01.AI, a China-based artificial intelligence startup, has embarked on a shopping spree and procured enough AI GPUs from Nvidia to keep growing for 18 months, its founder Kai-Fu Lee told Bloomberg in a brief interview. This was a smart move, as the U.S. has just restricted exports of these GPUs to China.
"We have stockpiled a lot of Nvidia chips," said Lee.
Truth be told, it is not completely clear what "18 months of Nvidia" GPU means. As companies develop larger models, they need more compute GPUs to train them and even more GPUs to use them. Perhaps Kai-Fu Lee, who used to work at Apple, Microsoft, and Google, meant that the company has a roadmap with larger models and GPU requirements, and the company will have enough GPUs to execute the roadmap for about 1.5 years. Then again, it is unknown how complex these models are and how many GPUs they need.
In any case, for now, 01.AI relies on Nvidia's hardware and CUDA software framework. Kai-Fu Lee is skeptical about China's ability to develop its own AI processors and underlying software that would be competitive with Nvidia's AI and HPC GPUs within the next year and a half, which is when 01.AI's stockpile of Nvidia chips will run out.
"The jury is out on whether China in 1.5 years can make equivalent or nearly as good chips," said Lee.
Lee regrets that because of the ongoing tensions between China and the U.S.. The technology world will be largely divided as China will rely on its own models and, perhaps, standards, whereas the U.S. and the rest of the world will develop their own set of technology.
"We will have two parallel universes," Lee said. "Americans will supply their products and technologies to the U.S. and other countries and Chinese companies will build for China and whoever else uses Chinese products. The reality is that they will not compete very much in the same marketplace."
01.AI has developed its own large language model called Yi-34B, which has outpaced Lllama 2 in some areas. The company is planning to make money by selling specialized versions of Yi-34B and applications to customers.
