Particle.news

DeepSeek Developing Its Own AI Inference Chip

The reported project seeks hardware control to cut reliance on Nvidia and Huawei by using tailored, lower‑power processors for running models.

Overview

  • Reuters and multiple outlets reported Tuesday that DeepSeek has quietly worked for about a year on an in‑house inference chip and that the effort remains at an early stage.
  • The chip is intended for inference, the step when a trained model generates user responses, which uses different hardware than the GPUs used to train models.
  • Sources say DeepSeek has privately increased hiring of chip‑design engineers and held talks with chip‑design firms, foundries and memory suppliers to build the project.
  • The plan faces major hurdles because U.S. export controls block Chinese firms from buying Nvidia’s most advanced processors and limit access to leading overseas foundries and high‑bandwidth memory.
  • Markets and strategy could shift if the chip succeeds: Nvidia shares slipped about 2% after the report, DeepSeek is opening to outside funding in a large round, and a homegrown chip could squeeze Huawei’s domestic position while requiring years and large capital to scale.