Huawei Cloud Rebuilds Infrastructure for the Agent Era

🔑 Enhanced Key Takeaways

•Huawei Cloud's 'Agentic Infra' introduces specific products like the AI Cluster Service (AICS), built on an ultra-high bandwidth UnifiedBus (UB) network, supporting clusters with over 100,000 cards and delivering up to 200 EFLOPS, with token generation latency reduced to less than 10 milliseconds.
•The infrastructure includes Agentic Memory Storage (AMS), which leverages NPU passthrough to Context Memory Storage (CMS) hardware, creating PB-scale memory space and supporting tiered KV-cache pooling for efficient AI agent operations.
•Huawei Cloud's strategy is underpinned by CloudMatrix384 supernodes connected via a MatrixLink network, creating a hybrid system that combines general-purpose and intelligent compute, specifically targeting Mixture of Experts (MoE) models for improved inference speed.
•The Pangu Models 5.5, central to this agentic shift, feature a 718-billion parameter deep thinking model utilizing a Mixture of Experts (MoE) architecture with 256 specialists, enabling adaptive fast and slow thinking integration for an eightfold improvement in overall model inference efficiency.
•Huawei Cloud has launched the Versatile platform, an enterprise-grade agent platform designed to integrate with existing business systems and cover the full application lifecycle from development to deployment, usage, and management, addressing the specific workflow demands of enterprise AI agents.

📊 Competitor Analysis▸ Show

While the article focuses on Huawei Cloud, the broader market for AI agent infrastructure includes major hyperscalers. Here's a comparison:

Feature/Platform	AWS Bedrock Agents	Google Cloud (ADK/Vertex AI Agent Builder)	Azure AI Foundry Agent Service	Huawei Cloud (Agentic Infra/Versatile)
Approach	Managed service, configured agents with built-in guardrails and orchestration.	Open-source Agent Development Kit (ADK) for building, Vertex AI Agent Builder for managed runtime.	Aligns Semantic Kernel and AutoGen under a shared framework, integrates with Microsoft 365.	Supernode architecture (CloudMatrix384), industry-specific model training, Versatile platform for enterprise deployment.
Integration	Deeply integrated with AWS services (Lambda, S3, DynamoDB, Step Functions).	Integrates with BigQuery, Cloud Run, Dataproc; native RAG with Google Search grounding.	Strong enterprise integration with Microsoft ecosystem (SharePoint, Fabric).	Integrates with Huawei Cloud's AI compute, models, data platforms, and tools; focuses on industry-specific solutions.
Key Strengths	Operational simplicity, handles orchestration, memory, safety; extensive ML toolkit (SageMaker).	Open-source flexibility, optimized for Gemini, comprehensive AI ecosystem, cost-efficient Vector Search.	Broad strategy spanning SDKs, low-code platforms, end-user products; strong for Windows-based environments.	Full-stack hardware-software synergy, high-performance supernodes (CloudMatrix384), memory-centric AI-Native Storage, Pangu Models 5.5 with MoE.
Pricing Model	Pay per model token and per tool invocation; no separate charge for orchestration layer.	Pay-as-you-go; charges for Agent Engine Runtime, stored sessions, code execution.	Not explicitly detailed, but generally pay-as-you-go with flexible scaling.	Not explicitly detailed in search results, but focuses on enterprise-grade solutions.
Target Audience	Teams already invested in AWS needing managed orchestration with enterprise compliance.	Teams on GCP wanting managed infrastructure with visual agent building tools or those building agents needing cross-organizational interoperability.	Teams in Microsoft ecosystems.	Enterprises seeking industry-specific AI solutions, particularly in manufacturing, agriculture, scientific research, and smart driving.

