Can the Heterogeneous Computing Power Scheduling Platform Become a New Growth Driver in the AI Infrastructure Track? ——In-depth Analysis of Migua Intelligence's HAMi Open Source Ecosystem Model

Your question involves two key dimensions:

1. Technical Dimension
   : Can the heterogeneous computing power scheduling platform become a core growth driver in the AI infrastructure track?
2. Business Dimension
   : Can HAMi’s open source ecosystem model replicate VMware’s successful path in the CPU virtualization era?

According to the latest market data, China’s AI infrastructure platform market is experiencing

:

• Market Size
  : In 2024, China’s AI Infra platform market size reached
  3.45 billion yuan
  , and it is expected to jump to
  6.73 billion yuan
  in 2025, with a year-on-year growth rate of
  95.1%
  [2]
• Computing Power Management Layer
  : Accounts for 64.6% of the AI Infra platform market, being the main component
• Policy Driver
  : Domestic substitution accelerates, GPU virtualization and heterogeneous computing power scheduling become rigid demands[3]

:

• Moore Threads, Muxi, Cambricon, etc., have been listed or passed the Hong Kong Stock Exchange hearing
• The localization rate of China’s general-purpose GPU market increased from 2% in 2022 to 3.6% in 2024, and is expected to exceed
  50%
  by 2029[4]
• However, the market share of each manufacturer is small (Muxi about 1%, Moore Threads and Biren Technology less than 1%), and the market is highly fragmented[4]

:

1. Severe fragmentation of heterogeneous chip ecosystem
   :
   • There are many domestic GPU manufacturers (Huawei Ascend, Muxi, Moore Threads, Cambricon, etc.), but the software stacks are not unified
   • Different chips have huge differences in architecture design and software stacks, significantly increasing the complexity of unified training and scheduling[5]
   • Leading enterprises like SenseTime and Baidu Smart Cloud need to adapt to multiple chips simultaneously, leading to high system complexity

2. Low GPU resource utilization
   :
   • Traditional Kubernetes Device Plugin only supports
     static GPU allocation
     , which monopolizes the entire card once assigned to a Pod[6]
   • Small-scale inference tasks may only require partial memory or computing power, but monopolizing the entire GPU leads to resource waste
   • Industry data shows that after adopting HAMi, GPU utilization can increase from 60% to over 90%[1]

3. Huge cost pressure
   :
   • Using the special chip H20 to complete the same scale of AI training requires
     40%-60% more computing time
     and
     over 35% higher power cost
     compared to H100[3]
   • The Frost & Sullivan report points out that China’s GPU cloud market has entered a new stage of “usable, easy to use, and sustainable”, and the competition focus has shifted from single hardware performance to
     full-stack capability
     [3]

According to industry report analysis:

:

• 2024: 3.45 billion yuan
• 2025: 6.73 billion yuan (95.1% growth)
• The computing power management layer accounts for 64.6%, which is about
  4.35 billion yuan
  in 2025[2]

:

• Frost & Sullivan report shows that China’s self-developed GPU cloud market has entered the “10,000-card era”[3]
• The global GPU cloud computing market is expected to reach
  47.24 billion US dollars
  by 2033, with a CAGR of
  35%
  from 2025 to 2033[3]

, meaning the domestic market space can reach in the next 3-5 years.

:

Type	Representative Enterprises	Core Products	Advantages	Disadvantages
Cloud Vendor Self-developed	Alibaba Cloud cGPU, Tencent Cloud qGPU	Internal solutions of cloud vendors	Deep integration with cloud services, deep technical accumulation	Mainly serve own cloud customers, weak cross-cloud support
Professional AI Infra Vendors	SenseTime Large Device, Baidu Baige	End-to-end AI computing power platform	Full-stack capability, scenario-based solutions	Closed ecosystem, limited chip adaptation
Open Source Project	HAMi	Heterogeneous computing power virtualization middleware	Open source ecosystem, cross-chip compatibility, community-driven	Commercialization path not yet verified
Virtualization Vendors	QingCloud ZStack AIOS	Full-stack AI Infra platform	Enterprise-level virtualization experience	Focus on private cloud, weak cross-cloud capability

:

• CNCF Certification
  : The world’s only CNCF project focusing on heterogeneous computing power virtualization[1]
• Cross-chip Compatibility
  : Supports more than 9 chips including NVIDIA, Huawei Ascend, Muxi, Moore Threads, Cambricon
• Cloud-native Architecture
  : Based on Kubernetes, naturally compatible with the cloud ecosystem
• Team Background
  : The founding team comes from DaoCloud and 4Paradigm, and are core contributors to Kubernetes

:

1. Technical Leadership
   :
   • Solved the hardware virtualization problem of x86 architecture in the early stage
   • ESXi hypervisor performance is close to native, laying the foundation for enterprise adoption

