DeepSeek AI Startup Develops Secret Proprietary Inference Chips To Cut Dependence On Nvidia And Huawei Amid Geopolitical Challenges And External Funding Shifts
A highly,visible Chinese AI start,up, DeepSeek, is working in secret on its own proprietary chip, according to 3 sources with knowledge of the matter who talked to Reuters. The project aims to cut back the company's dependence on chips from Nvidia and Huawei, the sources said. The hardware effort is a big change for a company which has been successful worldwide in a short space of time thanks to its super,efficient
models of software.
Instead, the startup is concentrating its design on inference hardware: a subclass of processor designed explicitly to host existing models and produce responses to users’ inputs as quickly as possible. Serving active applications uses a fraction of the energy required to train a model, allowing cheaper custom inference chips to serve very high volumes of queries. This shift brings DeepSeek in line with other international competitorsOpenAI and Anthropic, for examplewho have also begun designing custom silicon for the same reason.
The development project commenced about 1 year ago and has only recently moved into a higher activity phase. The start,up has covertly trained up and recruited specialist chip design engineers, avoiding public advert sites and making the project as secret as possible. The company has also started talking to other design companies, foundries and memory manufacturers to enable a viable production pipeline.
Proprietary hardware may have great geopolitical implications. Existing export controls in the US do notallow China companies to buy Nvidia's latest graphics chips. DeepSeek, who already trained the muchpublicised R1 reasoning system on Nvidia H800 silicon, have had to involve more and more of Huawei's Ascend chips due to tightening export restrictions. Huawei holds roughly half of China's 50 billion dollardomestic market for chips, but faces competition from Alibaba and Baidu, who have also produced customprocessors.
Rick Windsor at Radio Free Mobile said Nvidia is effectively at 0 in China because of the export bans, so DeepSeek has zero chance of any export sales for its own custom siliconand the company will have to get into a good overseas manufacturing facility if it is to have any success outside of China. Foundries are extremely difficult to get into, and it's not simply the manufacturing you need, but also the high bandwidth memory you need for modern AI processors.
The expansion in hardware is also timed with an underlying paradigm shift in the company's financial approach. Previously staunchly opposed to external funding mechanisms, the startup had begun to gear up for a mammoth 7,000,000,000 USD fundraising. Such a huge initial injection of capital would have valued the firm at between USD52 billion to 59 billion, giving it the huge financial war chest that is indispensable for the long term development of semiconductor chips. Although the startup refused to officially comment on the chip project, the quiet recruitment of hordes of engineering manpower indicates a genuine desire to back its own hardware channels.
