CAPLAI: AI-assisted Lifecycle Provisioning for GPU data centers

The accelerating deployment of artificial intelligence (AI) workloads, driven by recent advancements in AI technology, has significantly increased the demand for computing resources and supporting infrastructure. However, the physical and electrical capacity of data centers cannot scale at the same pace, which introduces a new challenge: accommodating rising compute demand under stringent power and space constraints. New high performance GPUs offer better power efficiency compared to previous generations, and liquid cooling systems are significantly more efficient than traditional air cooling. These advancements create an opportunity to upgrade data centers that accommodate more AI workloads with limitations in physical infrastructure and power availability.We propose CAPLAI, Capacity-Aware PLanning for AI infrastructure, an AI-assisted stochastic optimization framework for lifecycle planning in GPU data centers. Our method adopts large language models (llMs) to generate diverse and plausible future scenarios that capture demand growth, hardware efficiency decay, electricity prices, and resale market trends. These scenarios feed into a stochastic optimization model that determines GPU purchase, retirement, and cooling infrastructure upgrades. We evaluate our framework using real-world traces and constraints derived from a large-scale AI data center at Brookhaven National Lab. Compared to conventional threshold-based heuristics, our approach increases the effective GPU computing capacity within the same power limit by up to 36 % and reduces lifecycle operating cost by up to 32 %. Results demonstrate that capacityaware, AI-guided planning significantly enhances efficiency and robustness with the escalating demand and infrastructural limits.