Bloomberg Tech→ original

DeepSeek develops its own AI chip for its neural networks — Reuters

DeepSeek, a Chinese AI startup known for creating cheap alternatives to OpenAI models, has begun developing proprietary semiconductor chips for AI systems…

AI-processed from Bloomberg Tech; edited by Hamidun News
DeepSeek develops its own AI chip for its neural networks — Reuters
Source: Bloomberg Tech. Collage: Hamidun News.
◐ Listen to article

DeepSeek, a Chinese artificial intelligence startup, has begun developing its own semiconductor chips to power AI systems. This was reported by Reuters on July 7, 2026, citing unnamed sources. No official confirmation has been received from DeepSeek.

Why does DeepSeek need its own chip?

The main reason is limited access to advanced accelerators. American export restrictions, consistently tightened since 2023, have cut off Chinese organizations' access to Nvidia's H100, H200, and A100 chips — the very accelerators used by OpenAI, Google DeepMind, and Anthropic to build their infrastructure. Even the H20 series, created by Nvidia specifically for the Chinese market with reduced specifications, was added to the restrictions list in 2025.

DeepSeek has already demonstrated the ability to squeeze maximum value from limited hardware: in early 2025, the startup released a series of models comparable in quality to OpenAI's flagship solutions but significantly cheaper to train. This approach — algorithmic efficiency instead of computational power — made DeepSeek a global sensation. Its own chip is the logical next step: optimizing hardware architecture for its own algorithms is far more efficient than adapting algorithms to someone else's hardware.

A proprietary hardware platform would potentially give the company:

  • Independence from Nvidia supplies and American export policy
  • The ability to optimize microarchitecture for the specifics of its own transformers
  • Long-term cost reduction in training and inference
  • Competitive advantage amid ongoing technological blockade

Who has already taken this path

Creating one's own chip is a practice that the largest Western players adopted long ago, but each reached results over years and billions of dollars. Google has been developing its TPU series since 2016 — today they form the foundation of Gemini's training. Amazon released Trainium for model training and Inferentia for inference. Meta develops MTIA accelerators, Microsoft invests in Maia chips for Azure.

In China, other companies are following a similar path. Huawei created the Ascend accelerator series, used in major domestic AI projects. Baidu develops multiple generations of Kunlun chips. Cambricon specializes in neural network processors and is traded on an exchange. All these examples show: the Chinese industry is capable of moving toward chip sovereignty, although the technological gap with Western leaders remains significant.

A fundamental problem for any Chinese chip program is manufacturing. TSMC, which produces the world's most advanced chips, is inaccessible to Chinese clients due to U.S. sanctions pressure. SMIC, China's largest manufacturer, is technologically limited to older process nodes — which significantly affects energy efficiency and performance of the future chip.

What remained behind the scenes

Reuters does not disclose the technical specifications of the future chip, does not name development timelines, and does not specify the investment volume in the project. It is unknown whether the work is at the stage of early architectural research or already manufacturing prototyping. It is unclear whether the chip will be designed for model training, inference, or both tasks. No official comments have been received from DeepSeek.

This is a typical picture for articles about Chinese technology projects in semiconductors: companies rarely announce developments before the product is ready — especially in a strategic area that has become the center of technological confrontation between the United States and China.

What this means

If the project comes to fruition, DeepSeek will join a narrow list of companies controlling the full AI stack — from algorithms to silicon. For the global industry, this signals: the fragmentation of AI infrastructure into Western and Chinese ecosystems is becoming increasingly irreversible, and the struggle for technological sovereignty is expanding beyond software to the hardware level.

*Meta is recognized as an extremist organization and is banned in the Russian Federation.

⧉ Story
ZK
Hamidun News
AI news without noise. Daily editorial selection from 400+ sources. A product by Zhemal Khamidun, Head of AI at Alpina Digital.

Want to stop reading about AI and start using it?

AI News is a curated feed of AI/tech news. Hamidun Academy teaches you to use AI systematically in your work.

What do you think?
Loading comments…