-->

🎤 ThakiCloud @ KCD Seoul 2025 Presentation Information

📅 Date

May 22, 2025 (Thursday)

🔗 Related Links

📜 Presentation Script

🎤 ThakiCloud Introduction and xPU-based Agentic AI Infrastructure Platform

1. Intro (Slide 1)

Hello. I’m [Name] from ThakiCloud.
Today, I’ll talk about the Kubernetes-Native Agentic AI Platform that changes the infrastructure paradigm of the AI era, and the future of xPU-based AI infrastructure that we propose.

2. Company Introduction & Mission (Slide 2)

ThakiCloud is an AI infrastructure platform company that realizes the flexibility and scalability of public cloud level in private and hybrid environments.

Mission: Supporting all companies to transform to AI First

Core Technology Areas:

  • LLM & Agentic AI Infrastructure
  • Heterogeneous accelerator integrated management (xPU management)
  • Kubernetes-Native based xPU servitization

3. Why xPUaaS and Agentic AI? (Slide 3)

Current Problems:

  • GPU-centered cost (TCO) increase and supply chain instability
  • Difficulty in hardware optimization according to various workloads
  • Complex orchestration unique to Agentic AI
  • Data sovereignty issues

ThakiCloud’s Solutions:

  • xPUaaS providing various accelerators as services
  • Turnkey Agentic AI PaaS centered on developer experience
  • Data regulation response through Sovereign Cloud

4. AI Workload Optimization Flow (Slide 4)

Looking at the diagram, our platform automatically allocates the most suitable xPU according to AI workload types.

Examples:

  • Large-scale training → NVIDIA GPU Cluster
  • Real-time inference → High-performance GPU or domestic NPU
  • Batch inference → Cost-optimized hybrid structure

These pipelines are automatically optimized through continuous monitoring and feedback.

5. Cloud-Native AI Infrastructure Configuration (Slide 5~6)

Our xPUaaS is designed based on Kubernetes extension architecture:

  • Various device plugins
  • Integrated inference runtime
  • Intuitive API through xPU SDK Wrapper
  • Monitoring environment based on Prometheus, Grafana, Loki

Various clients including SDK, web, and mobile access through API Gateway.

🎤 Slide 6 Detailed Presentation Script

✅ Overall Configuration Flow

ThakiCloud’s xPUaaS architecture is a structure that visualizes the entire flow from client requests to inference accelerators.

1. Client Layer

  • Web, mobile, SDK clients access AI services through API Gateway
  • xPUaaS API Gateway centrally routes requests

2. Core Service Layer

  • Inference Service: Real-time inference processing
  • Model Management: Model registration, version management
  • xPU Resource Pools: Accelerator pool configuration
  • Autoscaling: Automatic scale adjustment according to demand

3. Kubernetes Orchestration Layer

  • Device Plugins: Register accelerators by vendor (NVIDIA, Rebellions, Furiosa, etc.)
  • Custom Scheduler: Optimized resource placement
  • Inference Runtime / SDK Wrapper: Backend integration
  • Resource Isolation / Observability: Build isolation and monitoring systems

4. Hardware Layer

  • Real-time integration with NVIDIA GPU, FuriosaAI NPU, Rebellions NPU, etc.
  • Includes driver, power, health check, firmware update management

📌 Summary Emphasis

  • Single API Gateway
  • Kubernetes-based automated infrastructure
  • Flexible xPU connectivity
  • Strong monitoring and stability assurance

6. Flexible Cloud Operation Strategy (Slide 8~9)

  • GitOps + Helm based declarative deployment
  • Multi-cloud support: On-premises, AWS EKS, GCP GKE, Azure AKS
  • Serverless scalability: Integration with ACA, Cloud Run
  • Realizing public cloud level automation and scaling in private environments

7. Reasons to Join Us (Slide 10)

ThakiCloud:

  • Leads AI infrastructure innovation
  • Pursues open source contribution-centered engineer culture
  • Grows together with the domestic NPU ecosystem

We’re waiting for partners and colleagues who will design the future of AI infrastructure together.

🔚 Conclusion

Thank you for listening.
We look forward to having more conversations during the Q&A session after the presentation.

참고