Economic Observer reporter Ren Xiaoning

On July 7th, the big model companies DeepSeek and Zhipu were revealed to have developed their own inference chips on the same day. According to reports, DeepSeek is in talks with chip design companies, wafer foundries, and storage enterprises. Zhipu has made preliminary inquiries to some domestic chip design companies to explore joint development of specialized AI processors.

Regarding the matter of self-developed chips, the reporter sought confirmation from Zhipu, but as of press time, no response has been received. DeepSeek has also not made a public response.

The self-developed chips exposed by both companies are inference chips, not training chips. The training chip is mainly used for AI model training, with more computing cores, high-speed memory interfaces, and a large amount of storage space; The inference chip is used for the deployment and application of AI models, and has the characteristics of low latency, high energy consumption ratio, and miniaturized design.

A senior engineer from a leading domestic chip company told Economic Observer that Big Model Company focuses on inference chips for two main reasons:

At the current stage, it will be relatively easier for domestic companies to independently develop inference chips. Because training chips require high accuracy and computing power, once the accuracy collapses during training, it means that a long period of money will be wasted. The inference chip is based on the existing model weights, and even if the chip's ability is insufficient, the loss is relatively small, which can be gradually adjusted.

Another reason is that inference chips are more practical nowadays. After the outbreak of agents this year, the demand for large model inference has grown rapidly, and domestic chips are in short supply. If we can successfully develop our own inference chip, it can fill some gaps.

The senior engineer stated that currently, domestic large model manufacturers are in a state of computing power shortage.

On social media platforms, users of Zhipu often report issues with new models reporting errors and the need to queue up to purchase personal packages, making it difficult to grab them. DeepSeek has collapsed "often becomes a hot search topic. According to user statistics, DeepSeek has crashed more than 18 times this year. Tencent executives have repeatedly mentioned during earnings calls that Tencent hopes to buy more cards. A source from Volcano Engine told reporters that the API service for the Seedance model is currently in short supply.

Zhang Faen, a senior artificial intelligence engineer in Beijing, believes that large model companies cannot meet the demand for self-developed chips in the short term, and during the transition period, they need to continue to buy cards for mixed use. Building chips is something that big model companies must do, because big model companies can use their smartest AI to polish the underlying hardware that is most suitable for them, squeezing performance to the extreme. Once the model is successfully validated, future profit margins and ecological initiative will be guaranteed.

OpenAI has proven the cost-effectiveness of self-developed chips. On June 24th, OpenAI and Broadcom jointly released the first customized inference chip. Broadcom CEO Chen Fuyang stated that the performance of this chip in inference tasks is comparable to that of Nvidia Blackwell chips and Google TPU, and the cost can be reduced by about 50%.

However, senior engineers from the top domestic chip companies mentioned above believe that there is uncertainty about the sustainability and success of domestic large model companies' self-developed chips. Self developed chips are a high investment, long-term business, and the revenue level of domestic large model companies cannot currently compare with OpenAI.

The chip manufacturers with mature products in China, whether Internet manufacturers or independent chip companies, have been deployed for many years. Baidu Kunlun Chip was established in 2011 and has been in existence for 15 years; The layout of Alibaba's Pingtou Brother began in 2018 and has been around for about 8 years now; Independent chip companies such as Cambricon, Suiyuan Technology, Tiantian Zhixin, and Biren Technology have also been investing for more than 7 years to achieve large-scale deployment.

There have also been cases of chip development failing halfway. OPPO's chip company ZEKU will shut down in 2023, and its team of over 3000 people will be disbanded. JD.com once recruited chip talents with high salaries last year. In less than half a year, in April 2026, the JD.com chip team revealed the news of the company's main body switching.