Analysis of the Localization Level of Baidu Kunlunxin Chips

Based on the searched information, the analysis of the localization level of Baidu AI chips (Kunlunxin) is as follows:

Baidu Kunlunxin has formed a complete product matrix, with localization level ranking in the first tier of domestic AI chips:

Product	Mass Production Time	Process Technology	Architecture	Localization Features
Kunlunxin 1st Gen	2018	14nm	XPU Architecture	Independent Design + Domestic Foundry
Kunlunxin 2nd Gen	August 2021	7nm	XPU-R Architecture	Independent Design + Domestic Foundry
Kunlunxin P800	2024	Advanced Process	XPU-P Architecture	Independent Architecture Design

: Kunlunxin adopts a fully self-developed XPU architecture, supporting mixed computing of multiple data precisions such as FP32, FP16, and INT8, with high throughput and low latency characteristics [1][2].

: The P800 chip uses the self-developed XPU-P architecture, supporting High Bandwidth Memory (HBM) and DDR4 memory, compatible with the PyTorch ecosystem, and supporting large model training scenarios [3].

: Kunlunxin P800 has undertaken most of the computing tasks for Baidu Search inference, and has achieved large-scale deployment in scenarios such as Baidu Smart Cloud and Intelligent Computing Centers [4].

According to public information, the localization of Kunlunxin is mainly reflected in:

1. Chip Design
   : Fully independent and controllable, using self-developed XPU architecture
2. Software Ecosystem
   : Compatible with mainstream AI frameworks, reducing migration costs
3. Supply Chain
   : Although process technology still relies on external foundries, it has achieved the maximum domestic substitution within the current domestic foundry capacity

According to relevant data, Kunlunxin ranks second in shipment volume in China’s domestic AI chip market, second only to Huawei HiSilicon [5]. In the 2025 domestic AI chip market pattern, Huawei accounts for 62% of the share, domestic AI chips account for about 35% overall, and Kunlunxin is an important part of it [6].

Kunlunxin is accelerating the localization process:

• The next-generation M100 chip is expected to be launched in 2026, optimized for large-scale inference scenarios
• The M300 chip is expected to be launched in 2027, targeting ultra-large-scale multimodal model training and inference
• The 10,000-card-level AI cluster has been successfully delivered, and the domestic cluster capability continues to improve [5]

---

Baidu Kunlunxin has achieved a high localization rate in the AI chip field, mainly reflected in:

1. Independent design of core architecture
   (XPU series architecture)
2. Independent control of key algorithms and software stacks
3. Realization of large-scale domestic substitution in inference scenarios
4. Steady market share ranking second among domestic AI chips

Although the foundry link of advanced process technology is still restricted by the global industrial chain, Kunlunxin has been in a leading position in the domestic AI chip localization process. With the gradual maturity of the domestic semiconductor industrial chain, the localization rate of Kunlunxin is expected to further improve.

---

:
[1] Supplyframe Sifangwei - “DeepSeek: Helping Chinese Chips Break Through” (https://cn.supplyframe.com/article/8309.html)
[2] QbitAI - “Baidu’s AI Chip Unit Rushes for IPO: Shipment Volume Ranks Second Among Domestic Players” (https://www.qbitai.com/2026/01/366454.html)
[3] Guosen Securities In-depth Report - Baidu Group-SW (09888.HK) In-depth Report (http://testtoo1.oss-cn-hangzhou.aliyuncs.com/eastmoney_pdf/AP202509191746456759.pdf)
[4] CNBC - “Baidu’s semiconductor unit Kunlunxin files for Hong Kong listing” (https://www.cnbc.com/2026/01/02/baidus-semiconductor-kunlunxin-hong-kong-ipo-ai-chips-listing-china.html)
[5] Reports from financial media such as 36Kr and Business Insider