It’s now available on all major platforms
Well ahead of Washington’s prohibition on exporting high-performance graphic processing units to China, the nation’s tech giants had already begun stockpiling these units in anticipation of an escalating tech conflict between the two countries.
Baidu, a leading tech company involved in developing China’s versions of OpenAI, has successfully acquired enough AI chips to sustain training its counterpart to ChatGPT, Ernie Bot, for the next one or two years, as revealed by the firm’s CEO Robin Li during an earnings call this week.
Li stated, “Moreover, for inference, less potent chips suffice, and we are confident that our chip reserves, along with other alternatives, will adequately support numerous AI-centric applications for end-users. However, in the long term, the challenges in accessing the most cutting-edge chips will inevitably impede the pace of AI advancement in China. Hence, we are actively seeking alternative solutions.”
Other well-financed Chinese tech firms have also taken proactive steps in response to U.S. export regulations. According to the Financial Times in August, Baidu, ByteDance, Tencent, and Alibaba collectively placed orders for approximately 100,000 units of Nvidia’s A800 processors to be delivered this year, amounting to a cost of up to $4 billion. They also secured $1 billion worth of GPUs scheduled for delivery in 2024.
These substantial upfront investments could discourage numerous startups from entering the field of advanced language model (LLM) development. However, exceptions exist if a young business swiftly secures significant investments. For instance, 01.AI, established in late March by prominent investor Kai-Fu Lee, managed to obtain a considerable number of high-performance inference chips through loans and subsequently cleared its debt after raising capital that valued the company at $1 billion.
Armed with its GPU reserves, Baidu recently introduced the Ernie Bot 4, which Li confidently claimed is “on par with GPT-4 in every aspect.”
Evaluating LLMs poses challenges due to their intricate nature. Many Chinese AI firms resort to improving their rankings by meticulously meeting the criteria outlined in LLM charts. However, the practical effectiveness of these models in real-life applications remains under evaluation.
Smaller AI entities, lacking the financial capacity to hoard chips, might have to settle for less potent processors not subjected to U.S. export controls. Alternatively, they could wait for potential acquisition opportunities. Li anticipates that a combination of factors—including the scarcity of advanced chips, high demand for data and AI expertise, and significant upfront investments—will lead the industry into a “consolidation stage” in the near future.
