• Expanded from 500 GPUs in Phase One to More Than 1,000 in Phase Three, Demonstrating Backend.AI’s Large-Scale AI Training Infrastructure Operations

Lablup Inc., an AI infrastructure platform company led by CEO Jeongkyu Shin, announced that the Upstage Consortium, for which Lablup serves as an infrastructure partner, has advanced to the third phase of the Ministry of Science and ICT’s Sovereign AI Foundation Model Project.

The Sovereign AI Foundation Model Project is a national strategic initiative aimed at developing globally competitive AI foundation models using domestic technology. Following phased evaluations, the Upstage Consortium has qualified to participate continuously from the first through the third phase of the program.

Lablup has operated the consortium’s end-to-end training infrastructure through Backend.AI, its AI infrastructure platform. During the first phase, Lablup operated a cluster of more than 500 NVIDIA B200 GPUs to support training for Upstage’s Solar Open 100B model.

In the second phase, Lablup supported training for Upstage’s Solar Open 2 250B model on a 1,000-GPU cluster comprising NVIDIA H100 and B200 GPUs. The company also provided stable GPU resource operations so consortium partners could use the infrastructure for a range of work, including model compression, evaluation-pipeline enhancement, industry-specific proof-of-concept services, foundational research, and AI safety initiatives.

The third phase will expand the cluster to more than 1,000 GPUs. As post-training, including reinforcement learning (RL), becomes increasingly central to improving foundation-model performance, Lablup will introduce full Slurm interface support in Backend.AI during this phase.

This enables users to run Slurm-based workflows, widely used in large-scale training environments, directly on Backend.AI. Organizations can preserve compatibility with their established reinforcement-learning pipelines while leveraging Backend.AI’s resource isolation, scheduling, and heterogeneous accelerator-management capabilities.

“Operating clusters that grew from 500 to more than 1,000 GPUs reliably across three phases demonstrates Backend.AI’s ability to support large-scale AI training infrastructure in a real national R&D program,” said Jeongkyu Shin, CEO of Lablup. “In the third phase, we will further advance the platform through Slurm interface support and broader NPU compatibility, enabling us to meet the needs of a wider range of large-scale training and inference environments.”