Tech companies hoarded high-performance GPUs before Washington banned their exports to China in anticipation of a tech war.
In an earnings call this week, Baidu CEO Robin Li said the company has enough AI chips to train Ernie Bot for “next year or two.”
“Inference requires less powerful chips, and we believe our chip reserves and other alternatives will support many end-user AI-native apps,” he said. Limited access to advanced chips slows China’s AI development. So we are looking for alternatives.”
Wealthy Chinese tech companies have taken precautions against U.S. export controls. In August, the Financial Times reported that Baidu, ByteDance, Tencent, and Alibaba bought 100,000 Nvidia A800 processors for $4 billion. Also ordered: $1 billion GPU for 2024.
Due to high upfront costs, many startups may avoid LLM. Young businesses that quickly secure large investments are exceptions. In late March, prominent investor Kai-Fu Lee founded 01.AI, which borrowed to buy many high-performance inference chips and paid off its debt after raising $1 billion.
Li said Baidu’s GPU-powered Ernie Bot 4 is “not inferior in any respect to GPT-4.”
The AI model’s complexity makes LLM ratings difficult. LLM chart criteria help many Chinese AI firms rank higher, but their real-world efficacy is unknown.
Smaller AI players without the cash to hoard chips will have to use less powerful processors without U.S. export controls. They could also wait for acquisitions. With the scarcity of advanced chips, high demand for data and AI talent, and huge upfront investments, Li expects the industry to enter a “consolidation stage.”