A South Korean industry association estimate puts China about three years behind in HBM, two in DRAM and one in NAND.
Huawei envisions 900 billion AI agents by 2035, with global annual token consumption rising 100,000 fold and agents generating more than 90% of AI token traffic.
APXInf is an open source inference engine developed by Infinigence AI with Tsinghua University and Shanghai Jiao Tong University to run embodied AI models near a robot’s sensors and actuators.
PhysBrain 1.5 uses one autoregressive model to produce spatial answers, robot motion and predicted future states.
Huawei unveiled the liquid cooled Atlas 960E SuperPoD, claiming it can connect up to 4,096 NPUs for AI workloads involving 10 trillion parameter models.
HERON CRA combines task history memory, corrective reinforcement learning and cross robot pretraining, with a proposed learning loop using imagined rollouts.
LightNav 0 is an open sourced robot navigation model built on Qwen3 VL 4B. Light Origins says training on more than 2,000 reconstructed scenes and 4,000 hours of simulated experience helps one policy transfe...
LimiX 2 is a 400 million parameter structured data model released with weights and inference code on September 16, 2026.
Chinese brands rose from under 29% of domestic industrial robot market share in 2020 to about 55% of shipments in 2025.
At Huawei Connect 2026, Huawei positioned Lingqu UnifiedBus as the shared interconnect for AI compute and memory across SuperPoDs and larger clusters.
Huawei says OceanStor M900 extends SuperPoD KV cache into a shared SSD backed tier with up to 64 PB per cluster.
ByteDance’s September 16 update made Doubao-Seed-2.1-pro 0915 available through Volcano Ark and emphasized production tasks that combine visual input, code, tools, and long-running agents. The demonstrations...