Conference Presentation, Keynote
Industrializing Intelligence: Building the AI Factory as a System | Vertiv | RAISE Summit 2026
- The AI industry is shifting from a focus on "proven AI" to maximizing "useful AI" output, where physical infrastructure serves as the critical constraint for AI factories.
- Traditional metrics like Power Usage Effectiveness (PUE) are becoming obsolete in favor of "tokens per megawatt" and "tokens per second per megawatt" as primary efficiency indicators.
- Five primary forces are reshaping data center economics: density, speed of deployment, scale, optimization, and increasing complexity in thermal and current management.
- Rack-level power density is projected to surge from traditional kilowatt ranges to 100 kW, 600 kW, and eventually 1 MW per rack.
- Industry success is now defined by "time to token," requiring rapid capacity deployment and extreme optimization of available power.
- The prevailing construction model of assembling discrete, optimized components (switchgear, UPS, chillers) creates "seam" inefficiencies that limit overall system performance.
- Vertiv advocates for a "converged system" approach, where the powertrain, thermal chain, and compute are engineered holistically as one seamless unit to eliminate capacity leakages.
- Design is moving toward "grid-interactive" powertrains that enable behind-the-meter generation and actively support grid stability.
- White space deployment is transitioning from complex, on-site integration to pre-engineered, modular solutions that reduce deployment time by an order of magnitude.
- The concept of "manufacturing data centers" is emerging, treating infrastructure as a single SKU produced in a controlled factory environment rather than a bespoke site-built project.
- Modularization allows for seamless scaling of infrastructure to hundreds of megawatts or gigawatts as compute demand and power availability expand.
- Reference designs are increasingly optimized for specific silicon models, ensuring maximum efficiency and reducing the need for custom on-site engineering over the asset's 15–20 year lifespan.
- Full modular deployment can halve the time-to-deployment, reduce the physical footprint, and significantly lower the Total Cost of Ownership (TCO).
- Lower TCO directly translates to improved "cost per token," which is identified as a critical lever for competitiveness in the AI sector.
- Close collaboration with silicon providers and hyperscalers is now mandatory, as infrastructure must be future-proofed to host next-generation silicon released two to three years down the line.
- Vertiv emphasizes a global service organization to maintain efficiency and operational stability for massive, long-duration infrastructure deployments.