Air Cooling at the Limit

Air cooling has not left the AI hall, but its job has shrunk. A GB200 or GB300 NVL72 rack on direct-to-chip liquid still sends part of its heat to room air: HPE specifies the GB200 at about 115 kW to liquid and 17 kW to air. Air-qualified networking, storage, CPU servers and modest-density inference need moving air too. In Vera Rubin NVL72, NVIDIA liquid-cools the compute and switch trays and removes their fans, but the power-shelf PSUs stay on air, so a hall's air load now follows its actual equipment list. Air also remains the answer for a brownfield hall where structural rework would cost more than the retrofit and outage budget allows, and wherever the cheapest correct design is a fan and a coil.

The decisions here are mechanical, and their consequences show up as PUE points, fan watts and stranded floor space. Four structure the chapter: raised floor versus slab (the distribution substrate), containment and the air-handling plant (CRAC, CRAH, in-row, fan-wall), how warm you run the supply air (the ASHRAE envelope as an economic dial), and how far you can push a rack before air loses on cost. A last section covers the load that stays on air. The heat-flux physics and the cooling hierarchy are in Chapter 5.1; this chapter covers the air side of that wall.

Cold air goes under the IT or over it, a choice that predates AI and shapes everything after it. The raised access floor (perforated tiles over a pressurized plenum, the iconic 600 mm void) was the default for thirty years because it hid cabling and let one CRAC pressurize the whole room. Cold air escapes through tile cutouts and gaps without passing through any rack, and the floor assembly caps the weight it can carry. A loaded 48U NVL72 rack is heavy. HPE lists the GB200 NVL72 at 1,472 kg / 3,245 lb and the GB300 NVL72 at about 1,497 kg / 3,300 lb fully loaded (HPE QuickSpecs, September 2026), and NVIDIA puts Vera Rubin NVL72 at roughly 1,800 kg / 4,000 lb (March 2026, coolant fill not specified). That is before the CDU or the row's charged pipe. That weight, along with airflow needs, is why many dense AI halls use a slab.

Slab-on-grade with overhead air distribution (supply ducted down from above or pushed by perimeter or in-row units, return through a ceiling plenum) takes the access-floor assembly out of the support path. It also avoids the underfloor obstruction that upsets plenum pressure once liquid pipe and busway fill the void. The slab and the move route still need the reactions and load combinations in Chapter 6.2. Both schemes use fans, pressure control and airflow balancing, so the slab does not add powered air movement. What it removes is the plenum as a free distribution path: you design a supply and return path unit by unit. Choose on the constraints you can demonstrate: the floor's load path, plenum capacity, move route and service access. A greenfield hall that will carry liquid distribution often lands on slab, and a raised floor full of coolant manifolds hides leaks you would rather see. An existing raised floor does not by itself cap a brownfield retrofit; those checks decide, a thread picked up in Chapter 5.10.

Before you buy more cooling, find out what mixing between hot and cold air is costing you. In an uncontained room, supply and return blend, you over-cool to compensate, and your fans and chillers work against your own bypass. Containment seals either the cold aisle (a roof and end doors over the supply aisle) or the hot aisle (capturing the exhaust and ducting it back). It reduces bypass and recirculation, which can allow warmer supply air or lower fan flow, as long as the worst-served rack inlet stays within its allowed range. The two differ in practice. Cold-aisle containment keeps the working aisle comfortable and is cheaper to retrofit, but it bakes the rest of the floor, and the people on it, in return air. Hot-aisle containment keeps the room at supply temperature and pipes the hot exhaust away. It suits warm-supply operation and mixed liquid/air halls, but costs more and complicates the ceiling. If the hall will take rear-door heat exchangers or a liquid retrofit, choose containment for the exhaust that remains, the door outlet temperature, and the maintenance and failure case. A room-neutral rear door can remove the need for hot-aisle containment, while a partially capturing door or a hybrid rack still needs a path for the remaining room heat. Chapter 5.3 covers door selection.

