AI infrastructure is no longer a single design problem. It’s a spectrum.At one end are hyperscale “AI factories”—massive, tightly integrated systems aligned with next-generation platforms like NVIDIA’s Vera Rubin roadmap, pushing toward extreme rack densities and tightly coupled power, cooling, and compute architectures. At the other are smaller-scale enterprise and edge AI deployments, where inference, latency, and cost constraints often favor lower-density configurations that can be deployed flexibly in existing environments. Between these poles, the industry is being pulled in multiple directions at once. The roadmap suggests 100 kW racks are arriving, 200–300 kW is being engineered, and some projections point to 1 MW+ per rack—yet sustained production environments above 50–60 kW remain limited, and each step upward multiplies challenges across:

• Power delivery and distribution (including emerging 800 VDC and DC architectures)

• Liquid cooling adoption, integration, and serviceability

• Structural design and white space planning

• Operations, safety, and maintainability at scale This panel examines where density is actually landing in production today, and how different classes of AI workloads—training vs. inference, centralized vs. distributed—are shaping real-world design decisions. Topics include:

• When high-density is essential, and when it’s an over-optimization

• Where retrofits still work—and where only purpose-built high-density sites make sense • How emerging electrical architectures and rack-level designs are influencing new builds

• Whether the market will converge on a dominant high-density model, or fragment into multiple architectures aligned to different AI workload classes