How Huawei Cloud Versatile Tackles Enterprise AI Agents

Huawei Cloud has developed infrastructure and tooling to address the computational and workflow requirements of enterprise AI agents. The company’s approach combines supernode architecture, industry-specific model training and an agent platform that integrates with existing business systems.

Zhang Yuxin, CTO of Huawei Cloud, outlined the shift this creates in computing architecture at Huawei Connect 2025. “Unlike traditional systems, with their fixed processes and resources, agentic AI makes decisions independently,” Zhang says. “It adapts dynamically, reshaping how computing systems interact and allocate resources.”

This evolution has prompted Huawei Cloud to develop a strategy spanning infrastructure, foundation models, tools and agent platforms. The framework addresses both the computational demands of agentic AI and the practical requirements of businesses implementing these systems.

Huawei Cloud CloudMatrix384 supernodes address compute demands

The computational requirements for foundation model training and inference have exposed limitations in existing architectures. Huawei Cloud’s AI Compute Service uses CloudMatrix384 supernodes connected through a MatrixLink network. The architecture creates a hybrid system that combines general-purpose and intelligent compute.

The supernode structure targets Mixture of Experts models specifically. By enabling expert parallelism inference, the design reduces NPU idle time during data transfers. Huawei reports single-PU inference speed increases of four to five times compared to other models in the market. The company has paired this with memory-centric AI-Native Storage designed for the access patterns typical in training and inference workloads.

ModelBest has integrated its MiniCPM series with this infrastructure. When running on Huawei’s infrastructure, training energy efficiency improved by 20%, with performance exceeding industry standards by 10%. Deployments now include cars, smartphones, embodied AI systems and AIPCs.

Huawei Cloud industry models use incremental training workflows

Creating models for specific industries requires more than access to compute. Huawei Cloud has developed processes for data preparation, incremental training and evaluation that help companies build models suited to their domains. The incremental training workflow adjusts parameters based on core model characteristics and industry objectives. Huawei states this approach boosts model performance by 20% to 30%.

Shaanxi Cultural Industry Investment Group has applied this approach to cultural tourism. Working with Huawei, the organisation combined datasets spanning history, film, and intangible heritage. Huang Yong, Chairman of Shaanxi Cultural Industry Investment Group, described the scope of the work. “Using Huawei Cloud’s data-AI convergence platform, the group combined diverse cultural tourism data to create comprehensive datasets across areas like history, film and intangible heritage, strengthening Shaanxi’s cultural tourism foundation,” Huang says.

The partnership established a data space for cultural tourism on Huawei Cloud. The collaboration produced the Boguan cultural tourism model, which powers tools including a cultural tourism intelligent brain, smart management assistant, intelligent travel assistant and a short video platform.

Huawei Cloud Versatile platform enables enterprise agent deployment

The shift from personal user agents to enterprise systems introduces different requirements. Enterprise agents must integrate with existing workflows, handle situations that span multiple systems, and meet standards for reliability and auditability. Huawei Cloud developed Versatile to address these needs.

The platform covers the application cycle from development through deployment, release, usage and management. It integrates with Huawei Cloud’s AI compute, models, data platforms, tools and ecosystem capabilities.

Conch Group used this approach to create agents for the cement industry. Working with Huawei, the company built models for production, operations and management. The system predicts clinker strength at three and 28 days with deviations of less than 1 MPa from actual results, achieving over 90% accuracy. For cement calcination, the model generates process parameters under varying conditions and identifies operational solutions that reduce standard coal usage by 1% compared to class A energy efficiency standards.

Xu Yue, Assistant to Conch Cement's General Manager, discussed the implications for the industry. “The model’s success with quality control, production optimisation, equipment management and safety lays the groundwork for end-to-end collaboration and decision-making for cement agents,” Xu says. “These advanced agents are moving the cement industry from relying on traditional expertise to being fully driven by data across all processes.”