China’s IT behemoths had been stockpiling Nvidia’s high-performance GPUs even before Washington stopped the company from exporting them, fearing an intensifying tech war between the two countries.
During an earnings call this week, Robin Li, the CEO of Baidu, one of the major companies creating China’s OpenAI rivals, stated that the company has acquired enough AI processors to continue training its ChatGPT comparable Ernie Bot for the “next year or two.”
Additionally, he stated, “We believe our chip reserves, along with other alternatives, will be sufficient to support lots of AI-native apps for the end users. Inference requires less powerful chips.” “Eventually, China’s speed of AI development will be impacted by its inability to obtain the most cutting-edge chips. We are therefore actively looking for alternatives.
Proactive steps have also been taken by other wealthy Chinese IT businesses in reaction to U.S. export regulations. According to a story published in August by the Financial Times, Baidu, ByteDance, Tencent, and Alibaba bought approximately 100,000 A800 processors from Nvidia to be delivered this year, at a total cost of up to $4 billion. In addition, they invested $1 billion in GPUs that will be delivered in 2024.
Such large initial outlays would easily discourage many startups from competing in the LLM space. There may be an exception if the startup company is able to swiftly raise sizeable funding. 01.AI was established by well-known investor Kai-Fu Lee in late March. After receiving $1 billion in funding, the company was able to purchase a sizable number of high-performance inference chips through loans. The company has now paid off its debt.
Baidu recently introduced the Ernie Bot 4, which Li said is “not inferior in any respect to GPT-4,” using its reserve of GPUs.
Because these AI models are so complicated, rating LLMs is difficult. Numerous Chinese artificial intelligence companies have turned to rank increases by rigorous adherence to LLM chart criteria; nevertheless, the practicality of these models in real-world scenarios remains to be determined.
Smaller AI players will have to make do with less powerful processors that are exempt from US export restrictions since they lack the cash flow to hoard chips. Alternatively, they could wait for possible purchase chances. Li predicts that the sector will soon enter a “consolidation stage” due to a number of issues, such as the dearth of sophisticated chips, the strong need for data and AI talent, and significant upfront investments.
