HAMi Enters CNCF Incubation, Founder Says Still Treading on Thin Ice

Two years ago, Li Mengxuan looked at the booths of mature open source projects at KubeCon in Europe, when HAMi was still a 'small fry' with little community recognition. On July 2, 2026, HAMi passed the CNCF Technical Oversight Committee review and officially entered incubation. At this point, the project had been adopted directly or indirectly by hundreds of enterprises, gathering developers from GPU vendors, cloud providers, financial institutions, and internet companies.

HAMi's starting point was a specific production problem: traditional GPU management allocates whole GPUs, causing significant memory waste for small models and inference tasks. HAMi adds a 'resource manager' between GPU hardware and upper-layer applications, dividing a single GPU into multiple virtual resources allocated on demand, enabling sharing and isolation.

The project was initially initiated by Zhang Xiao and Li Mengxuan from different organizational backgrounds, and gradually formed an engineering system. After entering incubation, the project faces practical challenges in community governance, long-term maintenance, and enterprise-level deployment.

HAMi's market is not empty; NVIDIA, cloud providers, and commercial companies have solutions, but they often bind to their own hardware or platforms. The heterogeneity of GPUs (different generations, different vendors) makes it difficult for a single company to continuously adapt to all products.

HAMi attempts to achieve 'dependency reversal': attracting chip vendors to actively root in the community and maintain device support themselves. According to Zhang Xiao, a significant number of GPU vendors have invested R&D resources to strive for Day 0 or Day 1 support.

Entering CNCF is to prove neutrality, handing over code, trademarks, etc., to foundation governance, reducing the risk for enterprise users of project discontinuation or closed-sourcing. Zhang Xiao compares CNCF to a school, where incubation indicates the project has begun to be used as production infrastructure by enterprises.

One of HAMi's earliest production users was a listed internet finance company, deploying on 3 clusters, 16 nodes, and 128 GPUs, running recommendation systems and OCR tasks, with GPU utilization efficiency more than doubled. Subsequently, cooperation with SF Technology led to the release of a logistics industry white paper.

What truly showed the team the possibility of a commercial closed loop was a large joint-stock bank, whose need was to uniformly manage heterogeneous computing (existing NVIDIA and newly added domestic GPUs) while meeting stability and risk control requirements. The bank, GPU vendors, community, and Melon AI formed a closed loop, with Melon AI providing adaptation, delivery, and production assurance and receiving revenue.

Production needs drove the expansion of the technical roadmap: from NVIDIA GPU virtualization to heterogeneous computing, Kubernetes control plane, multi-cluster management, observability, etc. Zhang Xiao stated that achievements were not set initially but were driven step by step by production needs.

As users increased, relying solely on the open source community became insufficient to bear responsibilities in enterprise production environments, such as stable versions, fault response, and on-site support. Melon AI was established to fill this gap and become the responsible entity.

Melon AI formed a basic boundary: cross-vendor compatibility and general capabilities go into the HAMi community; production assurance, version management, and scenario optimization for specific customers are handled by enterprise products. Zhang Xiao believes customers purchase products that are runnable, upgradable, and accountable.

The two founders focus on technical evolution and commercial reality respectively. Li Mengxuan emphasizes avoiding HAMi becoming a closed technology stack; Zhang Xiao must answer why customers pay. Establishing a company and handing the project to CNCF address sustainability and neutrality issues respectively.

Melon AI positions itself above devices and drivers, using GPU pooling, virtualization, and scheduling to convert whole GPUs into fine-grained resources. Its value ultimately comes down to the economic equation: improving GPU utilization and reducing idle resources.

Li Mengxuan gave a restrained judgment: 'If HAMi is not successful, Melon will definitely not succeed, but after HAMi succeeds, Melon still needs to find a way to support itself.' This reveals the core of the relationship between open source projects and commercial companies: project success is a necessary but not sufficient condition.

After entering incubation, HAMi faces long-term challenges of continuous maintenance, community governance, and enterprise adoption. Melon AI needs to find a balance between open source contributions and commercial revenue to ensure project sustainability.