Conference Presentation, Keynote
Industrializing Intelligence: Building the AI Factory as a System | Vertiv | RAISE Summit 2026
- Persistent constraints on compute capacity and available power will define a "constrained world," necessitating a focus on deployment speed and the efficiency of converting power into tokens.
- Traditional data center capabilities will become inadequate as rack density requirements shift from kilowatt levels to targets of 100, 600, and eventually one megawatt per rack.
- Market success will be determined by the "time to token," driving a strategic pivot from bespoke designs to rapid deployment and scale to manage thermal loads.
- Operational efficiency will depend on engineering the powertrain and thermal chain as a single seamless system to eliminate capacity leakages inherent in discrete elements.
- Data centers are expected to evolve into "grid interactive" and potentially "autonomous" systems with behind-the-meter power generation to address power limitations.
- Pre-engineering and modularization in white spaces are projected to reduce deployment time by "an order of magnitude," accelerating time to revenue and time to token.
- Data centers will transition to being manufactured as a "one SKU" product in factories to ensure repeatability and eliminate on-site variability.
- Fully modular, converging infrastructure is anticipated to be deployable in "approximately half the time" of traditional builds while occupying a smaller footprint.
- Modular products will possess the scalability to expand "all the way to hundreds of megawatts or gigawatts" in response to growing compute needs or power availability.
- Infrastructure must be designed today to host "new silicon" arriving two to three years in the future, requiring close collaboration with silicon providers and hyperscalers.
- Assets within massive capacity infrastructures are expected to operate for "15, 20 years," requiring service organizations to maintain efficiency over that extended lifespan.
- A comprehensive portfolio of physical infrastructure technologies will be required, ranging from traditional powertrains to new 100 volt DC powertrains and thermal solutions.
- The shift toward pre-engineered, optimized, and scaled deployment is expected to dramatically lower the total cost of ownership and the "cost per token."