Long before Washington banned exports of Nvidia’s high-performance graphics processing units to China, the country’s tech giants had been hoarding them in anticipation of an escalating tech war between the two nations.
Baidu, one of the tech companies building OpenAI’s Chinese counterparts, has secured enough AI chips to continue training its ChatGPT equivalent, Ernie Bot, for “the next year or two,” the company’s CEO said. Robin Li, on an earnings conference call this week.
“In addition, inference requires less powerful chips, and we believe that our chip reserves, as well as other alternatives, will be sufficient to support many native AI applications for end users,” he said. “And in the long term, difficulties in acquiring the most advanced chips inevitably impact the pace of AI development in China. Therefore, we are proactively looking for alternatives.”
Other deep-pocketed Chinese tech companies have also been taking proactive steps in response to U.S. export controls. Baidu, ByteDance, Tencent and Alibaba collectively ordered around 100,000 units of Nvidia A800 processors to be delivered this year, costing them up to $4 billion, the Financial Times reported in August. They also purchased $1 billion worth of GPUs scheduled for delivery in 2024.
Such large initial investments could easily deter many startups from entering the LLM career path. There are exceptions if the young company quickly obtains good investments. 01.AI, founded in late March by prominent investor Kai-Fu Lee, acquired a substantial amount of high-performance inference chips through loans and has already paid off its debt after raising capital that valued it at $1 billion .
With its GPU stockpile, Baidu recently launched the Ernie Bot 4, which Li says is “in no way inferior to the GPT-4.”
Grading LLMs is complicated thanks to the great complexity of these AI models. Many Chinese AI companies have resorted to improving classification by diligently meeting the criteria of LLM graphs, but the effectiveness of these models when applied to real-life real-world applications is still up for trial.
Smaller AI players, lacking the cash flow to stockpile chips, will have to make do with less powerful processors that are not under US export controls. Alternatively, they can wait for potential acquisition opportunities. Li hopes that with a confluence of factors, including a shortage of advanced chips, high demand for data and AI talent, and huge upfront investments, the industry will soon move into a “consolidation stage.”