2. Business Model
   :
   • Enterprise-level subscription services: Product combinations like vSphere and vCenter
   • Tiered pricing: Gradient pricing from free version to enterprise-level functions
   • Service ecosystem: Certified training, partner network

3. Ecosystem Construction
   :
   • Deep integration with hardware manufacturers (servers, storage, networks)
   • Established a huge technical partner network

[8]:

:

1. Open Source Core
   : HAMi’s core functions are provided for free in open source
2. Enterprise Value-added Services
   :
   • Enterprise-level technical support
   • Advanced function modules (such as deep monitoring, automated operation and maintenance)
   • Private deployment and customized services
   • Training and consulting services

:

• Use community power to accelerate product development and innovation
• Reduce market promotion costs; open source itself is a powerful marketing tool
• Improve product transparency and credibility, facilitating security audits
• Avoid vendor lock-in, enhance user trust
• Cultivate developer ecosystem, expand potential user base[8]

:

• Uncertain commercialization path, need to balance open source and commercial interests
• Difficult to divide the boundary between core functions and value-added functions
• Face competition from cloud service providers, may be “hijacked”
• Complex open source community management, requiring professional teams to maintain[8]

:

Dimension	VMware	HAMi	Similarity
Technical Positioning	CPU Virtualization Layer	GPU Virtualization Layer	⭐⭐⭐⭐⭐
Architecture Position	Infrastructure Layer	Infrastructure Layer	⭐⭐⭐⭐⭐
Business Model	Enterprise-level Subscription	Open Core + Enterprise Services	⭐⭐⭐
Ecosystem Dependence	Hardware/Software Ecosystem	Chip/Cloud Ecosystem	⭐⭐⭐⭐

:

1. Changes in Technical Environment
   :
   • VMware Era: Enterprise data centers were mainstream, private deployment was the mainstay
   • HAMi Era: Public cloud, hybrid cloud, and multi-cloud have become mainstream, making cross-cloud scheduling more complex

2. Different Competitive Landscape
   :
   • VMware Era: Few competitors, high technical threshold
   • HAMi Era: Cloud vendors’ self-developed solutions are strong (Alibaba cGPU, Tencent qGPU), and competition is more intense

3. Maturity of Open Source Ecosystem
   :
   • VMware Era: Open source ecosystem was not mature
   • HAMi Era: CNCF/K8s ecosystem is mature, and open source has become the mainstream development model[9]

4. Different Market Rhythms
   :
   • VMware: It took more than 10 years to reach mature commercialization
   • HAMi: AI development is extremely fast, requiring faster commercial verification

.

:

1. Clear Technical Rigid Demand
   :
   • Heterogeneous computing power scheduling is a bottleneck problem for AI industry implementation
   • Strong customer pain points: Low resource utilization, high cost, complex management
   • Leading enterprises like SenseTime and Baidu have verified market demand[5]

2. Open Source Ecosystem Advantages
   :
   • CNCF certification provides community trust endorsement
   • Kubernetes ecosystem is mature, easy to integrate
   • Open source model can quickly accumulate users and developer communities

3. Driven by Domestic Substitution Policy
   :
   • GPU domestic substitution accelerates, requiring a unified scheduling platform
   • Key industries like finance and government have high requirements for independent controllability
   • Policy orientation supports domestic AI infrastructure construction

4. Team Gene Matching
   :
   • DaoCloud background: Deep accumulation of cloud-native technology
   • 4Paradigm background: AI enterprise service experience
   • Kubernetes core contributors: Technical authority

:

• Obtained
  2 million yuan product order contracts
  within one quarter of establishment (customers include SF Technology, Vietnam PREP EDU)
• Completed angel round financing of tens of millions of yuan, led by Fosun Capital[1]

:

1. Complexity of Cross-chip Adaptation
   :
   • Large differences in GPU architectures (NVIDIA CUDA vs. Huawei CANN vs. domestic chips)
   • High difficulty in performance optimization and stability assurance
   • Need to continuously adapt to new chips (such as Muxi, Moore Threads, Iluvatar CoreX)

2. Performance Overhead Problem
   :
   • GPU virtualization itself brings certain performance loss
   • Time slice scheduling has a greater impact on performance
   • Need to balance isolation and performance

3. Cluster Scale Challenge
   :
   • The scheduling complexity increases exponentially from single cluster to 10,000-card cluster
   • The stability of cross-domain training is a key variable[5]
   • Need to support large-scale commercial deployment

:

1. Competition from Cloud Vendors
   :
   • Alibaba Cloud cGPU and Tencent Cloud qGPU are mature and deeply integrated with cloud services
   • Cloud vendors may choose self-development instead of third-party solutions
   • HAMi needs to position itself as a “multi-cloud neutral” scheduling platform

2. Open Source Monetization Path
   :
   • How to draw the boundary between open source core and enterprise version functions?
   • How to avoid cloud vendors “free-riding” open source code?
   • Can the subscription model support continuous R&D investment?

