Huawei
2026-07-10
Huawei Ascend 10K-Card Cluster Goes Live, UnifiedBus Protocol Pools All Resources
Summary
Key Takeaways
Huawei officially launched the Ascend 10,000-card AI computing cluster in the Guangdong-Hong Kong-Macao Greater Bay Area (Shaoguan), deploying over 10,000 Ascend AI accelerators to support domestic large model development. At WAIC 2026, Huawei will showcase the Atlas 950 SuperPoD with its proprietary UnifiedBus interconnect, supporting up to 8,192 NPUs with an aggregate bandwidth of 16.3 PB/s, targeting trillion-parameter model training and inference.
Huawei Cloud was named a Leader in the Gartner 2026 Cloud AI Infrastructure Magic Quadrant, highlighting its vertical integration capabilities: UnifiedBus enables unified pooling of CPU, NPU, memory, storage, and DPU resources; the new AICS supports thousand-node cloud supernodes and hundred-thousand-node clusters; Elastic Memory Service (EMS) achieves 95% hit rate for long-context reads.
Huawei also launched an AI DreamWorks industry platform with dedicated zones for healthcare and embodied AI. The Ascend-aware AI container cluster (CCE Volcano) leverages topology-aware scheduling and distributed inference acceleration. Huawei VP Wang Lei predicted China's AI computing industry to exceed 30 trillion yuan by 2030, noting current token circulation efficiency is only 60% and needs to reach 90%+. Huawei is collaborating with CAICT and three major telecom operators to promote 'network-as-compute' architecture changes. Partnerships were signed with Chengdu entities for computing-electricity infrastructure and AI applications.
Why It Matters
Huawei's proprietary UnifiedBus and Ascend ecosystem aim to build a fully closed stack to defend against NVIDIA's CUDA ecosystem and encircle domestic AI chip startups (e.g., Cambricon, Biren). Adoption locks users into CANN and MindSpore, creating high migration costs.
Hidden lock-in: UnifiedBus pooling centralizes control in AICS, preventing users from choosing standard RoCEv2 or InfiniBand components, forcing full Huawei hardware procurement.
Physical limitations: The 16.3 PB/s is theoretical aggregate bandwidth; actual effective bandwidth may suffer from congestion control and topology constraints. Atlas 950 SuperPoD has extreme power and cooling demands, raising deployment costs. Huawei lacks direct performance comparisons with NVIDIA H100/B200, so enterprises must scrutinize actual training throughput and ecosystem compatibility.
PRO Decision
[Vendors] Competitors (NVIDIA, AMD, Cambricon) should highlight Huawei's performance uncertainty and ecosystem lock-in risks. NVIDIA can promote its mature CUDA ecosystem, open InfiniBand networks, and offer migration tools to attract potential Huawei customers. Domestic chip makers should accelerate open interconnect standards (e.g., OCP) and software stacks to avoid being encircled.
[Enterprises] CIOs and architects must conduct zero-trust technical audits: demand independent benchmarks comparing actual training throughput, energy efficiency, and linear scaling of Huawei's cluster vs. NVIDIA H100/B200 on models like GPT-3. Evaluate TCO including power, cooling, and network upgrades. Ensure contracts include cross-platform portability clauses to avoid lock-in to MindSpore and CANN. Consider multi-cloud strategies retaining standard RoCEv2 or InfiniBand options.
[Investors] See through Huawei's PR: the Ascend cluster may win government contracts short-term but faces growth limits due to ecosystem closure. Monitor Huawei Cloud's actual market share and profitability. Compare the open ecosystem progress of other domestic AI chip makers (e.g., Cambricon) to assess Huawei's vertical integration sustainability. Beware of supply chain concentration risk.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)