Amidst the preclusion of Nvidia's exportation of high-powered graphic processing units to China by Washington, the tech behemoths of the nation had foreseen a burgeoning technological clash between the two countries. In anticipation, tech titan Baidu, a progenitor of China's analogs to OpenAI, strategically amassed a reservoir of these coveted AI chips. CEO Robin Li, in an earnings call this week, revealed that Baidu has acquired a sufficient cache of AI chips to fuel the training endeavors of its ChatGPT equivalent, Ernie Bot, for the foreseeable "next year or two."
Li articulated, "Furthermore, the exigencies of inference demand less potent chips, and we hold the conviction that our chip stockpile, alongside viable alternatives, will prove adequate to underpin a plethora of AI-native applications for end-users. Over the long term, grappling with difficulties in acquiring the apex chips inexorably hampers the momentum of AI development in China. Therefore, we are actively exploring alternative avenues."
Notably, other well-heeled Chinese tech entities are also undertaking preemptive measures in response to stringent U.S. export controls. Baidu, ByteDance, Tencent, and Alibaba collectively placed orders for approximately 100,000 units of Nvidia's A800 processors, incurring a staggering cost of up to $4 billion, as reported by the Financial Times in August. Additionally, they secured GPUs amounting to $1 billion, scheduled for delivery in 2024.
This substantial upfront investment could potentially dissuade nascent players from entering the linguistic model race, unless said startups promptly secure substantial funding. A prime example is 01.AI, established in late March by prominent investor Kai-Fu Lee, which procured a significant number of high-performance inference chips through loans. This financial endeavor was repaid swiftly after garnering capital that bestowed a valuation of $1 billion upon the company.
Armed with its arsenal of GPUs, Baidu recently unveiled the Ernie Bot 4, a creation that Li proudly asserts is "devoid of any inferiority when compared to GPT-4."
Appraising Language Models (LLMs) proves to be a complex task due to the intricate nature of these artificial intelligence constructs. Many Chinese AI enterprises have resorted to elevation in rankings by diligently adhering to the criteria laid out in LLM charts. However, the efficacy of these models in practical real-life applications remains subject to ongoing scrutiny.
For smaller players in the AI landscape, lacking the financial fluidity to stockpile chips, the recourse is to settle for less potent processors not subject to U.S. export controls. Alternatively, they may bide their time for potential acquisition opportunities. Li anticipates that a convergence of factors, encompassing the scarcity of cutting-edge chips, heightened demand for data and AI talent, and substantial initial investments, will propel the industry into a forthcoming "consolidation stage."