Retrofitting Air-Cooled Facilities for Liquid

The building you are retrofitting was sized for a different physics: 3 to 7 kW racks on a raised floor, air handlers tuned to a 1.4–1.6 PUE, and an electrical bus built for a rack power density that one GB300 NVL72 rack now outstrips by 20x. The retrofit question is never 'can liquid cool this hardware' (it can). It is 'what does this building let me do before one of its original constraints runs out.' About half of the facilities in Uptime Institute's 2024 Global Data Center Survey were more than 10 years old (reported October 2024): air-era stock built before AI-class power density and cooling. Whether such a hall is worth converting depends on its remaining building or lease life and on its power, structure and cooling capacity, which the assessment below measures.

The decision runs in the order the building forces on you. The brownfield assessment inventories four fixed quantities before any capital is committed. The retrofit paths then rank by disruption, because in a live hall the cost that matters is downtime per rack and risk to the racks already earning, not capex per kW. Phased migration through mixed air-and-liquid operation is where the hardest engineering sits: keeping a half-converted hall thermally and electrically stable while you commission charged piping over running equipment. Last comes the stranded-capacity math, which reconciles power, cooling and floor to show how much of the building you made usable and how much you paid to leave empty.

Every retrofit feasibility study reduces to four inventories, and a candidate building passes only if it clears all four. They are an AND, not an OR. Miss any one and that constraint becomes your density ceiling, however generous the other three are. Find that ceiling before design, because it decides whether you are buying a liquid-capable AI hall or a more expensive air hall.

Floor loading is an independent structural gate. It is structure, not equipment. A GB200 NVL72 weighs about 1.36–1.50 t (3,000–3,300 lb wet; HPE's filled profile is 3,245 lb), and GB300 NVL72 sits in the same band. NVIDIA describes the Vera Rubin NVL72 as roughly 4,000 lb (about 1.8 t; March 2026) without stating its fill state or footprint, so a Vera Rubin floor assessment starts from the OEM's installed mass, footprint and support reactions. Spread the GB200 mass band over the HPE GB200 cabinet's 0.64 m² (600 × 1,068 mm) and the footprint average comes to about 2,100–2,300 kg/m². Use it as a first screen; it is neither a point load nor an acceptance test. Floor ratings come in separate concentrated, rolling and distributed forms. For example, Tate rates its ConCore 1000 access-floor panel for a 1,000 lb concentrated design load when it sits on Tate's own understructure. Get the OEM foot and wheel reactions and compare them with the complete panel, pedestal, stringer, slab and move route. A shortfall can be relieved three ways. Steel plates can spread the load. Removing the raised floor lands racks on a structural slab you have verified. Strengthening is the third, and the fix follows the weak element. Structural members take FRP or steel, a weak slab-on-grade gets a designed overlay or replacement, and a foundation that governs needs underpinning or micropiles. Each brings access, cost and outage consequences, so price the chosen scope from the engineer's bill of quantities and your live-site outage plan. → the structural basis is engineered in Chapter 6.2.

Plenum and distribution volume is the second. A raised-floor plenum that was an air path is now a contested corridor. It must carry coolant supply and return mains, CDU connections and leak-zoning drainage, and it is often already congested with decades of power and data cabling. A slab-on-grade hall with no plenum forces overhead pipe racks and a different leak-containment geometry. The plenum question decides where the CDU lives (in-row, end-of-row or perimeter gallery), how the secondary loop is routed, and whether liquid distribution competes with the air cooling you still need for the rest of each rack's heat. That share depends on the rack and generation: Lenovo's GB300 NVL72 guide puts about 10% of rack heat on air, while a Vera Rubin NVL72 has liquid-cooled, fanless compute and switch trays but keeps its power-shelf PSUs on air. Ask the OEM for the rack's air and liquid heat schedule, and add network and storage gear separately. → spatial planning for liquid in Chapter 6.1.

