AI and GPU rack power and cooling readiness field guide

What AI and GPU rack readiness is

AI and GPU rack readiness is the assessment and the work that prove a space can take a high-density rack before the rack arrives. Readiness covers four things at once: the power delivered to the rack, the cooling that pulls the heat back out, the floor that carries the weight, and the network fabric that ties the accelerators together. When all four are sized for the real density and recorded against the design, the space is ready and the GPUs can land.

This is its own discipline because of the gap between a legacy rack and an AI rack. A general-purpose compute rack ran somewhere around 5 to 12 kW for years, and a room built for that has feeders, cooling, and a floor rated for it. An AI training rack pulls 40 kW at the low end and 120 to past 130 kW at the top of the current generation, with the next chip generation already quoted at 180 to 220 kW per rack. Drop that load into a room built for 10 kW and no part of it holds.

So readiness is not a punch list you run after the rack is in. It is the gate you clear before the rack ships. Every shortfall, a feeder that is too small, a floor that is not rated, a chilled water plant with no spare tons, is cheap to find on paper and brutal to fix with hardware in the room.

Why does one AI rack change the whole room?

One AI rack changes the room because the load it presents is roughly ten times the legacy density, and every system in the building was sized for the old number. The power, the cooling, the structure, and the network were all balanced around a rack that drew 5 to 10 kW. The AI rack breaks all four balances at the same time.

Start with heat, because heat follows power almost one for one. A 130 kW rack rejects about 130 kW of heat into the room, and the air handlers, the chilled water plant, and the heat rejection outside were never sized to move that from a single cabinet. Then the weight: the same dense rack can run to 3,000 lb, which a raised floor or even a slab may not be rated to carry as a point load. Then the power feed, which now needs hundreds of amps to one cabinet instead of one row. Then the network, where the GPUs talk to each other over a high-speed fabric that demands far more fiber than a legacy top-of-rack switch ever pulled.

The lesson crews learn the hard way is that you cannot fix one pillar in isolation. Add the power and the cooling plant is still short. Add the cooling and the floor still cannot take the weight. Readiness treats the four as one problem, because the rack does.

How much power does an AI rack use?

An AI rack uses roughly 40 kW at the low end and 120 to over 130 kW at the top of the current generation, against the 5 to 12 kW a legacy general-purpose rack drew. A current top-end rack such as the GB200 NVL72 class is specified around 120 kW and has been measured drawing 130 to 132 kW under full load, while an air-cooled H100 rack tops out near 40 kW. The 2026 generation is quoted around 180 to 220 kW per rack, and the generation after it, in 2027, is quoted higher still, on the order of 600 kW per rack.

Two numbers frame the jump. The fleet average rack density is still in the single digits, on the order of 8 kW in 2025 and climbing, and most operators still run no rack above 30 kW, even as new AI rows run far higher, while a room designed a decade ago was planned around 5 to 10 kW. The AI rack is not an increment on that. It is a different class of load on a single floor tile.

Use the manufacturer's rack power spec as the design number, not a rule of thumb, because the spread between nominal and measured peak is real and it lands on your feeder and your cooling. Size the readiness around the peak the rack actually pulls under training load, and confirm it against the current edition of the manufacturer's site planning document.

Why do AI racks need liquid cooling?

AI racks need liquid cooling because air runs into a physical wall somewhere around 30 to 50 kW per rack, and the AI rack sits far past it. You can push a well-designed air rack with containment to about 35 to 50 kW, but above that the volume of air you would have to move through the cabinet is more than the fans and the floor can deliver, and the chip heat flux climbs past what an air-cooled heat sink can shed.

The practical thresholds the industry uses are worth carrying. Above roughly 20 kW per rack, direct-to-chip liquid cooling becomes the minimum sensible approach. Above about 50 kW it is no longer optional, and above 100 kW the design moves to dense direct-to-chip or immersion. A 130 kW rack is liquid-cooled, full stop, with at most a small fraction of the load left on air for the parts of the chassis the cold plates do not reach.

Most current AI deployments are hybrid: direct-to-chip liquid takes the bulk of the heat off the CPUs and GPUs through cold plates, and a residual air load, often 10 to 20 percent of the rack, still goes to the room or a rear-door heat exchanger. The commissioning of that liquid loop is its own job. The liquid cooling loop commissioning guide covers the flush, the leak strategy, the flow balancing, and the coolant spec, so this guide stays on the readiness around it rather than repeating it.

Liquid readiness: facility water, the CDU, and leak detection

Liquid readiness means the building can supply, condition, and contain the coolant the racks need before any rack is plumbed. The core pieces are the facility water system that the chiller plant produces, the coolant distribution unit that isolates the clean rack loop from that facility water, the rack manifold that splits coolant to each node, and the leak detection that catches a drip before it reaches a live cabinet.

