Baidu expands AI chip business as US Nvidia limits persist

Baidu is expanding its AI chip business in China as US restrictions limit Nvidia’s ability to supply the market.

The company’s Kunlunxin chip unit designs and sells processors for data centers, AI models, and cloud services.

Baidu uses a mix of its own chips and Nvidia products, and this month shared a five-year plan for Kunlun AI chips, starting with the M100 in 2026 and the M300 in 2027.

Earlier in 2025, Kunlunxin secured orders from suppliers to China Mobile.

Analysts from Deutsche Bank and JP Morgan say Baidu is well-placed to meet rising domestic demand as Chinese tech giants like Alibaba and Tencent face semiconductor shortages.

🔗 Source: CNBC

🧠 Food for thought

Implications, context, and why it matters.

Kunlunxin’s foundry and node unknown

Baidu announced M100 and M300 for 2026 and 2027, pitched as low-cost, controllable AI computing power 1, without naming the foundry (a contract chip manufacturer) or the process node (the manufacturing generation, measured in nanometers).
Missing data on the manufacturing partner or yields (the percentage of usable chips per wafer) clouds volume plans for Baidu’s Kunlunxin (its AI chip unit), while Huawei’s Ascend chips (Huawei-designed AI accelerators) have a roadmap through 2028 with in-house High Bandwidth Memory (HBM) 2.
Baidu’s Tianchi256 supernode (an internal AI compute cluster) claims over 50% better performance than its prior cluster 1, though that internal yardstick blurs comparisons with Huawei’s CloudMatrix 384 (a Huawei data center AI cluster configuration) or Nvidia’s systems 3.

Software firms can profit

Chinese graphics processing unit (GPU) vendors like Kunlunxin ship software stacks inspired by Compute Unified Device Architecture (CUDA), Nvidia’s GPU programming platform. Some claim native CUDA support 4, though compatibility and performance at scale remain unproven on Kunlun chips.
Third-party software firms can review Kunlunxin’s PyTorch fork (a customized version of the PyTorch machine-learning framework) and Software Development Kit (SDK) documentation 5 6 to spot support gaps. They can build migration and optimization services for Chinese enterprises moving workloads from Nvidia to domestic chips.
Alibaba and Tencent face semiconductor shortages, and the Chinese government directed companies to limit foreign chips 3. That creates near-term demand for services to ease moves to domestic AI accelerators for AI infrastructure specialists (consulting and systems-integration firms).