A single NVIDIA NVL72 rack draws 132 kilowatts. The hyperscale data centers built between 2018 and 2023 were sized for racks drawing somewhere between 8 and 15 kilowatts. The buildings did not get an upgrade. The chips inside them did. The gap between what the silicon now requires and what the concrete around it can deliver is the largest unspoken liability on the balance sheet of the cloud industry, and it cannot be closed by retrofit.
The number that matters more than the kilowatts is the floor loading. An air cooled data center is engineered for roughly 250 pounds per square foot. A liquid cooled facility, with its coolant distribution units, manifolds, dielectric fluid reservoirs, and reinforced piping, needs between 500 and 1,000 pounds per square foot. Concrete does not negotiate. You cannot pour additional capacity into a slab that was finished six years ago. You can demolish, or you can build new.
The Density Curve
In 2019 a typical hyperscale rack was a 10 kilowatt unit. Facebook, Google, and Amazon were running fleets of standardized cabinets at 8 to 12 kilowatts, cooled by hot aisle containment and CRAC units pushing chilled air at supply temperatures around 75 degrees Fahrenheit. The economics of the building were sized to that envelope. Power distribution, transformer sizing, busway capacity, cable trays, and crucially the air handling load were all designed against a ceiling that nobody expected to move quickly.
It moved quickly. By 2022 the first generation of training clusters for large language models pushed rack densities into the 25 to 35 kilowatt range. By 2024 the NVIDIA H100 reference designs were landing at 40 to 60 kilowatts. In 2025 the GB200 NVL72 reference architecture shipped at 132 kilowatts per rack, and the next generation announced under the Rubin codename is targeting somewhere between 200 and 300 kilowatts per rack at scale. AMD MI300 and MI325 deployments sit in the 50 to 80 kilowatt band. The curve is not flattening.
Ten to thirty to one hundred to three hundred kilowatts. Five years. The chips compounded. The buildings did not.
Where Air Cooling Physically Fails
Air is a poor coolant. The specific heat capacity of air at room temperature is roughly 1.005 kilojoules per kilogram per Kelvin. Water is roughly 4.18. To remove a given quantity of heat, water requires about a quarter of the mass flow. When you account for the density difference, water moves heat per unit volume about 3,500 times more efficiently than air. This is not an engineering preference. It is a thermodynamic floor.
At 10 kilowatts per rack the airflow required is large but manageable, and the temperature rise across the rack is acceptable. At 30 kilowatts the air mass flow needed to keep junction temperatures within manufacturer specifications climbs to the point where the velocity through the rack starts producing acoustic and turbulence problems, and the chilled water plant feeding the air handlers grows substantially. Somewhere between 30 and 40 kilowatts per rack the curve stops being asymptotic and becomes a wall. You can engineer rear door heat exchangers, in row coolers, and overhead supplemental cooling to push the envelope into the 50 to 70 kilowatt range for a single rack in a half empty hall, but you cannot do it for a full hall packed to that density. The chilled water plant required to feed enough air to cool a full hyperscale floor at 100 kilowatts per rack does not fit on the site.
This is the part that gets missed in the trade press. The limit is not the chip. The chip will accept whatever cooling you give it. The limit is how many watts of heat you can dissipate per square foot of floor area before the air itself becomes the bottleneck. Above roughly 30 kilowatts per rack the answer is that you cannot do it with air, regardless of how much money you spend on the air handling system.
What A Liquid Cooling Architecture Actually Requires
Direct to chip liquid cooling replaces the heatsink on the processor with a cold plate. A fluid, usually a water glycol mix or a dielectric coolant, flows through micro channels milled into the plate, picks up the heat, and exits to a manifold. The manifold feeds a coolant distribution unit, the CDU, which is the heart of the system. The CDU performs the heat exchange between the technology cooling system loop, which runs inside the rack and across the chips, and the facility water system loop, which connects to the cooling towers or dry coolers outside the building.
The CDU is not a commodity item. Vertiv, Schneider Electric, Stulz, Motivair, and JetCool are the volume suppliers. Lead times in the second quarter of 2026 sit between 40 and 60 weeks. A megawatt of cooling capacity in CDU form factor costs several hundred thousand dollars at the unit level, and a full hyperscale deployment requires dozens of them, redundantly configured, with isolation valves, leak detection, fluid quality monitoring, and tertiary loop integration.
The piping is the part that the retrofit estimates underprice. A liquid cooled facility needs supply and return manifolds running through every row, with quick disconnect couplings at every rack position, leak containment trays under every connection, and a leak detection system that can isolate a fault to a single rack without taking down the row. The pipe sizing scales with the cooling load. At 132 kilowatts per rack the flow rates require pipe diameters that do not fit through the conduit penetrations that were core drilled into a 2019 air cooled facility. You can cut new penetrations, but every one of them is a structural review.
Two phase immersion is the other end of the architecture. The entire server is submerged in a dielectric fluid that boils on contact with the hot components, carries the heat away as vapor, and condenses on a coil at the top of the tank. Submer, Iceotope, GRC, and LiquidStack are the names that matter here. The thermal performance is exceptional. The capital cost is high and the operational maturity is still building. For racks above 150 kilowatts the trade off increasingly favors immersion despite the cost.
