Back
Key Takeaway
Building a scalable cloud architecture for GPU-intensive LiDAR AI SaaS
Through a Provisioning Ensemble structure combining Terraform, AWS CDK, and Karpenter, we secured stability and scalability for LiDAR AI SaaS (VueX), and achieved approximately 30-45% reduction in operational costs through GPU auto-scaling.
Vueron Technology
Client :Vueron Technology
Industry :Software / Mobility Tech
1. Overview (Project Background)
Vueron Technology launched a new SaaS-based LiDAR AI Platform VueX to the global market this year, and internally built a cloud-based architecture to operate it stably and scalably.
Since VueX is a platform where high-performance GPU workloads are essential for large-scale LiDAR data processing, Auto-labeling, and model training, we technically adopted a Provisioning Ensemble structure (combination of Terraform, AWS CDK, and Karpenter) to meet these requirements and implemented it on the AWS environment.
Megazone Cloud participated in the process of verifying whether the architecture configuration of VueX conforms to AWS best practices after its launch and performed a role in reviewing structural stability.
After VueX's launch and listing on AWS Marketplace, full-scale global market expansion was needed, and in particular, securing product promotion channels in the Middle East region, where interest in mobility and deep tech industries is high, was important.
Through this process, through the 「2025 AX (Artificial Intelligence Transformation) Solution SME Overseas Expansion Support」 Middle East program in which Megazone Cloud participated as the lead organization, Vueron was provided with an opportunity to participate in the ADIPEC 2025 joint pavilion, allowing VueX to be introduced and promoted in the local market.
2. Challenge (Problem Definition)
① Standardization and consistency between infrastructure layer and service layer
As base infrastructure (VPC, EKS, RDS, etc.) and application layer (API, data pipeline) change and expand at different speeds, it is difficult to meet both stability and agility with a single IaC
Maintaining deployment reproducibility and configuration consistency across multi-environments (dev/stage/prod) is challenging
② Unpredictable resource demand according to LiDAR and AI workload characteristics
Large-scale LiDAR uploads, Auto-labeling, and model training cause mixed GPU and CPU workloads to spike suddenly, making it difficult to respond with only existing fixed node configurations
When operating multi-tenant SaaS, the usage gap between customers increases, requiring real-time scaling and cost control
③ Need to meet scalability, cost, and stability standards for global SaaS operations
GPU nodes have a high-cost structure → risk of operational costs skyrocketing if inefficient scaling occurs
Since we target global OEMs and Tier-1 customers, we must meet high operational standards such as uptime, security, and scalability, and infrastructure instability directly leads to decreased customer trust
3. Solution (Resolution)
Vueron Technology and Megazone Cloud configured a Provisioning Ensemble structure based on AWS for stable supply of VueX, and built a SaaS operating system that meets both stability and scalability.
Component | Key Implementation Details |
AWS Infra | We configured VueX's base infrastructure including VPC, Subnet, EKS, and RDS on the AWS environment, designed to meet the stability and security standards required for global SaaS. |
Terraforming | We codified immutable infrastructure (accounts, networks, clusters, etc.) with Terraform to establish a highly reproducible deployment system. |
AWS CDK | We structured VueX's service features (API Gateway, Lambda, S3, data pipeline, etc.) based on CDK, |
Karpenter | For workloads requiring high-performance GPUs such as Auto-labeling and model training, we implemented real-time auto-scaling with Karpenter. |
Monitoring & Cost Optimization | We built a dashboard for real-time monitoring of key metrics including EKS, GPU nodes, Auto-labeling, and model training, |
4. Result (Achievements)
GPU-based Auto-labeling and model training performance have been optimized, significantly improving processing speed compared to existing standards for equivalent tasks.
With Karpenter-based real-time scaling, GPU costs have been reduced by approximately 30-45%, significantly improving operational efficiency.
By establishing a standardized deployment system based on Terraform and CDK, deployment stability and reproducibility across environments (dev/stage/prod) have been greatly improved.
By securing AWS Marketplace SaaS structure consistency, trust with global OEM and Tier-1 customers has been strengthened, and onboarding of overseas customers has become smoother.