Electrical headroom is the third, and it is frequently the true ceiling. A hall sized for 5 kW racks has a bus, transformers, switchgear and PDU topology to match. Liquid does not reduce the IT load. It does not remove heat rejection or cooling auxiliaries either, so credit only a reduction you have calculated for your path, climate and required state. Meanwhile the racks draw far more: HPE's GB200 NVL72 at 132 kW nominal TDP is 26x a 5 kW rack, and NVIDIA's GB300 NVL72 at up to 142 kW is 28x. For the electrical check itself, use the OEM's peak and provisioning basis rather than nominal TDP. HPE's GB300 NVL72 QuickSpecs (September 2026) give about 155 kW EDPp and recommend 192 kW busway provisioning, 38x a 5 kW rack; Lenovo's GB300 NVL72 peaks at up to 155 kW. Vera Rubin NVL72 draws more than GB300's 155 kW peak. Pegatron rates its RA4803-72N3 at 188 kW in Max Q mode and 228 kW in Max P, between 38x and 46x, and NVIDIA designs DSX around a 330 kW cabinet, 66x. If the upstream feed, busway ampacity or UPS capacity caps out at, say, 8 MW, you can cool far more density than you can power, and the retrofit is power-bound the moment the cooling works. The worst version succeeds thermally and then strands half the floor because there is no megawatt left to energize it. → the power chain in Chapter 4.5; the binding-constraint framing in Chapter 1.1.

Viable heat rejection is the fourth, and its plant is a site constraint. A full liquid-to-liquid retrofit needs a facility-side liquid loop and a heat-rejection plant sized to the new load. Dry coolers and air-cooled chillers avoid process-water use but must clear ambient-bin capacity, electrical, acoustic and space constraints. Evaporative rejection also depends on water supply, treatment and discharge. → water sourcing and climate-driven rejection in Chapter 3.7 and Chapter 5.8.

There are three real paths from an air hall to liquid, and you rank them by how much they disturb a building that is already earning revenue, not by peak density. In a live hall, downtime per rack and risk to the neighbors cost more than dollars per kW. The paths form a disruption ladder. Closed-loop liquid-to-air units (L2A) sit at the bottom: no facility plumbing, and they work with the hall's air plant. Water-fed rear-door heat exchangers (RDHx) sit in the middle: a room water loop, with the servers untouched. Full direct-to-chip liquid-to-liquid (L2L) sits at the top: a facility-side loop, CDUs, in-rack manifolds and the largest civil works.

L2A: closed-loop liquid-to-air (the drop-in). The cold plates and in-rack loop are real direct-to-chip liquid cooling, but the heat goes back into the room air through a liquid-to-air exchanger in the rack or in an adjacent sidecar, with no facility water piping at all. Air-assisted liquid cooling (AALC) with cold plates belongs here: Meta runs GB200 in its Catalina racks this way (September 2025) because its traditional halls have no facility liquid. Microsoft retrofits liquid-to-air heat exchanger units ("sidekicks") into existing Azure halls to cool GB200 (November 2024), and in October 2025 described a next-generation unit it says doubles performance with no data center modifications. Google's Brazos sidecar, announced in June 2026, is now generally available; its OCP specification (v0.75.4) took effect in September 2026 and was published by October. Its three modular units share a nominal 60 kW of rack thermal capacity between them, and they sit beside the rack they cool or inside it. The limit is the room. Every watt the rack captures lands on the hall's air system, along with fan and pump heat from the exchanger, so L2A serves as many racks as your verified air distribution and ultimate heat rejection can carry in every required state. The servers still need liquid-ready trays and an OEM-supported configuration: HPE's GB200 NVL72 configuration specifies L2L, for example.

