Huawei's ambitious 3-year AI roadmap: Chips, HarmonyOS, and self-reliance
At its Connect 2025 event in Shanghai this September, Huawei rolled out a bold three-year roadmap to build next-gen AI systems.
The plan centers on combining its new Ascend AI chips with ARM-based Kunpeng processors, and debuting tech like the CloudMatrix384 supernode and UnifiedBus protocol to power clusters of up to a million chips.
Why the shift to homegrown tech?
With US export bans making NVIDIA GPUs tough to get, Huawei is doubling down on homegrown tech.
By integrating its own CPUs and HarmonyOS across devices, it's aiming for self-reliance—so China's AI future isn't tied to foreign suppliers.
CloudMatrix384 supernode connects 384 chips for massive performance boost
The CloudMatrix384 supernode connects 384 Ascend chips and delivers up to four times better performance than earlier setups.
Plus, the UnifiedBus 2.0 protocol promises chip-to-chip speeds that are an incredible 62 times faster than NVIDIA's NVLink—making huge "SuperClusters" possible.
Ascend 970 chip on the horizon
Huawei also teased the Ascend 970 chip for release by 2028, targeting blazing-fast interconnect speeds of four terabits per second thanks to its telecom know-how.
It's all about scaling up AI—even with some manufacturing hurdles in the way.