In recent years, as generative AI has shifted from model training toward large-scale inference, an increasing number of AI model developers have turned their attention to underlying hardware.
Chinese AI startup DeepSeek has launched an in-house AI chip project focused on inference workloads, according to Reuters. The company aims to reduce inference costs through custom-designed processors while lowering its reliance on overseas suppliers such as NVIDIA.
Currently, DeepSeek has yet to comment publicly.
The report said the project remains in its early stages and is designed primarily for AI inference rather than model training. As generative AI adoption accelerates, inference has become one of the fastest-growing segments of the AI market.
Unlike training, which requires concentrated bursts of computing power, inference must continuously serve large volumes of user requests, making cost efficiency, power consumption and system reliability increasingly important.