RDHx: water-fed rear doors (the bridge). A rear-door heat exchanger replaces the rack's back door with a water coil that intercepts the hot exhaust and sends the heat to a room water loop. Capacity depends on the door, water temperature and airflow, so quote it per named model. nVent's active 800 mm RDHX PRO door, with fans, is rated 78 kW at 14 °C water and 24 °C air, and 44 kW at 24 °C water and 27 °C air. A passive door has no fans of its own and relies on the servers to push air through the coil. nVent's passive 800 mm RDHX PRO door is rated 38 kW at 14 °C water and 24 °C air, and 34 kW at 24 °C water and 27 °C air. SemiAnalysis puts typical doors at 30–40 kW and active doors above 50 kW (Chapter 5.3 compares the ratings and their test conditions). A rear door needs a water supply and return that meet its temperature, chemistry, flow and pressure limits. Feed it directly from a compatible facility loop, or through a CDU-fed secondary loop where the facility water would run at or below the room dew point, and hold exposed coil and hose surfaces above that dew point. In an RDHx-only deployment, the servers can remain air-cooled, with no chip-level cold plates or liquid connections. That makes it the natural bridge for a hall ramping density gradually, or for a mixed fleet where some racks are dense and some are not. → engineered in full in Chapter 5.3.

L2L: full direct-to-chip, liquid-to-liquid (the destination). Cold plates on the hot silicon, rack manifolds, and a CDU that isolates the technology-cooling loop from the facility-side loop, with heat rejection sized to the new load. Use L2L when the room's air budget cannot carry the rack population and the facility-liquid path can. HPE specifies L2L for its own GB200 NVL72 configuration, at 132 kW nominal rack TDP. Rack power alone does not choose the path for GB300 NVL72 (up to 142 kW per rack in NVIDIA's reference architecture) or for denser successors: follow the configuration your OEM supports and the room-air and facility-liquid budgets you have verified. L2L can justify its cost in a training-shaped hall when the OEM-supported hardware and the installation, service and operating plan repay the disruption. It is also the most disruptive: it needs a facility-side liquid loop and a heat-rejection plant sized to the load, plus the largest civil and electrical works. Site water supply and discharge become extra gates only when that plant uses evaporation. Ask the OEM for the rack's connection schedule rather than assuming a quick-disconnect count. → engineered in Chapter 5.4; CDUs and the secondary loop in Chapter 5.6.

The table is a disruption ladder with a density tax attached at the bottom. L2A avoids a facility-water tie-in, so it is the least negotiation with the building's plumbing. It does not avoid the structure: the dense rack, its filled exchanger and any ancillary unit still need verified support and a move route. Vertiv's CoolChip CDU 350, a row-format liquid-to-air unit piped to nearby racks, weighs 920 kg dry on its own. The ~60 kW figure is Brazos's, not the architecture's: Vertiv rates that row unit at 350 kW at a 15 °C approach temperature difference, or ATD (Chapter 5.6), a capacity shared across the racks it serves rather than a per-rack rating. Every L2A variant still loads the air system, so a hall that goes all-L2A is ultimately limited by air rejection. Water-fed RDHx buys headroom for a door swap and a room water loop, but stops where its air and water capture no longer covers the rack's duty. L2L serves racks whose cold-plate heat cannot fit the room's air budget, and it demands the facility water and structural capacity the assessment may have told you the building lacks. In practice the tightest limit chooses the path. Where facility liquid is missing but structure, power and room-air rejection hold up, look at an L2A configuration your OEM supports. Water-fed RDHx needs its water path. Every option still has to pass the structural gate.

You almost never convert a live hall in one outage. You convert it in phases, and for a long, awkward middle period the hall runs mixed: some rows on air, some on liquid, sharing a room, a power chain and an air-handling plant that was balanced for neither. The engineering difficulty of a retrofit lives in that mixed state, not at either endpoint.

The first problem is thermal balance, and what a converted row does to the room depends on its path. An L2L row, and a water-fed rear door for the heat it captures, send heat out through facility liquid, so the room's air load falls, unevenly. An L2A row returns the full rack heat to the room, plus the fan and pump heat of its exchanger, so its air load does not fall at all. The surviving air rows still need full airflow and supply temperature. Where the room load does drop, CRAH units sized for the old load short-cycle or over-cool, and containment tuned for a uniform air hall must cope with rows that give off almost no exhaust. Containment, airflow and CRAH/CRAC setpoints have to track the conversion row by row, or you get hot spots on the surviving air rows while the liquid rows run cold. → hybrid containment in Chapter 5.2.

