According to exclusive news from Reuters, DeepSeek, which has made a name for itself with efficient large models, has secretly advanced its self-developed AI chip project, focusing on inference specific chips and abandoning the path of general-purpose training GPUs. Once the project is implemented, this AI company known for its models will officially cross over into the semiconductor industry, completing a major strategic shift that will impact the industry landscape.

The project is low-key and stealthy, with a layout cycle of over a year

According to foreign media reports, DeepSeek's self-developed chip program was launched a year ago and is currently in the early stages of implementation, maintaining extremely low exposure throughout the process. The full chain cooperation negotiation of the hardware link is being synchronously promoted: the company has actively connected with chip design, wafer foundry, and storage leading manufacturers, and cooperated with multiple parties to polish the overall solution of the inference chip. The actions of the talent side are also secretive. In recent months, DeepSeek has significantly expanded its recruitment of chip design engineers, but all of them have been smuggled out of the interview channels and no relevant positions have been released on any public recruitment platforms, deliberately avoiding industry attention.

Industry analysis suggests that DeepSeek's decision to pursue low-key research and development is partly due to the long semiconductor research and development cycle, high uncertainty, and unwillingness to release expectations in advance; On the other hand, it is also to reduce the game pressure of upstream computing power manufacturers and smoothly promote supply chain cooperation.

Abandoning training and specializing in reasoning: stepping on the core turning point of AI computing power demand
Unlike most manufacturers that prioritize training high computing power GPUs, DeepSeek directly bets on the inference track, accurately seizing the huge changes in the current AI industry's computing power structure.
In the past, the industry generally focused on investing computing power in the model pre training stage. However, with the widespread use of chatbots, code assistants, enterprise search, intelligent office, and AI agents, the logic of computing power consumption in the industry has completely reversed: massive computing power is no longer used for one-time training, but continuously consumed in the inference process of user real-time interaction and model response.
Universal GPUs balance training and inference, with high architecture redundancy, energy consumption, and hardware costs; And dedicated inference chips can perform targeted architecture optimization for fixed models and vertical scenarios, naturally possessing three major advantages of low power consumption, low cost, and low latency.
For DeepSeek, which focuses on lightweight and cost-effective models, its self-developed inference chip has unique collaborative value: the hardware architecture is deeply adapted to its own large-scale model, which can compress cloud and local deployment costs from the root, further widen the API pricing gap with competitors, and consolidate its commercial advantages.

Behind self-developed chips: a strategic self rescue to reduce computing power dependence
Currently, top AI companies in China are generally facing dual constraints on computing power supply: limited supply of high-end NVIDIA GPUs and continuously rising procurement costs; Although chips such as Ascend continue to adapt, they rely on external hardware for a long time, making it difficult to achieve the optimal solution for chip mold integration.
DeepSeek has launched its self-developed inference chip, with a clear core demand: to reduce the dual dependence on externally purchased computing hardware and gain autonomy in AI infrastructure.
Cost autonomy and controllability: Get rid of the premium of general GPU, significantly reduce the hardware cost of inference services through integrated software and hardware optimization, and provide more competitive AI services for B2B enterprise customers;
Performance exclusive optimization: No longer limited by third-party chip architectures, hardware computing units can be customized based on their own MoE architecture, multi head latent attention, and other model characteristics;
Supply chain risk hedging: Connect the chip design and OEM cooperation links, establish self owned computing hardware backup solutions, and avoid external chip supply fluctuation risks.
According to Reuters, if DeepSeek's inference chip development is implemented, it will mark the official completion of the full stack layout of "model self-developed → chip self-developed" by China's top model manufacturers, which is a significant strategic shift in the company's development path.

New Changes in the Industry: Collective Off site Manufacturing of Chips by Large Model Companies has become a Trend
Looking at the world, Google, Amazon, and Meta have all launched self-developed inference dedicated chips, and splitting the hardware routes of training and inference is a recognized long-term trend in the industry; Domestic manufacturers such as Cambricon and Suiyuan have been deeply involved in the development of inference chips for many years. Now, large model native enterprises are entering the market across borders, and competition in the field is further intensifying.
The entry of DeepSeek will form a new competitive logic: in the past, chip manufacturers adapted to various large models, while in the future, top model manufacturers will hold self-developed chips, achieve a "model+chip" binding loop, and build cost and performance barriers that are difficult to replicate.

In the short term, the DeepSeek chip project still has a long R&D, chip fabrication, and mass production cycle, and will not quickly replace existing Nvidia and Huawei computing power bases in the short term; In the long run, this low-key chip manufacturing plan may reshape the supply pattern of AI computing power in China.