Baidu answers China’s call for home-grown silicon with custom AI accelerators
Summary
Baidu has unveiled a new generation of custom AI accelerators as China pushes to reduce dependence on Western chips. The company’s Kunlunxin unit developed an inference-optimised chip, the M100, planned for 2026 and offered in rack-scale clustered systems called Tianchi256 (early 2026) and Tianchi512 (late 2026). A more powerful training chip, the M300, is slated for 2027 to target multi-trillion-parameter model training. Baidu also announced ERNIE 5.0, its latest multimodal foundation model.
Key Points
- Baidu revealed two families of custom AI accelerators: the M100 (inference) and the M300 (training).
- The M100 is designed with MoE (mixture-of-experts) inference in mind and will ship in clustered rack-scale systems (Tianchi256 and Tianchi512) to tackle interconnect bottlenecks.
- Tianchi256 (256 M100 chips) is expected in early 2026; Tianchi512 (512 chips) planned for late 2026.
- The M300 training chip targets multi-trillion-parameter model training and is due in 2027.
- Baidu’s hardware push accompanies ERNIE 5.0 and reflects Beijing’s pressure on domestic firms to reduce reliance on Western accelerators such as Nvidia’s.
- Other Chinese vendors — Huawei, Biren, Cambricon, MetaX — are also advancing home-grown GPU and rack-scale solutions, prioritising scale over peak efficiency in some designs.
- Nvidia has reportedly paused active discussions around selling its latest Blackwell accelerators in China, accelerating the domestic shift.
Why should I read this?
Quick version: China’s big players are building their own AI muscle. If you care about AI costs, geopolitics, or how big models will be served and trained, this matters — and Baidu’s moves give a clear signal that the market is splitting into domestic hardware ecosystems. Short, sharp and worth a glance if you want to stay ahead.
Context and Relevance
This announcement sits at the intersection of two major trends: the global race to scale AI infrastructure for ever-larger models, and geopolitically driven hardware sovereignty. Baidu’s rack-scale approach addresses real engineering problems for MoE and other large architectures where latency and interconnect bandwidth kill efficiency. For organisations tracking cloud strategy, supply chains or AI deployment cost curves, Baidu’s timeline (M100 in 2026, M300 in 2027) signals growing viability of non-Western accelerator stacks and further diversification of the AI hardware market.