Next is the plant that moves and cools the air. A CRAC (computer-room air conditioner) is a packaged DX unit with its own compressor and refrigerant. A CRAH (computer-room air handler) has only a chilled-water coil and fans, with refrigeration done remotely by a chiller plant. CRAH is the hyperscale default because chilled water separates heat rejection from the room and rides the economizer. CRAC survives where there is no chilled-water plant: smaller rooms, edge sites and some retrofits. Where the cooling sits changes the air path. In-row coolers sit between the racks, where they can shorten the air path and reduce mixing. Vertiv, for example, rates its Liebert CRV chilled-water in-row units at 27.1, 46.3 and 53.2 kW of net sensible cooling each (CR032, CR040 and CR050), with return air reaching the coil from the hot aisle at 35 °C (95 °F). The CR050 figure also assumes warmer entering water than the smaller two (Vertiv, May 2025). How many racks one unit serves depends on airflow distribution, containment and the redundancy case. Fan-wall (thermal-wall) units use a high-efficiency fan array and can deliver across the room from the perimeter, from a service corridor, or from a close-coupled position. Do not rank these by distance alone. Compare the rack inlet conditions each candidate delivers at normal and fault duty, and its full fan electrical input: airflow times fan pressure rise, divided by the combined fan, motor and drive efficiency. The remaining tradeoffs are return path, containment, floor space and service, plus putting water nearer the IT with CRAH, in-row and fan-wall units.

Containment and plant selection set the ceiling; airflow management decides how much of it you reach. The failure modes are mundane and they compound. Bypass is cold air that returns to the cooler without doing work: leaking tiles, gaps under racks, misplaced perforated tiles in a hot aisle. Recirculation is hot exhaust looping back to the inlet over the top or around the ends of a rack. Open U-space, such as a missing blanking panel, lets the two streams short-circuit straight through the rack. Each forces you to lower supply temperature to protect the worst inlet, which spends chiller energy and erases the free-cooling hours you meant to bank. Blanking panels, brush grommets, floor-leakage sealing and correct tile placement fix these. Rank the fixes by the inlet and flow losses you measure and the energy each saves against its installed cost.

The metric that captures this is the air-delivery temperature rise: the gap between terminal supply and equipment inlet, distinct from the exchanger approach in 5.1. A well-managed contained aisle holds that gap to a few degrees. A leaky uncontained room can lose 5–10 °C to mixing, which you must claw back at the chiller. Delivery, pressure, containment and recirculation decide how much of the coil's rated capacity reaches the racks, as much as the coil schedule does. Approach and NTU/effectiveness are defined in Chapter 5.1; live-floor instrumentation of inlet temperatures, ΔT and bypass belongs to operations in Chapter 14.2.

Supply-air temperature is the dial most operators under-exploit. ASHRAE TC 9.9 sets a recommended inlet band of 18–27 °C for classes A1–A4. The high-density H1 class is narrower, at 18–22 °C recommended and 15–25 °C allowable, and the equipment OEM decides which class applies. The allowable envelopes widen by class: A1 15–32 °C, A2 10–35 °C, A3 5–40 °C, A4 5–45 °C. Warmer supply air is a lever, not a concession. It can permit warmer chilled water, which cuts compressor lift at a given duty and widens the hours the plant can run on free / economizer cooling, where the climate and plant allow. Condenser temperature is a separate optimization against rejection energy and equipment limits, and warmer supply does not raise it. A hall designed and run at A2–A3 in a temperate climate can gain hundreds to thousands of economizer hours a year over one held to the recommended band, which lowers mechanical PUE materially. For scale, Uptime's 2026 survey puts the average annual PUE at 1.52 per facility and 1.36 weighted by capacity, because larger, newer sites pull the weighted figure down. Neither number is an air-cooled benchmark (Uptime Intelligence, August 2026).

Warmth costs you in three ways. First, server fan power: above the server's fan-control knee, on-board fans ramp non-linearly, and past a point the rack's own fans eat the chiller savings. Roughly 25–27 °C inlet is a fair first guess for where that knee sits; the measured fan curve tells you where it actually is. Total energy against supply temperature is a curve with a minimum, so warmer stops paying past a point. Second, reliability margin: warmer silicon fails more often and leaves less headroom for a cooling excursion. That matters more for synchronous training, where a thermal trip that throttles or drops a node restarts the job, than for stateless inference. Third, acoustics and the human floor: ramped fans are loud, and a hot aisle in an A3 hall is a workplace-safety question. The right answer depends on climate and workload. A cool-climate inference hall can buy more free-cooling hours by running as warm as its total-energy minimum permits. A tropical training hall has fewer such hours and gives up trip margin it needs when a dropped node can restart the job. Either way, stay inside the A-class or H1 envelope with fault and service margin. Chapter 5.8 covers how supply temperature shapes plant selection and economizer mode, and Chapter 15.1 is canonical for how PUE and ITUE are defined and reported.

