THE FLOOR QUESTION IS SETTLED A Data-Driven Case for Slab-on-Grade Construction in Modern, AI-Era, Liquid-Cooled Data Centers
A Master Reference White Paper Pros and Cons of Slab versus Raised Access Floors for Data Centers
Abstract
For more than four decades, the raised access floor has been the default underpinning of the modern data center. Originating in the IBM System/360 era, when mainframes required a dedicated chilled-air plenum and a labyrinth of bottom-entry cabling, the raised floor became, through accident of history, the visual and structural shorthand for what a data center is. Today that default is wrong.
This paper assembles the evidence — drawn from BICSI 002-2024, ANSI/TIA-942-C, the Uptime Institute Tier Standard: Topology, ASHRAE Technical Committee 9.9 thermal guidelines (5th edition), EN 50600, ISO/IEC 22237, NFPA 75 and 76, IEEE Standard 1100, FM Global Data Sheet 5-32, peer-reviewed computational fluid dynamics studies, the Uptime Institute Annual Outage Analysis series, NASA Goddard zinc whisker research, and the published practice of every major hyperscale and AI factory operator — for the conclusion that slab-on-grade construction is the correct default for new data center white space and that legacy raised floors should be retired during the next major refresh in all but a small set of legitimate edge cases.
The argument has six pillars. First, no major standard requires a raised floor; both Uptime and EN 50600 explicitly state the contrary. Second, underfloor air distribution wastes between forty-eight and sixty percent of conditioned air to bypass leakage on average, a problem that thirty years of grommet, blanking-panel, and tile-management product development has improved but not solved. Third, raised-floor systems hit a hard density wall at approximately eight to ten kilowatts per rack, far below the fifty-to-one-hundred-thirty kilowatt densities now shipping in NVIDIA GB200-class AI racks. Fourth, the failure modes of the underfloor void — chilled-water leaks, zinc whisker contamination, cable abandonment, fire detection complexity, and structural under-rating relative to modern rack weights — are catalogued, photographed, and quantified in the public literature. Fifth, the total cost of ownership advantage tilts decisively toward slab once insurance burden under FM Global DS 5-32, NFPA 75 sub-floor detection requirements, and continuous airflow tuning labor are accounted for. Sixth and finally, the industry has already voted: the Open Compute Project encodes slab as a baseline, every named hyperscale operator builds slab, and every named AI factory in operation today is slab-on-grade by necessity, not preference.
Four edge cases survive serious scrutiny: special-security facilities under ICD 705 where inspection regimes favor accessible voids; legacy telco central offices with sunk-cost cabling networks; a narrow set of OEM equipment with bottom-only cable entry; and customer-facing aesthetic environments where the appearance of a raised floor remains a marketing signal. These are documented and the boundaries of each are drawn explicitly.
The paper closes with a four-step decision framework, a retrofit pathway for operators committed to legacy raised-floor sites, and an explicit set of recommendations for greenfield builds, brownfield refreshes, and the engineering community that designs both.
Full paper below