Single phase direct to chip remains the volume technology through 2027. Asetek and CoolIT are the volume suppliers of the cold plate hardware. Immersion is the technology that matters for the frontier.
Why Retrofit Math Does Not Work
A hyperscale operator looking at a 2020 vintage air cooled colocation facility and considering whether to retrofit it for 100 kilowatt liquid cooled racks runs the numbers and gets the same answer every time. The retrofit costs between 60 and 80 percent of new build, and at the end of it you have a building that is constrained by the original structural envelope, the original electrical service capacity, and the original site water supply.
The structural problem dominates. A slab rated for 250 pounds per square foot cannot accept the weight of fully provisioned liquid cooled racks plus the manifold infrastructure plus the CDU plant without reinforcement. Reinforcement means either supplemental steel framing under the existing slab, which requires shutting the facility down and excavating the underside, or pouring a new slab on top, which lowers the ceiling height and forces relocation of overhead infrastructure. Neither option is cheap and neither option is fast. Most existing buildings cannot be reinforced to the 1,000 pounds per square foot needed for high density immersion at any reasonable cost.
The electrical service is the second constraint. A 36 megawatt facility built in 2020 has the transformers, the switchgear, and the utility interconnection sized for that load. Doubling the rack density without expanding the building doubles the load on the electrical infrastructure. The utility upgrade alone, if the local substation can accommodate it, is a multi year process. If it cannot, the conversation moves to a new substation, which is a different category of project.
The water is the third constraint. Air cooled facilities using evaporative cooling towers consume between 5 and 15 million gallons per day of makeup water at the gigawatt scale. Closed loop liquid cooling, paired with dry coolers or hybrid coolers, can cut that consumption by an order of magnitude, but only if the cooling tower infrastructure is replaced. Retrofitting from open loop evaporative to closed loop dry is, again, a substantial fraction of new build cost and requires significant outdoor equipment yard space that the original site plan may not have reserved.
What the retrofit calculation produces, almost without exception, is a number that exceeds the cost of building a purpose designed liquid cooled facility on a greenfield site, on a faster schedule, with a longer useful life. The hyperscalers know this. The colocation operators with hundreds of megawatts of 2018 to 2023 air cooled inventory know this too, and the implication for their balance sheets is the part of the industry conversation that is happening quietly.
The Steam To Electric Parallel
Factories in 1890 ran on steam. A single central engine drove an overhead lineshaft, and every machine in the building was connected to that shaft through belts and pulleys. The architecture of the factory followed the architecture of the power transmission system. Buildings were long and narrow to keep belt runs manageable. Floors were heavily reinforced to handle the vibration. Machine layout was constrained by proximity to the shaft.
When electric motors arrived, the obvious thing to do was to replace the steam engine with an electric motor that drove the same lineshaft. This is what most factories did between 1890 and 1910. The productivity gain was modest. Buildings were still long and narrow, machines were still tied to the shaft, layout was still constrained.
The productivity revolution came when factories rebuilt around unit drive. Each machine got its own motor. The lineshaft disappeared. Buildings could be shaped to the workflow rather than the power transmission. Ford Highland Park, opened in 1910 and refined through 1914, was the iconic example. The moving assembly line was possible because each station had its own power source and the building did not have to be organized around belt geometry.
The economic historian Paul David framed this in his 1990 paper on the dynamo and the computer. The lag between the arrival of a general purpose technology and the realization of its productivity benefit comes from the time required to redesign the surrounding architecture. The technology arrived in 1880. The productivity surge arrived in 1920. Forty years of lag, almost entirely structural.
The data center industry is in the early phase of the same transition. Liquid cooling is the electric motor. The buildings constructed for air cooling between 2018 and 2023 are the steam factories with electric motors bolted onto the old lineshaft. They will continue to operate, at degraded competitive position, until they are replaced by purpose designed liquid cooled campuses. The replacement cycle is the productivity surge that the industry is now beginning.
What Becomes True By 2028
Direct to chip liquid cooling is the default specification for any rack above 50 kilowatts. New build hyperscale campuses underwrite liquid cooling as a baseline assumption rather than a feature. The CDU supply chain widens to include three or four new entrants beyond the current incumbents. Lead times compress from 60 weeks to 20.
Two phase immersion expands from the experimental tier into volume deployment for racks above 150 kilowatts. The dielectric fluid supply chain, currently concentrated in a handful of specialty chemistry suppliers, becomes a strategic constraint.
The legacy air cooled colocation inventory built between 2018 and 2023 becomes a measurably depreciating asset class. Cap rates widen between liquid ready and air cooled facilities. The widening shows up in transaction comps by 2027 and in published valuation indices by 2028. A subset of air cooled facilities, particularly those with adequate site power and water but inadequate structural capacity, get demolished and rebuilt on the same parcel.
The campuses being entitled and designed now, in 2026, are the ones that will define the competitive landscape for the back half of the decade. The ones being retrofitted are catching up to where the frontier was two years ago. The ones built before 2023 and left in place are the steam factories.
The building lost. The chip won. The replacement cycle is the trade.