Chinese company Baidu has announced two new AI processors and several large-scale computing products at its annual Baidu World conference. The company's new M100 and M300 chips, developed by its Kunlunxin Technology unit, target inference and training workloads, Reuters and South China Morning Post reports. The M100 chip optimized for inference workloads, is expected to launch in early 2026. The M300, capable of training massive multimodal models with trillions of parameters will follow in 2027. Both aim to deliver "powerful, low-cost, and controllable AI computing power," according to Baidu Cloud president Shen Dou. Baidu has been developing its own processors since 2011 and now joins domestic chipmakers such as Huawei's HiSilicon, Cambricon, MetaX, and Biren in reducing reliance on NVIDIA hardware. Similar to Huawei's CloudMatrix 384 system based on Ascend 910C chips, Baidu is building new "supernode" clusters that combine hundreds of its in-house processors. The Tianchi256 system, integrating 256 Baidu P800 chips, is set to debut in the first half of 2026, followed by the Tianchi512 with 512 chips later in the year. Baidu said the configuration offers over 50% higher performance than its previous generation and plans to scale these clusters to "millions" of chips by 2030.

Baidu founder and CEO Robin Li called the current AI industry structure "unhealthy," criticizing how most profits concentrate at the chip and foundation-model level. He suggested a "reversed industry pyramid," where value shifts toward real-world AI applications. The company also introduced Ernie 5.0, a multimodal large language model with 2.4 trillion parameters, twice the size of Alibaba's Qwen3-Max and Moonshot's Kimi K2. Baidu says Ernie 5.0 improves in text, image, video, coding, and agent tasks through unified multimodal training.

More...