SAN FRANCISCO, Jan 27 (Reuters) - Nvidia announced on Monday the significance of Chinese AI company DeepSeek's progress in demonstrating the capability of its chips for the Chinese market. According to Nvidia, there will be a growing need for their chips in the future to meet the increasing demand for DeepSeek's services.
Following concerns from investors that DeepSeek had achieved similar results as competitors like OpenAI using fewer Nvidia chips than U.S. companies, Nvidia responded with a statement on Monday. This led to Nvidia's shares dropping to $118.58, while Advanced Micro Devices' shares also decreased by over 6% to $115.01.
Nvidia's statement highlighted DeepSeek's innovative approach in creating new models and utilizing widely-accessible models and computing resources that comply with export control regulations. One of DeepSeek's research papers revealed their use of around 2,000 of Nvidia's H800 chips, specifically designed to adhere to U.S. export control rules introduced in 2022.
The U.S. export controls on microchips aimed to restrict China's advancements in supercomputing for military and AI purposes. Jimmy Goodrich, a technology analysis senior adviser at RAND Corp, noted that China possessed several major supercomputers with a substantial number of Nvidia chips legally acquired before DeepSeek's utilization for efficiency improvements.
DeepSeek faced challenges in managing a surge of new users on Monday, a process known in the AI industry as "inference." Nvidia emphasized that the continued demand for its chips is evident in the inference workload.
Nvidia's latest chip offering, the H20, is designed to align with current export control policies. Although these restrictions impact the chip's applicability for AI training, Goodrich acknowledged it as potentially the top chip globally for inference tasks.
In contemplating the longevity of permitting the sale of the premier inference chip to China, Goodrich questioned the stance of Washington on this matter, reflecting on the geopolitical implications.