of OpenAI was a great inspiration to them, and the flood of capital certainly helped give them the boost they needed to start their own ventures, the investor said.

The latest frenzy has also spurred on tech specialists and investors to go looking for good ideas to back.

Retired Meituan co-founder Wang Huiwen announced a personal investment of US$50 million to set up Beijing Lightyear Technology and went on a talent hunt to “build China’s OpenAI”. Soon after, his firm acquired OneFlow Technology, a developer of open-source AI frameworks for machine learning, through a stock swap. The two companies merged for continued financing at a valuation of US$1 billion.

Wang’s venture inspired Lee Kai-Fu, the chairman and chief executive officer of Sinovation Ventures, to launch his own startup.

Lee announced in March that he was setting up Project Al 2.0, a company focused on developing platform-based AI and related applications. The AI veteran and former Google China chief is also looking for investment opportunities in the field.

Many of the new AI projects in China are being worked on behind closed doors, making it difficult to track the exact number. But some industry insiders expect more than 50 startups focused on building large-scale models could be launched within three months.

But creating these models is a costly business, and a challenge few can manage.

Yin Qi, co-founder and CEO of AI software developer Megvii Technology, said for a company to build a GPT large language model, it would need at least 10,000 Nvidia A100 graphics processing chips — which are considered to be one of the best chips for powering machine learning tasks — and to invest around two billion yuan (S$383.8 million) in hardware.

This prices many AI companies out of the race, which, like most tech startups, typically run at a loss for years. But if they don’t jump on the ChatGPT bandwagon, they fear they may be vulnerable to new entrants, particularly if they have recently gone public and have shareholders to please.

SenseTime decided to take on the challenge. The AI software developer unveiled in April a suite of new services including its large AI model SenseNova and a chatbot called SenseChat, which has around 180 billion parameters.

Parameters are key to machine learning algorithms and used to measure how well an AI model is performing. The company hopes SenseNova could become a “supermarket of AI big models” for clients and partners.

SenseTime is one the Chinese AI sector’s more well-established developers yet, financially, it pales in comparison to its big tech rivals. In 2022, the firm booked 3.8 billion yuan in revenue and a loss of more than 6 billion yuan — whereas Baidu turned over revenue of 123.7 billion yuan and 20.7 billion yuan in net income.

Baidu unveiled its Kunlun AI chips in 2018, and has already deployed them on a large scale, while SenseTime only unveiled its first hardware offering, its AI-ISP chip, last year.

Moreover, SenseTime is operating under US sanctions that inhibit its access to capital as well as crucial American components. Still the company is unfazed, with its co-founder Yang Fan telling Caixin that what matters in creating large models is experience in hardware engineering, algorithms and big data.

Overall, China’s AI investment is expected to reach US$26.7 billion in 2026, accounting for about 8.9 per cent of global investment, according to IDC in an October 2022 report. It predicts that hardware will be the largest primary market in China’s AI market over the next five years and investment will reach more than US$15 billion in 2026.

Where are all of these models going to being used? Zhou Ming, who was at Microsoft Asia Research for 21 years, before resigning in 2020 to start his own AI venture, sees a lot of potential for application in finance.

Given that startups typically have limited resources, Zhou believes they should build their models based on demand. He picked the finance industry, where there’s plenty of usable data, potential clients and money, but not a big supply of companies that can provide ChatGPT-like, multi-round dialogue capabilities.

His startup, Beijing Langboat Technology, provides functions including public sentiment analysis, research report analysis, and professional financial translations based on its self-developed Mengzi large language model.

Other firms seeking a lower entry threshold are targeting the consumer sector. Zhou Zhifeng, a partner at Chinese venture capital firm Qiming Venture Partners, said that while the first wave of AI frenzy lacked supporting infrastructure, large model APIs have emerged.

API, which stands for application programming interface, is a software intermediary that allows two applications to “talk” to each other.

Startups don’t need to recruit a large number of software engineers to develop the underlying models. With small-scale financing, they can form a relatively small-scale team to build applications based on existing APIs, Zhou said.

Chinese AI image generation startup ZMO.AI is doing just this, operating with a team of about 20 people. Founder and CEO Zhang Shiying told Caixin that her firm launched a marketing content generation product in September 2022 and racked up more than one million monthly active users in just about five months. It launched a paid product in February and is already making a profit.

Startups don’t need to build models and should instead focus on developing applications, where “brand new entrepreneurial opportunities that are 10 times greater than WeChat and Douyin will emerge”, said Baidu co-founder and CEO Robin Li.

In other words, leave the large models to the likes of Baidu. CAIXIN GLOBAL