The facility-water capacity is the first thing to size and the easiest to underestimate. The CDU only moves heat from the secondary loop to the facility water as fast as the facility water can carry it away, so a hall full of liquid-cooled racks needs the chilled water flow and the supply temperature the CDUs were rated against. ASHRAE TC 9.9 classifies facility water supply temperature in W-classes, from W17 through W+ in degrees Celsius, and the warmer classes are what make economizer hours and heat reuse possible. Pick the class the CDU and the cold plates are rated for and confirm the plant can hold it.

Leak detection has to be commissioned and proven, not just installed, because a leak onto an electronics cabinet of GPUs is a six- or seven-figure event. The plumbing, the flush, the hydrotest, the flow balance, and the leak-detection proof live in the liquid cooling loop commissioning guide. For readiness, the gate is simpler: the facility water capacity, the CDU placement and rating, the manifold and quick-disconnect plan, and a leak-detection scheme are all designed and confirmed before the racks are scheduled.

Power delivery to the rack

Power delivery for an AI rack means getting hundreds of amps to a single cabinet, redundantly, without cooking the busway or the neutral. A 130 kW rack at 415 V three-phase pulls on the order of 180 A per cabinet, and a row of ten racks at 40 kW already draws over 500 A from the overhead busway. New AI rows are being planned around 800 A to 1,600 A busway where older halls ran 400 A, so the busway, the taps, the whips, and the rack PDUs all step up together.

The distribution chain is the same shape as a legacy hall but rated far higher. Medium-voltage power steps down to 415 V or 480 V three-phase, feeds the UPS, and runs out to the hall on busway. From the busway, a tap and a whip land at the rack, into a remote power panel or rack PDU sized for the cabinet's full draw on both an A and a B feed. Dual-corded redundancy still applies, but now each side carries a much larger load, so the redundancy math has to account for one side picking up the whole rack on a failure.

Two details bite on AI power. GPU and accelerator power supplies draw heavy harmonic current, so neutrals and metering have to be sized and rated for it rather than assumed. And the industry is moving toward higher-voltage distribution, including 800 VDC architectures from several major vendors, to carry the current without unworkable copper. The power-density and feeder-sizing topics are covered in the electrical guides; here the readiness gate is that the busway amperage, the A and B feeds, the PDU or RPP capacity, and the redundancy scheme are confirmed against the rack's measured peak.

How heavy is a GPU rack, and will the floor hold it?

A fully built AI rack runs heavy, on the order of 3,000 lb (about 1.36 metric tons) for a current top-end rack, and that weight concentrates onto a footprint smaller than a square meter. That turns into a point load near 1,800 kg per square meter, which is well past what many existing data center floors were designed to carry.

The number that catches teams is the floor rating they already have. Plenty of existing halls were built to a floor loading limit around 2,000 to 2,500 lb per rack, and a raised-floor tile is commonly rated near 250 kg before you even count the rolling load of moving the rack into place. A 3,000 lb cabinet rolled across a floor rated for 2,500 lb static can crack a tile under the caster, or settle and stress the structure under the standing load weeks later.

So the weight check is a structural question, not a logistics one. Confirm the static load, the rolling load along the path from the dock to the spot, and the point load at the feet against the rated capacity of the raised floor or the slab, and reinforce or pour a dedicated slab where the rack exceeds it. The rack readiness and floor load guide covers the layout, leveling, anchoring, and floor-rating verification in detail; for AI readiness the point is blunt. Confirm the floor before the rack ships, because you cannot move a 3,000 lb cabinet off a cracked tile easily once it is loaded.

The network fabric and cabling density

AI racks change the network as much as they change the power, because the accelerators have to talk to each other at very high speed to train as one machine. That means a high-bandwidth GPU interconnect inside and between racks, a separate high-speed Ethernet or InfiniBand fabric across the cluster, and a fiber count per rack far above what a legacy top-of-rack switch ever needed.

The readiness issue is physical: pathway, fiber, and the structured cabling have to be in before the racks, because retrofitting hundreds of fibers into a live, liquid-cooled hot aisle is miserable. The cable trays, the fiber runs, the patch fields, and the optics all scale with the GPU count, and the cabling competes for the same overhead space as the busway and the liquid manifolds. Coordinate those three in the same overhead section early or they collide.

Plan the fabric topology with the compute team, not after them. The cluster's interconnect design drives how racks are grouped, how long the cross-rack links can be, and where the spine switches sit, and that in turn drives the rack pitch and the pathway. The detailed cabling and pathway work is its own topic; for readiness, the fabric and the fiber density are part of the gate alongside power and cooling, not an afterthought once the racks are placed.

Space, rack pitch, and CDU placement

The layout for AI racks has to make room for the liquid gear and the service access the density demands, which usually means fewer racks in more space, not more racks in less. The rack pitch widens to fit the manifolds, the heavier cabinets, and the work room a tech needs behind a loaded liquid rack. The aisles still run hot and cold for the residual air load, and containment still matters for the fraction of heat that stays on air.

CDU placement is the new variable. An in-row CDU eats a rack position and has to sit close enough to the racks it serves to hold the loop temperature and pressure, while a perimeter or external CDU frees floor space but lengthens the secondary loop. Either way the CDU needs its own service clearance, its own facility-water connection, and a path for a leak to drain somewhere safe.

