Backend.AI AI Infrastructure Operating Platform

Lablup Inc. is a Seoul-based AI infrastructure software company founded in 2015, with a US subsidiary in San Jose. It grew out of research-computing work and builds software that lets organizations share and operate expensive GPU and accelerator clusters efficiently instead of leaving them underused.

Backend.AI is described as an operating system for AI infrastructure. Its Sokovan orchestrator schedules containerized workloads, container-level GPU virtualization splits GPUs into fractions, and a hardware abstraction layer manages more than 11 accelerator types from NVIDIA, AMD, Intel, Rebellions, FuriosaAI and others. A web UI provides four-tier multi-tenancy, monitoring and governance, while FastTrack covers the pipeline from data preparation to model serving. It runs on-premises (including air-gapped), on public clouds or in hybrid setups, at more than 120 sites including Samsung Electronics, KT, Shinhan Bank and Korean research institutes.

• Sokovan orchestrator: schedules containerized AI workloads across single and multi-node clusters.
• GPU virtualization: patented container-level technology that splits GPUs into fractions.
• Hardware abstraction layer: manages more than 11 accelerator types including NVIDIA, AMD, Intel and Korean NPUs.
• Multi-tenancy: four-tier resource governance across organization, department, project and user.
• FastTrack pipelines: covers data preparation, training and model serving in one workflow.
• Reservoir: air-gapped model hub and package repository for closed networks.

A Vietnamese bank can run an on-premises GPU cluster shared by risk, fraud and chatbot teams with per-department quotas. Universities and research institutes can give students and labs fractional GPU access. Telecom and cloud operators can use it as the scheduling layer for GPU-as-a-service offerings.

Fractional GPU sharing raises utilization; Lablup reports utilization rising from 30% to 85% and GPU spending falling by 40%. One platform manages mixed accelerators from many vendors, reducing lock-in. Multi-tenant quotas let departments and projects share one cluster with clear governance. Air-gapped deployment suits banks, defense and public bodies with strict data rules.

Enterprises and governments are building private and sovereign AI clusters, but GPUs are scarce and often underused. Software that pools accelerators, supports several chip vendors and runs in closed networks is in demand, including in Southeast Asia where banks, telecoms and universities are setting up their own GPU capacity.

Founded in 2015 in Seoul; US subsidiary Lablup USA established in January 2025. NVIDIA DGX-Ready Software certified, Intel and Dell partner, holds GPU virtualization patents in Korea, the US and Japan. Infrastructure partner of the Upstage consortium in Korea's sovereign AI project. KOSDAQ listing planned for November 2026; reseller partnerships in the US, UK and Singapore.