Even before Washington restricted Nvidia’s exports of high-performance graphics processing units (GPUs) to China, the country's technology giants had begun stockpiling these critical components in preparation for an intensifying tech conflict.
Baidu, a leader among companies developing China's versions of OpenAI, has secured enough AI chips to continue training its ChatGPT alternative, Ernie Bot, for at least the next year or two, according to CEO Robin Li during a recent earnings call.
“Additionally, while inference tasks require less powerful chips, we are confident our chip reserves and other alternatives will adequately support a variety of AI-native applications for end users," Li stated. "However, over the long term, challenges in obtaining the most advanced chips will inevitably slow the pace of AI development in China, so we are actively exploring alternatives.”
In response to U.S. export restrictions, other financially robust Chinese tech firms have also taken preemptive measures. Companies like Baidu, ByteDance, Tencent, and Alibaba collectively ordered around 100,000 A800 processors from Nvidia for delivery this year, with a total investment of up to $4 billion, as reported by the Financial Times in August. They have also committed $1 billion for GPU purchases scheduled for delivery in 2024.
These significant upfront investments may deter many startups from entering the large language model (LLM) landscape unless they can attract substantial funding swiftly. One notable exception is 01.AI, founded in late March by prominent investor Kai-Fu Lee, which has acquired a substantial number of high-performance inference chips through loans. Remarkably, it has already repaid its debts after securing capital that valued the company at $1 billion.
With its ample GPU supply, Baidu recently launched Ernie Bot 4, which Li claims “matches GPT-4 in every respect.”
Comparing LLMs is challenging due to the inherent complexity of these AI systems. Many Chinese AI firms have resorted to ranking manipulation by meticulously meeting the criteria set forth in LLM evaluations, but the real-world effectiveness of these models remains uncertain.
Smaller AI players, often unable to amass significant chip reserves, may have to rely on less powerful processors that are not subject to U.S. export controls. Alternatively, they may consider waiting for potential acquisition opportunities. Li anticipates that the combination of advanced chip scarcity, high demand for data and AI talent, and substantial upfront investments will soon lead the industry into a "consolidation stage."
Ranking
