UCloud Raises up to 1.5B RMB for AI Center

🔑 Enhanced Key Takeaways

• •The Ulanqab center leverages a 'cold climate' advantage to achieve a design PUE (Power Usage Effectiveness) of less than 1.1, utilizing indirect evaporative cooling to reduce operational expenditure by an estimated 30% compared to traditional Tier-1 city data centers.
• •This 1.5 billion RMB injection is specifically earmarked for 'Phase 2' expansion, which focuses on high-density rack configurations capable of supporting 30kW+ per cabinet, a requirement for next-generation H200 and domestic Blackwell-class AI accelerators.
• •UCloud is transitioning its business model from general-purpose IaaS to a 'GPU-as-a-Service' (GaaS) provider, aiming to capture the demand from Beijing-based LLM (Large Language Model) startups that require low-latency (under 10ms) connectivity to the Ulanqab cluster.
• •The project includes the deployment of a proprietary 'AIGC Infrastructure Platform' that automates cluster management, fault isolation, and checkpointing for distributed training across heterogeneous hardware environments.

🛠️ Technical Deep Dive

• •Network Architecture: Implementation of a non-blocking RoCE v2 (RDMA over Converged Ethernet) network fabric to support massive-scale distributed training.
• •Cooling System: Advanced indirect evaporative cooling units combined with liquid-to-chip cooling readiness for high-TDP AI processors.
• •Storage Layer: Integration of high-performance parallel file systems (Lustre/GPFS) providing multi-terabit throughput for rapid data loading during epoch cycles.
• •Energy Sourcing: Direct-line integration with Ulanqab's wind and solar power grid, targeting 100% renewable energy utilization for the new cluster.
• •Virtualization: Use of lightweight container-based GPU virtualization to allow fine-grained resource slicing (vGPU) for inference tasks.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline