DeepSeek is said to attract talent to develop AI chip on its own due to difficult access to Western products.
Digitimes cited sources in the semiconductor industry in China said DeepSeek has initiated recruitment campaign to find talent design sales leads. The company specifically targeted to experts about high-performance computing (HPC) and optimization algorithm to accelerate efforts to technology development and commercialization.
DeepSeek has not officially announced plans to develop chips for the app, or cloud, but according to the sources, dynamics recruitment this clearly demonstrates the ambition of the company in the field of AI chip.
“Global competition related to model language (LLM) increasingly fiercer, self-control of the computer has become necessary. This put DeepSeek into position similar to OpenAI in self-development, processors private,” a source said.
Another source said the current strategy of DeepSeek is use of GPU low-level, lower level and processor computing inland, such Huawei Ascend, to maximize the likelihood of training AI. Recently, 16 chip company, ONE of China, which has Huawei, Moore Threads and Hygon, has announced the template's own chips with compatibility or integration with services DeepSeek, ensuring access support from the cloud providers such as Huawei Cloud Tianyi Cloud and Tencent Cloud.
Besides chip local DeepSeek still use the GPU has been buying from Nvidia. Previous SemiAnalysis get the Chinese company might have approached approximately 10,000 model Nvidia GPU H800 up performance to comply with the ban of American and about 10,000 GPU H100. Tell CNBC, Alexandr Wang, CEO của Scale AI, cũng tiết lộ ông nắm thông tin DeepSeek đang sở hữu 50.000 chip Nvidia H100.
“The chip samples in water is not able to help DeepSeek keep up the world, particularly in the application of training and inference, large-scale,” a source said. “Therefore, initiative, self-developed chip of DeepSeek can help reduce external risks, improve performance and power consumption for the pattern AI of the company and create differentiated solution in accordance with specific requirements”.
Even so, DeepSeek also are faced with many challenges on the self-developed chip, consisting need to establish a solid foundation in R&D technology and meet international standards of development, design and fabrication of core technology. Therefore, long-term strategy of the company for chip self-development is still uncertain.
DeepSeek not yet given comment.
The self-developed chips custom trend is in the field of selling leads to reduction depends on Nvidia – company holding more than 80% market share AI chip. The most recent, according to Reuters, OpenAI đang hợp tác với Broadcom và TSMC để phát triển chip suy luận tự thiết kế đầu tiên, dự kiến phát hành vào năm 2026.
DeepSeek được thành lập năm 2023 bởi Liang Wenfeng, đồng sáng lập của quỹ đầu tư chuyên về AI High Flyer. Công ty bắt đầu phát triển DeepSeek vào tháng 4/2023, trước khi ra mắt mô hình suy luận R1 hôm 20/1.
Source: vnexpress.net