360 and Huawei Ascend AI Unveil Joint Agent Infrastructure at Huawei Connect 2026
The AMW Read
Incremental update pairing a known CN agent-orchestration player with domestic Ascend silicon for inference-side tuning; sub-segment relevant to China's agent-compute stack but no debate resolution or new top-tier entrant.
360 and Huawei Ascend AI Unveil Joint Agent Infrastructure at Huawei Connect 2026
At Huawei Connect 2026 in Shanghai on September 17, Chinese security-software company 360 (Qihoo 360, 360集团) showed a joint solution with Huawei's Ascend AI computing platform, centered on 360's Agent Factory — a natural-language tool for generating and orchestrating AI agents, including a "swarm" multi-agent technique that lets agents nest, form teams, and share task memory. The two companies said Ascend AI's tuning for long-sequence inference, throughput, and KV-cache handling lowered first-token latency and raised throughput for agent workloads running DeepSeek, Qwen, and GLM models. In internal testing, Agent Factory executed more than 1,000 sequential task steps without interruption at a reported success rate above 95%, and cut a two-hour multi-role collaboration task to 20 minutes.
The showcase extends a partnership the two firms first demonstrated publicly at the World AI Conference earlier in 2026, and it is a concrete data point in China's effort to run agentic workloads — not just model training — on domestic Ascend silicon rather than Nvidia GPUs. As agent use shifts from single-turn chat to long-horizon, tool-heavy, multi-agent execution, the bottleneck moves from model capability to the serving stack's ability to sustain low-latency, high-throughput inference across many concurrent agent calls. Pairing an agent-orchestration layer directly with chip-level inference tuning is one way China's AI stack is trying to close that gap without depending on export-controlled hardware.
For builders and investors tracking China's AI compute stack, the more durable signal is Ascend's inference-side readiness for agentic workloads on the open models 360 is already running — DeepSeek, Qwen, and GLM — rather than 360's own agent product. That strengthens the case for CN enterprises standardizing agent deployments on Huawei's stack instead of imported GPUs, though the reported 1,000-step, 95%-success benchmark comes from the vendors themselves and has not been independently verified.
