Chips or not, Chinese AI pushes ahead

A host of Chinese AI startups are attempting to write more efficient code for large language models to cope with the limited number of training cycles that come with using less sophisticated semiconductors. Others are building smaller, specialized models or employing training methods that require less energy and time.

The net effect, Chinese AI executives and analysts say, is that while semiconductors widely available in China generally lack the computing power of those produced by global leaders such as Nvidia, the drive to monetize can make some of the differences moot.

01.AI, a unicorn backed by Alibaba and Xiaomi, employs a lower-precision training format that reduces the energy and time needed to train machine-learning models. The format, used in the U.S. by companies including Google, can accelerate a model’s output, according to Nvidia researchers.

In China, “we don’t have a lot of [graphics processing units], and that forces us to develop very efficient AI infrastructure and inference engines," 01.AI founder Kai-Fu Lee said, citing a lack of funds as a reason for low chip supplies.

The company has said its chip-cluster failure rate, a measure of how often groups of connected chips fail to work together, is lower than the industry average.

The goal of efficiency has spread among Chinese business leaders seeking to build out their AI footprints even as the U.S. restricts their access to the most advanced chips.

Robin Li, chief executive of internet and autonomous-driving company Baidu, has been one of the most vocal industry leaders in China warning of a waste of computing resources by a glut of companies each developing their own foundation models.

“Without applications," Li said at an AI industry conference last month, “having only foundational models, whether open source or closed source, is worthless."

In the quest for consumer buy-in, some Chinese companies seek to develop specialized applications rather than focus on creating the biggest and best models, analysts say. A recent KPMG report said AI investors in China in the second quarter “focused on AI-enablement rather than on LLM offerings," including in areas like robotics and improving workplace efficiencies.

Still, “the big question now is which of these companies will be able to produce results and successfully commercialize their offerings," Zoe Shi, a partner at KPMG China, said in the report.

Early efforts at turning profits off AI have been largely unclear in presentations of quarterly results in recent weeks, but some pioneers in applications are emerging. TikTok’s Chinese parent ByteDance, dubbed China’s “App Factory," has launched more than 20 apps, including a chatbot, an English tutor and a video creator powered by its in-house models. Companies including ByteDance and its local rival Kuaishou Technology have made publicly available equivalents of OpenAI’s Sora that generate video from text using AI.

Overseas, three of the top 10 most downloaded AI apps in the U.S. this year were developed by Chinese companies, according to market researcher Sensor Tower. 01.AI has tested two AI chatbots, Monaland and Shado, outside China through a Singapore-based company, people familiar with the projects said.

Some industry experts expect that smaller-size models that can power AI features on smartphones and laptops, or “edge AI models," will be the next game-changer.

“This year is about small models," said Winston Ma, an adjunct professor at New York University School of Law and an adviser at Dragon Global, an AI-focused family office. Smaller models using less training data are faster, benefiting real-time applications with specific functions, he said.

AI unicorn Baichuan is working with Qualcomm to integrate a smaller LLM in its AI PC in China, people familiar with the matter said. Samsung has used models from Baidu and ByteDance in its smartphones in China.

“Focusing on edge doesn’t mean lowering the tech requirement, just shifting to an area where higher computing power isn’t the most critical requirement," said Boris Van, an analyst at Bernstein Research.

As for computing resources, Beijing’s push for self-sufficiency on several fronts is spurring efforts to close the gap with Western chip designers.

“It is important to figure out how we can achieve better results through engineering instead of blindly investing in computing power," 01.AI’s Lee said at the AI industry conference in China.

Huawei Technologies’ Ascend chips have been used by Chinese tech giants, state research labs and national AI data centers. A Huawei executive said in June that nearly half of China’s LLMs were trained using the company’s chips.

Baidu, Alibaba and Tencent have been using in-house chips to operate their AI models and improving engineering and algorithms to offset the shortage of advanced computing power. Companies are also researching how to combine different types of chips to avoid relying on any one kind of hardware.

“We shouldn’t think that not having the most advanced AI chips means we won’t be able to lead in AI," Zhang Ping’an, a Huawei senior executive in charge of its cloud-computing business, said at the July AI conference. “We should abandon this viewpoint in China."

Write to Kimberley Kao at kimberley.kao@wsj.com and Raffaele Huang at raffaele.huang@wsj.com