Air cooling has no universal rack-kW ceiling. SemiAnalysis (February 2025) reports 30–40 kW typical RDHx capacity and more than 50 kW with active rear-door fans, which are door-specific reference conditions, not a product-independent boundary. A perimeter-cooled raised-floor room can carry only the load that its delivered airflow, plenum pressure and rack-inlet temperatures support. Close-coupled air with full containment and disciplined airflow management can carry materially higher duties at the conditions the system is rated for. Air often gives way for economic and operational reasons rather than at a single physical cutoff. As airflow demand rises, the fan energy per kilowatt removed climbs until the marginal kilowatt costs more in fan power and floor space than it would on a rear-door heat exchanger or a direct-to-chip cold plate.

The cooling methods overlap, so the same rack can have more than one valid answer. For every candidate, record the rack's heat split and flux, OEM airflow and inlet limits, TCS/FWS or entering-water conditions, residual-room rejection, design climate, service and redundancy case, and next planned refresh profile. Then compare the candidates that pass on modeled fan and plant energy, floor area, water or refrigerant interfaces, service, redundancy and expansion cost. HPE's own figures settle the GB200 NVL72: roughly 115 kW to liquid and 17 kW to residual air. Subtracting a generic air limit from ~132 kW does not. A design is stranded when it cannot support the next planned rack profile, whichever rack-kW band that profile falls in.

Workload labels help you locate likely equipment, but they do not select cooling. Treat the list below as starting hypotheses, and run each rack through the candidate check set out above.

• Networking and storage. Many current profiles stay on air, but not all. Check each product's airflow direction, inlet limits, fan power, acoustics and room duty instead of inferring suitability from the category.
• Inference fleets. Air, RDHx/AALC, hybrid and DLC can all occur at the same rack kW. Model, precision, batching, server packaging, OEM heat split and the site's thermal limits decide the candidate, not the word inference.
• Edge and micro-sites. Air or sealed/modular systems are common, but ambient range, filtration, acoustics, rejection, remote service, redundancy and equipment inlet limits must hold at the specific site.
• Batch or curtailable work in an inherited hall. Retry tolerance changes the outage consequence, not heat removal. Keep the existing air service only where the rack and its roadmap profile fit what that air system is rated to deliver.

Split zones by equipment and facility requirements, not by labels like permanent-air or not-yet-liquid.

Three air-side mis-scopes recur, each from reasoning about the room instead of the load:

• Using rack kW as the selector. A 45 kW air profile, a 45 kW cold-plate/AALC profile and a 45 kW rack that fits neither can all exist. Record the rack's heat split, airflow and inlet limits and water conditions before choosing.
• One zone for incompatible equipment. Draw zone boundaries where needs differ materially (inlet, airflow, heat split, water, rejection, service, refresh roadmap, fault response), not along workload labels.
• Chasing the widest ASHRAE envelope as a setpoint. Running A3/A4 to bank free-cooling hours, then watching server fans and acoustics eat the savings and the reliability margin past the fan-power knee.

Equipment airflow = 80.0/(1.20 × 1.005 × 12.0) = 5.53 m³/s. Terminal airflow = 5.53/0.900 = 6.14 m³/s, using the unrounded equipment flow. Fan input = 6.14 m³/s × 300 Pa / 0.600 = 3.07 kW. This terminal input is separate from server fans already inside the IT meter. If you start from a vendor's external-static schedule, add the unit's internal losses first and pair the fan pressure with its matching efficiency.

The allowed indicated inlet is 27.0 − 1.0 − 0.2 = 25.8 °C. The normal and unavailable-terminal cases pass; open containment at 26.3 °C fails by 0.5 K. Restore containment or revise delivery before accepting that service state. The crossover is an indicated 25.8 °C, not a rack-kW band. Require the supplier's fan curve and inlet map to show these duties before release. ASHRAE's data-center environmental guidance gives the inlet-envelope context; use 13.5 for installed acceptance.

Choose the terminal and containment arrangement that delivers your servers' inlet envelope in normal operation and in the required service state. If containment fails, you lose capacity or must cut load; a fan's nameplate airflow does not recover air that bypasses the racks.