🛠️ Technical Deep Dive

CloudMatrix384 Supernodes: Huawei Cloud's AI Compute Service utilizes CloudMatrix384 supernodes connected via a MatrixLink network. This architecture creates a hybrid system combining general-purpose and intelligent compute resources.
Mixture of Experts (MoE) Optimization: The supernode structure is specifically designed for Mixture of Experts (MoE) models, facilitating expert parallelism inference to reduce NPU idle time during data transfers. Huawei reports single-PU inference speed increases of four to five times compared to other models.
AI-Native Storage: Paired with the supernodes is a memory-centric AI-Native Storage system, optimized for the access patterns typical in AI training and inference workloads.
Pangu Models 5.5: The latest iteration of Huawei's foundation models, Pangu Models 5.5, includes a 718-billion parameter deep thinking model. This model employs a MoE architecture with 256 experts and integrates adaptive fast and slow thinking, which can improve overall model inference efficiency by eight times.
AI Cluster Service (AICS): AICS is built on the ultra-high bandwidth UnifiedBus (UB) network, supporting clusters with over 100,000 cards and delivering a total computing power of up to 200 EFLOPS. It achieves a token generation latency of less than 10 milliseconds and a throughput of 5 million tokens per second across 1,000 cards, with 99.95% online service availability.
Agentic Memory Storage (AMS): This solution leverages NPU passthrough to Context Memory Storage (CMS) hardware, providing a PB-scale memory space. It also supports tiered KV-cache pooling for efficient memory management in agent operations.
Agent Development Platform: Huawei Cloud offers an agent development platform that supports zero-code/low-code hybrid development. It allows users to orchestrate workflows using a visualized GUI, integrating nodes for LLMs, knowledge retrieval, and plug-ins, and provides nearly 100 standard components for tasks like natural language processing and code execution.
Confidential Computing: Huawei Cloud provides an AI confidential computing solution with capabilities such as confidential virtual machines (VMs), remote attestation, confidential computing key management, and confidential inference gateway, supporting confidential inference, pre-training, and federated learning.

🔮 Future ImplicationsAI analysis grounded in cited sources

Huawei Cloud's focus on 'Agentic Infra' will accelerate the adoption of autonomous AI agents in enterprise settings.

By providing specialized hardware (supernodes, AI-Native Storage) and software platforms (Versatile, AgentArts) designed for the unique computational and workflow demands of agents, Huawei Cloud aims to remove current deployment bottlenecks for businesses.

The emphasis on industry-specific AI Foundries will lead to more tailored and effective AI solutions for vertical markets.

Huawei Cloud's strategy includes dedicated zones for Smart Healthcare, Embodied AI, Smart Manufacturing, and Scientific Computing, indicating a move towards highly customized AI applications that address specific industry challenges.

Huawei Cloud's full-stack approach, from chips to cloud services, will enhance its competitive position in the global AI market.

By integrating its own AI chips (Ascend), network technologies (MatrixLink, UnifiedBus), and cloud platforms, Huawei aims to offer a highly optimized and efficient ecosystem for AI development and deployment, differentiating itself from competitors.

⏳ Timeline

2018-10

Huawei launched its Full-Stack, All-Scenario AI Solution at HUAWEI CONNECT, marking its formal introduction to the AI scene.

2021-07

Huawei officially launched the Pangu series of large AI models.

2023-07

Huawei introduced Pangu 3.0, a large language model tailored for sectors like government, finance, manufacturing, mining, and meteorology.

2024-06

Huawei announced upgraded Pangu 5.0 at HDC 2024, integrating with Harmony Intelligence.

2025-03

Huawei Cloud unveiled new cloud services and solutions at MWC 2025, aiming to expedite the evolution from cloud-native to AI-native.

2025-06

Huawei Cloud introduced Pangu Models 5.5 at HDC 2025, featuring significant upgrades across NLP, CV, multi-modal, prediction, and scientific computing, with enhanced agent capabilities.

2025-10

Huawei Cloud revealed its enterprise-grade agentic AI strategy at Huawei Connect 2025, introducing CloudMatrix384 infrastructure and the Versatile platform.

2026-03

Huawei unveiled the AgenticCore solution at MWC Barcelona 2026, designed for AI-centric networks in the agent era.

2026-06

Huawei Cloud officially introduced the new paradigm of Agentic Infra and unveiled a series of Agentic AI products at Huawei Cloud INSPIRE 2026.

Huawei Cloud Rebuilds Infrastructure for the Agent Era

⚡ 30-Second TL;DR

Key Points

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (22)

👉Related Updates

Build agentic AI semantic layers with Stardog and Bedrock

Xu Jinbo's team launches MoleculeOS for AI drug discovery