Chinese companies are being forced to get creative to overcome a U.S. tech embargo that's choking the country out of one of the most valuable resources of the day: semiconductors.
A package of measures issued by the Biden administration bans U.S. companies from exporting advanced chips - and equipment used to produce them - to China.
The move is aimed at hampering Chinese manufacture and development of advanced semiconductors, key in training artificial intelligence models like ChatGPT and GPT-4.
By the time the sanctions were issued in October, Chinese institutions were producing 4.5 times more research papers on the topic of AI than American ones, according to a report released by State of AI.
The release of ChatGPT in November 2021 by OpenAI fired the starting pistol of an all-out race to dominate AI development, and due to U.S. sanctions, Chinese companies are lagging behind.
Microsoft (NASDAQ: MSFT) was among the first major U.S. big tech companies to include a version of ChatGPT in one of its consumer products when it added natural language chat capabilities to its Bing search engine.
Nvidia's (NASDAQ: NVDA) A100 chip is quickly becoming the industry standard for tech companies looking to train large AI models to perform tasks in the neighborhood of what ChatGPT can offer. The chip is high above those used in regular consumer electronics and costs around $10,000 each. The newer, improved version, called H100, is sold for over $30,000.
According to analysts interviewed by the Wall Street Journal, training a large language model like ChatGPT required between 5,000 and 10,000 of these chips, which are banned from China by U.S. measures.
Meta (NASDAQ: META) owns over 21,000 A100s, according to State of AI, which tracks ownership of large supercomputers. Leonardo, a supercomputer developed by the European Union and hosted in a research center in Italy, considered the fourth-fastest computer in the world, uses 13,000 of these chips.
While Nvidia has developed chips that are not bound by U.S. restrictions to provide specifically to the Chinese market, the less powerful A800 is about 30% slower than the A100, according to reporting by TechCrunch.
This is pushing Chinese companies like Huawei, Baidu (NASDAQ: BIDU) and Alibaba Group (NYSE: BABA) to develop new techniques that allow them to compete in the global landscape without relying on state-of-the-art semiconductors.
While the Chinese government has announced large investments into the development of a domestic chip manufacturing industry, experts put China years behind Taiwan and Korea in its chip-making capabilities.
According to the WSJ, Chinese chips, as well as Nvidia's slower versions, are not enough to train large AI models in order to successfully compete with growing development coming from the west.