CNCF Publishes Official HAMi Case Study With China Merchants Bank, Validating Open-Source GPU Orchestration in Finance

The Cloud Native Computing Foundation (CNCF) has officially published a case study examining how China Merchants Bank (CMB) built its AI compute scheduling platform on HAMi, an open-source GPU orchestration technology originating from China. This marks the first CNCF-endorsed benchmark project since HAMi graduated from Sandbox to Incubation status.

CMB deployed HAMi on top of Kubernetes to create a unified heterogeneous AI compute platform featuring multi-vendor pooling, elastic sharing, and topology-aware optimization. Production data reveals the platform achieved 100% hardware pool utilization, reduced cross-node scheduling probability for distributed training by 30%, and effectively solved long-standing problems including compute silos, resource idling, and high operational costs.

HAMi originated from internal vGPU technology at Paradigm, a full-stack AI cloud services company. Facing fragmented heterogeneous GPU devices and inefficient resource scheduling, Paradigm open-sourced the technology to drive standardization. To date, HAMi has attracted over 500 contributors from 17 countries and serves more than 300 enterprises across cloud services, internet, finance, automotive, logistics, and education, with deployments extending into Southeast Asia and Europe.

Beyond CMB, HAMi has been deployed at NIO, Beike, SF Tech, SNOW Corp., DaoCloud, and PREP EDU for GPU utilization improvement, heterogeneous compute pooling, large-model inference traffic scheduling, and autonomous driving training and simulation. Enterprises report up to 80% savings on hardware procurement costs and 5-10x improvement in GPU utilization through heterogeneous accelerator sharing and isolation scheduling.

Earlier this year, HAMi was featured in the keynote session at KubeCon+CloudNativeCon Europe in Amsterdam, making it one of the few Chinese open-source projects to receive that speaking slot, signaling growing global interest in China-originated AI compute orchestration.

As the project’s primary contributor, Paradigm has pushed key engineering milestones including production-grade Kubernetes DRA (Dynamic Resource Allocation) support and hardware adapters for NVIDIA, Huawei Ascend, and Kunlun chips.

Since joining CNCF in August 2024, HAMi has advanced rapidly from Sandbox to Incubating status within two years. Industry observers view this progression as validation that the project meets international open-source standards in governance, security, adoption breadth, and maturity, positioning China-born infrastructure technology as a potential de facto standard in global AI compute orchestration.