The second problem is electrical sequencing. Each liquid row is a step change in rack power on a bus provisioned for far less. The conversion order has to respect the upstream feed, PDU loading and UPS capacity, so that energizing a new dense row does not starve or trip the rows already running. Here the electrical-headroom inventory stops being a feasibility number and becomes an operational schedule.

The third, and the one that keeps facility managers awake, is commissioning charged piping over live equipment. Flushing, filling and pressure-testing a coolant loop a few feet from running, revenue-bearing racks makes a first-fill leak an incident that can take down neighbors, not a maintenance event. The mitigations are procedural and physical: leak zoning so a failure stays inside its row, leak detection wired to fast-acting shutoffs, blind-mate and quick-disconnect couplings that minimize open-fitting time, hydrostatic acceptance on isolated sections before they go near live load, and a sequence that proves each loop dry-then-wet before any GPU depends on it. The retrofit's defining risk is wet work in an occupied building. → leak detection, containment and the commissioning sequence are engineered in Chapter 5.11 and accepted in Chapter 13.5.

The deliverable of a retrofit is usable megawatts of IT load on the floor, not 'liquid cooling installed.' A retrofit reshuffles the three quantities that gate usable load (power capacity, cooling capacity, and the floor area you can structurally and physically populate), and they almost never line up. Whichever is smallest is your real capacity. The gap between it and the others is stranded capacity you paid to create and cannot use.

Work it as three independent budgets for the same hall, then take the minimum:

• Power-limited load = the megawatts the upstream feed, busway, switchgear and UPS can actually deliver to IT after the new density. Liquid can reduce the cooling auxiliaries' share of the electrical budget, so credit only the reduction you have calculated for the chosen path and climate; the racks use far more than any saving.
• Cooling-limited load = the heat the chosen path can reject. For L2A this is bounded by what the room air system can carry away. A water-fed RDHx sends its captured heat to the facility loop, so the room plant carries only the uncaptured share and the door-failure case. For L2L it is bounded by the facility water loop and heat-rejection plant you installed.
• Floor-limited load = the racks you can structurally place (slab and floor rating after reinforcement) and physically fit once CDU galleries, pipe racks, leak zoning and service clearance consume white space.

Usable load = min(power, cooling, floor). The other two are stranded. A hall retrofitted to 12 MW of cooling but fed 8 MW has 4 MW of stranded cooling. A hall powered to 12 MW whose reinforced slab carries dense racks across only 60% of the area has stranded power sitting over a floor it cannot load. The pre-retrofit hall was typically cooling-bound, because air ran out first. A successful liquid retrofit usually moves the limit to power or floor, which is why the assessment has to inventory all four quantities up front, and why the stranded number, not the density headline, decides whether the retrofit justified its capital.

Counts are floor(7,000/100) = 70 electrical, floor(5,200/80.0) = 65 liquid, floor(1,400/20.0) = 70 air and 80 positions. Select 65 complete racks, 6.50 MW IT. Unused resources are 0.500 MW electrical, 0.100 MW air and 15 positions. Rack 66 requires 5.28 MW liquid and fails.

The first crossover is 5.28 MW liquid for rack 66. Adding 0.400 MW yields 5.60 MW and permits 70 racks at 7.00 MW IT and 1.40 MW air, so power and air now bind. More liquid alone buys no rack. Price that whole-rack step with the method in Chapter 1.8. Chapter 13.5 covers acceptance of the migration states, and Chapter 6.2 verifies the structure.

A retrofit pays back by raising the number of complete racks the hall can host, and it spends disruption and capital across every affected system to get there. In the assumed case, liquid capacity stops the hall at 65 racks. Expanding liquid can take it to 70, where power and residual air run out together, and past that point more liquid capacity alone buys no additional rack. Fund the upgrade that clears the next limit, hold irreversible substrate for the ramp you have funded, and include the cost of taking a live hall through the work. Anything else buys installed cooling and reinforced floor that cannot host another rack, after the disruption has already been paid.

The retrofit can sell as many racks as the first limit to run out allows: power, liquid, air, structure or the installation route. Expand that subsystem to release the next increment, and price the live-site work and outage before you compare that increment with new construction.