Baidu

Full briefing

Baidu, Inc. (NASDAQ: BIDU; HKEX: 9888) is headquartered in Beijing and is best known as China's dominant internet search engine, commanding approximately 60–70% of the Chinese search market. However, Baidu has positioned itself as one of China's leading AI companies since at least 2013, when it established a Silicon Valley-based AI lab under Andrew Ng — predating the LLM era by nearly a decade. Baidu's AI revenue comes from its AI Cloud division, autonomous driving unit Apollo, and AI-powered search advertising, with total company revenue of approximately $19 billion in fiscal year 2024. Baidu's ERNIE (Enhanced Representation through kNowledge IntEgration) language model series has been in development since 2019, making it one of the longest-running LLM development programs among Chinese companies. ERNIE Bot was launched publicly in March 2023, becoming one of China's first mainstream consumer-facing LLM chatbots. The ERNIE 4.0 model, released in October 2023, was positioned as competitive with GPT-4 on Chinese-language benchmarks. The Wenxin (文心) brand unifies Baidu's AI model portfolio — Wenxin Yiyan (文心一言) is the consumer chatbot, Wenxin Qianfan (文心千帆) is the enterprise developer platform offering API access to ERNIE and third-party models, and Wenxin Agent provides autonomous agent capabilities for enterprise workflows. Baidu's most distinctive supply chain asset is its in-house AI chip program. The Kunlun (昆仑) AI chip series was unveiled in 2018 and first mass-produced in 2020. Kunlun 1 was fabricated on Samsung Foundry's 14nm process; Kunlun 2, released in 2021, moved to Samsung's 7nm process. Unlike Alibaba's Hanguang 800, which is inference-only, Kunlun 2 targets both training and inference workloads. Baidu claims to have deployed over 200,000 Kunlun chips across its data centers. The Kunlun 3, targeting 4nm (likely Samsung Foundry's 4nm process), was in development as of 2024 but has faced delays given Samsung's 4nm yield challenges. The Kunlun program gives Baidu partial independence from foreign GPU supply — a strategic priority that intensified after the 2022–2023 export controls. The export control timeline significantly shaped Baidu's hardware procurement. Baidu had accumulated substantial NVIDIA A100 inventory before the August 2022 BIS rules restricted A100/H100 sales to China. After the October 2023 rule expansion, Baidu transitioned to NVIDIA H20 GPUs for workloads requiring CUDA compatibility, while scaling its Kunlun clusters for workloads that run on Baidu's proprietary PaddlePaddle deep learning framework. PaddlePaddle — Baidu's open-source deep learning framework, released in 2016 — is optimized for Kunlun chip execution and provides a degree of supply chain independence from NVIDIA's CUDA ecosystem that frameworks like PyTorch and TensorFlow do not. Baidu's Apollo autonomous driving unit is also a significant AI chip consumer, running real-time inference on LiDAR point clouds, camera streams, and sensor fusion pipelines in both data center training environments and edge vehicle compute units. Apollo has deployed over 700 robotaxis in commercial service in cities including Wuhan, as of early 2025, with Apollo Go accumulating over 7 million paid rides. This autonomous driving workload creates a diverse AI inference demand that encompasses both cloud training (on Kunlun/H20 clusters) and edge inference (on NVIDIA Drive Orin and proprietary edge NPUs) — making Baidu's supply chain exposure broader than a pure cloud AI company.