Baidu, Huawei dominate China’s GPU cloud market: report

Baidu and Huawei held over 70% of China’s GPU cloud market in H1 2025, according to a Frost & Sullivan report.

Baidu, which integrates its Kunlunxin-developed GPUs into its Baige AI platform, led with a 40.4% share, while Huawei followed with 30.1%.

Kunlunxin, started in 2012 as Baidu’s internal chip unit, shipped nearly 70,000 units in 2024, based on IDC data.

Frost & Sullivan said only a few domestic vendors have reached large-scale deployment due to ongoing challenges with hardware, software, and system integration.

Chinese GPU makers, including Baidu, Moore Threads, and Biren Technology, are seeking public listings as Beijing pushes for semiconductor self-sufficiency.

Baidu recently said it confidentially filed for a Hong Kong IPO for Kunlunxin, with details yet to be finalized.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Vertical integration shapes China’s GPU cloud market more than chip speed

The market definition sets a high bar. Only companies that run the full chip-to-cloud chain (from chip design to data center clusters to cloud services) get counted, which favors cash-rich firms 1.
That setup helps Baidu at 40.4% and Huawei at 30.1% dominate. They built full stack software and hardware ecosystems while chasing an alternative to Nvidia’s CUDA (its proprietary GPU programming platform) 1.
Counting only firms that control the chip-to-cloud chain pushes the share to 70%, which tracks uptake of integrated platforms, not total AI accelerator rollouts 1.
Frost & Sullivan said few vendors have reached large deployments 1. Hardware performance and software ecosystems lag. So do system integration and commercial operations 1.

AI infrastructure teams wrestle with moves to domestic accelerators

Domestic GPUs will move from “strategic backups” to “core pillars” in 3 to 5 years. Teams that run CUDA-based models must shift work to Kunlunxin, Huawei Ascend (its AI accelerator line), and other local chips 1.
Current transcompilation tools remain imperfect. Chinese chipmakers pour money into CUDA-compatible stacks, yet compatibility and performance swing across frameworks 2.
Clouds and AI platform vendors can earn by offering migration services. They can ship machine learning operations (MLOps) tools and performance layers that bridge CUDA code to domestic hardware 2.
New computing centers in China run under 30% utilization 3. Software makers can boost resource scheduling and workload tuning across heterogeneous chip clusters (mixing different types of accelerators) 3.