Long before Washington banned Nvidia’s exports of high-performance graphics processing units to China, the country’s tech giants were hoarding them in anticipation of an escalating tech war between the two nations.
Baidu, one of the tech companies building China’s counterparts to OpenAI, has secured enough AI chips to continue training its ChatGPT counterpart Ernie Bot for the “next year or two,” said the company’s chief executive Robin Li to you earnings call this week.
“Also, inference requires less powerful chips, and we believe our chip stocks, as well as other alternatives, will be sufficient to support many native AI applications for end users,” he said. “And in the long run, the difficulty in obtaining the most advanced chips will inevitably affect the rate of development of artificial intelligence in China. Therefore, we are proactively looking for alternative solutions.”
Other deep-pocketed Chinese tech companies have also taken precautionary measures in response to US export controls. Baidu, ByteDance, Tencent and Alibaba have collectively ordered about 100,000 units of Nvidia’s A800 processors to be delivered this year, at a cost of up to $4 billion, according to the Financial Times. mentionted in August. They also bought $1 billion worth of GPUs scheduled to be delivered in 2024.
Such heavy upfront investments could easily deter many startups from entering the LLM race. Exceptions exist if the new business manages to secure handsome investments quickly. 01.AI, which was founded in late March by prominent investor Kai-Fu Lee, acquired a significant number of high-performance inference chips through loans and has already paid off its debt after raising funds that valued it at $1 billion.
With its GPU inventory, Baidu recently released Ernie Bot 4, which Li claimed is “in no way inferior to GPT-4.”
Evaluating LLMs is difficult thanks to the sheer complexity of these AI models. Many Chinese AI companies have resorted to boosting rankings by diligently fulfilling the criteria of LLM charts, but the effectiveness of these models when applied to actual real-life applications is still pending.
Smaller AI players, lacking the cash flow to accumulate chips, will have to make do with less powerful processors that are not subject to US export controls. Alternatively, they can wait for potential acquisition opportunities. Li expects that with a confluence of factors, including the scarcity of advanced chips, high demand for data and AI talent, and huge initial investment, the industry will soon enter a “consolidation stage.”
