The brief

Analyst view

• Operates in 170+ countries across cloud, AI, telecoms, consumer devices and transport systems.
• Ascend roadmap: 950/960/970 series at one-year cycles, each doubling compute. 950PR Q1 2026, 950DT Q4 with proprietary HBM.
• Atlas 950 SuperPoD (8,192 NPUs) delivers triple H20 inference performance at quarter the cost, shown at MWC 2026.

Technical approach

As reported

Ascend NPUs

Custom NPU architecture for AI workloads

AI accelerators: PR (prefill/recommendation) and DT (decode/training) variants. Compete with NVIDIA GPUs.

HBM Technology

In-house high-bandwidth memory

Proprietary HBM reduces cost and supply-chain risk. HiBL 1.0 on 950PR (128 GB, 1.6 TB/s), HiZQ 2.0 on 950DT (144 GB, 4 TB/s).

Scale Integration

Optical interconnect at cluster scale

4,096 optical transceivers (800G each) connect 8,192 NPUs with 3-microsecond latency, enabling unified system-scale training.