Globalization 2.0 | Macau ICT Company Builds AI Private Cloud with ZStack Cloud

The ICT company in Macau, China, provides end-to-end AI implementation services for large enterprises and public institutions. This year, the concept of “AI Empowering the Future” has heated up. Various industries in Macau, China, have been keen to try model fine-tuning and vertical AI scenarios. However, due to the scarcity and high cost of local AI computing resources, low resource allocation efficiency, and difficulties in data privacy and security isolation, they face a computing power bottleneck of “available but not accessible”: a multi-card A100 GPU server is expensive and difficult to deploy in bulk, forcing users to queue for bare-metal, leading to scheduling conflicts; different models have varying demands for GPU quantity, memory, and VRAM, resulting in low resource configuration efficiency; and training on the same machine poses data privacy isolation challenges.

Therefore, the company deployed ZStack Cloud on its two A100 GPU servers, virtualizing physical GPUs into multiple independent computing units, achieving unified scheduling and multi-tenant isolation, and truly enabling “one server, multi-party reuse,” helping users to conduct AI model training research in parallel and securely.

Construction Plan

The ICT company in Macau, China, deployed ZStack Cloud on two servers equipped with multi-card A100 GPUs. By virtualizing GPU computing power, pooling computing resources, and integrating local storage, they built an efficient, flexible, and secure AI cloud infrastructure platform. The overall architecture includes:

Physical Resource Layer: Through the vGPU scheduling capability of the ZStack Cloud platform, physical GPUs are divided into multiple virtual GPU instances as needed, supporting elastic allocation of VRAM and computing power. For high-performance scenarios, full-card pass-through mode is supported to meet differentiated computing power needs and avoid the problem of “heavy tasks monopolizing resources,” allowing multiple users to share AI servers and improve GPU utilization.

Virtualization Layer: GPU resource passthrough and general computing resource CPU, memory virtualization allocation are completed through ZStack Cloud.

Storage Layer: Uses high-performance disk resources including local NVMe SSD (LocalStorage) and Fiber Channel Storage Network (FC-SAN) to meet the IO throughput requirements of AI training.

Security Isolation: Through ZStack Cloud’s tenant isolation mechanism, different users’ AI environments are deployed in isolation, ensuring data confidentiality, resource non-conflict, and permission clarity during the AI model training process.

Unified Management: ZStack Cloud centrally manages GPU resources and user permissions to avoid resource conflicts.

Customer Benefits

AI resource efficiency increased by 3 times, enabling parallel research across multiple departments:

GPU resource utilization increased by over 200%: The same physical GPU server can support multiple users for AI model training simultaneously, solving the queuing bottleneck;

Deployment cycle reduced by 50%: Users do not need to wait for resource scheduling, allowing rapid initiation of AI application experiments;

Privacy and security enhancement: The platform’s virtual machine-level isolation mechanism ensures data confidentiality for each department, supporting independent network and user permission configurations;

Significant TCO optimization: Without additional hardware investment, AI computing power is shared among multiple users, saving customers millions in GPU server procurement budgets.

The successful launch of this project provides a cost-effective, highly elastic, and highly isolated technical foundation for local AI development and research, and also validates ZStack Cloud’s strong adaptability in AI virtualization deployment scenarios. In the future, ZStack will continue to work with local partners to build a more inclusive, agile, and secure AI infrastructure platform, injecting cloud momentum into the development of the AI ecosystem in Macau, China.