China's search giant Baidu is building out every layer of the AI stack—from custom silicon to frontier models to cloud infrastructure—positioning itself as one of the few companies globally attempting true vertical integration in artificial intelligence.

The push comes at a pivotal moment. After years of US export restrictions that locked Chinese companies out of the most advanced chips, the door has cracked open. Nvidia H200 GPUs have been approved for export to roughly ten Chinese firms between January and May 2026. Baidu, ByteDance, and Alibaba are among the recipients.

At its Create 2026 conference, Baidu unveiled a chip roadmap that signals independence from Western semiconductor suppliers. The company's Kunlunxin division launched the M100 chip in early 2026, with the M300 scheduled for early 2027. Both are designed to handle training and inference workloads.

The financial trajectory is striking. Analysts project Baidu's chip sales could increase sixfold to approximately RMB 8 billion—roughly $1.1 billion—by 2026. Macquarie has valued the Kunlun chip unit at around $28 billion, and Baidu has floated plans for a potential separate listing of the Kunlunxin business. Baidu has pursued a full-stack AI approach since 2011.

The company released ERNIE 5.1, its latest frontier model, in May 2026. The model uses 94% less pre-training costs than its predecessor while running on one-third the parameters. It was trained on a Kunlunxin cluster with a 97% training rate.

China's AI landscape shifted meaningfully after DeepSeek released its low-cost frontier model last year, forcing a reckoning with the assumption that cutting-edge AI required cutting-edge hardware. Baidu is taking that lesson further. Rather than merely optimizing models to work on constrained hardware, Baidu is simultaneously improving the chips themselves while gaining selective access to Nvidia's best hardware. Baidu can deploy H200s where available while building domestic alternatives that reduce dependence on US export policy.

The ERNIE 5.1 efficiency gains are arguably the more consequential data point for the broader industry. A 94% reduction in pre-training costs challenges the prevailing narrative that AI progress requires ever-larger capital expenditures on compute.

US export policy remains the wildcard. The approval of H200 sales to Chinese firms represents a loosening from the strictest periods of the chip embargo, but these decisions are inherently political and subject to reversal. Baidu clearly treats the current access window as potentially temporary—which is precisely why it's spending billions on chips it could theoretically simply purchase from Nvidia.