3. Long Ecosystem Construction Cycle
   :
   • Need to establish deep cooperation with various chip manufacturers (such as completing compatibility certification with Muxi GPU[10])
   • Need to cultivate developer communities and technical partners
   • Need to establish industry benchmark customer cases

:

1. High Customer Education Cost
   :
   • Heterogeneous computing power scheduling is a new concept, requiring market education
   • Customers may choose to wait for cloud vendors’ solutions instead of independent procurement

2. Maturity of Domestic GPUs
   :
   • Domestic GPU product performance has not yet拉开明显差距[4]
   • Small market share, immature ecosystem
   • Dependent on the rise of domestic GPUs

3. Profit Pressure
   :
   • Angel round financing of tens of millions of yuan, but huge R&D and ecosystem investment
   • Need to achieve product-market fit before funds are exhausted
   • The initial 2 million yuan order proves demand, but needs to be scaled

Based on analysis, HAMi needs to:

:

1. Quickly launch stable enterprise version products
2. Establish industry benchmark customer cases (finance, government, large model companies)
3. Improve core chip adaptation (NVIDIA, Huawei Ascend, Muxi)
4. Achieve ten-million-level ARR (Annual Recurring Revenue)

:

1. Build developer community and partner ecosystem
2. Support more than 10 chips, covering mainstream domestic GPUs
3. Realize unified scheduling across clouds and clusters
4. Reach 100 million-level commercial revenue

:

1. Become the de facto standard for heterogeneous computing power scheduling
2. Establish an open source ecosystem similar to CNCF
3. Explore internationalization (such as cooperation with Vietnam PREP EDU[1])
4. Consider IPO or acquisition by strategic investors

:

• ✅ Huge market space and rapid growth (AI Infra platform is expected to grow by 95.1% in 2025, and the computing power management layer accounts for 64.6%[2])
• ✅ Clear technical rigid demand, strong pain points
• ✅ Open source ecosystem model has been verified by Red Hat
• ✅ Excellent team background, Kubernetes core contributors
• ✅ Initial customer verification, strong technical feasibility
• ✅ CNCF certification provides authoritative endorsement

:

• ⚠️ Fierce competition from cloud vendors, may be marginalized
• ⚠️ Risk of domestic GPU development falling short of expectations
• ⚠️ Open source monetization path not yet verified
• ⚠️ Need continuous large-scale R&D investment
• ⚠️ Long ecosystem construction cycle, difficult to profit in the short term

:

• Growth of GitHub Stars and Contributors
• Number of supported chips and adaptation quality
• Community activity and Issue response speed

:

• ARR (Annual Recurring Revenue) growth rate
• Customer retention rate and NRR (Net Revenue Retention)
• Number of benchmark customers (especially leading cloud vendors and large model companies)

:

• Number of partners (chip manufacturers, cloud vendors, ISVs)
• Size of developer community
• Completeness of technical certification and training system

.
Reasons are as follows:

1. The AI Infra platform market is expected to grow by 95.1% in 2025, and the computing power management layer accounts for 64.6%[2]
2. Heterogeneous computing power scheduling is the key technology to solve low GPU resource utilization (60%→90%+)[1]
3. Domestic GPU substitution accelerates, requiring a unified scheduling platform to support (localization rate exceeds 50% by 2029)[4]
4. Leading enterprises like SenseTime and Baidu have verified market demand[5]

.
Reasons are as follows:

1. Replicable Elements
   :
   • Technical positioning: Virtualization middleware at the infrastructure layer
   • Business model: Open core + enterprise services
   • Ecosystem construction: Partner network and certification system

2. Elements That Need Adjustment
   :
   • Shift from private deployment to cloud-native and multi-cloud architecture
   • Shift from enterprise-level subscription to more flexible SaaS/hybrid models
   • Shift from closed ecosystem to open source community-driven
   • Accelerate commercial rhythm (AI era develops much faster than virtualization era)

3. Success Probability Evaluation
   :
   • Technical feasibility: ⭐⭐⭐⭐ (4/5)
   • Market demand: ⭐⭐⭐⭐⭐ (5/5)
   • Team capability: ⭐⭐⭐⭐ (4/5)
   • Business model: ⭐⭐⭐ (3/5)
   • Competitive environment: ⭐⭐ (2/5)
   • Comprehensive Score: ⭐⭐⭐⭐ (4/5) - Has good potential for success

:
HAMi is in a market. If it can:

1. Establish industry benchmark cases within 12-18 months
2. Rapidly expand open source community influence
3. Establish