Service clearance is where retrofits get tight. A legacy hall laid out for 10 kW racks on a 4 ft cold aisle may not have the rear clearance to pull a liquid node, swap a quick-disconnect, or roll a 3,000 lb cabinet into place. Walk the actual move path and the actual service envelope before you commit the layout, because the drawing that fits on paper often does not fit a rack on a pallet jack.

Does the room have the electrical headroom?

Before any of the rack-level work, the question is whether the room, the feeders, the UPS, and the generator have the spare capacity for the new density at all. An AI deployment can multiply a hall's connected load several times over, and the upstream gear, the switchgear, the UPS modules, the generators, the utility service, may simply not have the headroom. That is a load study, not a guess.

Run the study against usable capacity, not nameplate. A UPS rated for a number on the label may be carrying existing load, reserving redundancy, and derated for its environment, so the headroom you can actually commit to new racks is smaller than the sticker. The same goes for the generator plant and the utility feed. Stranded capacity, power that exists somewhere in the chain but cannot reach the new racks because a transformer or a breaker or a busway in between is the bottleneck, is the thing that quietly kills AI retrofits.

This is where the building's monitoring earns its keep. An electrical power monitoring system shows the real load on each feeder and UPS over time, and a DCIM platform ties that to rack-level capacity, so you can see usable versus stranded headroom instead of arguing from drawings. Those monitoring topics are covered elsewhere; the readiness point is that the capacity assessment uses measured load and usable capacity, and it happens before anyone commits a rack count.

Cooling capacity and heat rejection

The cooling-capacity assessment asks whether the chilled water plant and the heat rejection outside can absorb the new heat, because every kilowatt into the racks comes back out as heat the plant has to move. A hall adding several megawatts of AI load is adding several megawatts of heat, and the chillers, the cooling towers or dry coolers, the pumps, and the piping were sized for the old number.

Convert the load to the plant's language and check it honestly. Cooling capacity is often counted in tons, where one ton is about 3.517 kW, so a single 130 kW rack is roughly 37 tons of rejection on its own, and a row of them is a plant-scale number. Confirm the chiller plant has the spare tons at the design wet-bulb, the pumps have the flow, and the piping has the capacity to the CDUs, at the worst-case outdoor condition, not the average day.

Many existing rooms need a plant upgrade before they can host AI at scale, and that upgrade has a long lead time, often longer than the racks. The chillers, the towers, the pumps, and the structural and electrical work to support them can run a year or more. If the cooling plant is short, that is the schedule driver for the whole project, so size it first and order it early.

Can you put AI racks in an existing data center?

You can put AI racks in an existing data center, but usually only at lower density, in a limited zone, and after upgrades, not by sliding them into the open white space. The limits are the floor rating, the available power and cooling capacity, the lack of a facility-water loop, and the service clearance the layout never planned for. Each of those can be worked, but together they cap how far a retrofit goes.

The honest version is a spectrum. A legacy hall can often host a handful of moderate-density liquid racks with a liquid-to-air CDU, no facility-water plant, and a reinforced floor zone, which suits a pilot or an inference cluster. Pushing the same hall to rows of 130 kW racks means a new chilled-water plant, new busway, structural floor work, and a fabric build, at which point you are rebuilding the room around the racks.

A purpose-built AI hall starts from the density: high-capacity busway, a facility-water loop sized for liquid, a floor rated for the weight, wide pitch, and the fabric pathway designed in. It costs more up front and it is the cheaper path at scale, because the retrofit hits a wall where the next rack needs a plant you do not have. Decide which one you are doing before the first rack, because the retrofit that quietly turns into a rebuild is the most expensive way to learn the difference.

Commissioning the high-density deployment

Commissioning a high-density deployment proves the power and cooling carry the real load and heat before the GPUs do, because the failure modes are fast and expensive enough that you do not want to discover them under live training. The work runs from component checks up to an integrated systems test, where load banks stand in for the racks and the whole power and cooling chain is driven to design load at once.

Heat is the part that has to be proven, not assumed. Use load banks or heater loads to put the rated kilowatts on the cooling and watch the chip-supply temperature, the CDU approach, the facility-water return, and the plant all hold at design. Run the integrated systems test through a utility-loss and generator-start sequence with the load on, because that is the moment the cooling has the least margin and the most to prove.

The liquid loop has its own commissioning that comes first: the flush to a cleanliness target, the hydrotest, the leak-detection proof, and the per-rack flow balance, all covered in the liquid cooling loop commissioning guide. The rack-level readiness, the bonding, the A and B power, the containment, sits in the rack readiness and floor load guide. The high-density commissioning ties them together and proves the assembled system holds load before the hardware that cannot tolerate a miss shows up.

Redundancy and the fast failure

High-density failure is fast, and that changes how you design redundancy. A direct-to-chip loop at full load has very little thermal ride-through, on the order of seconds, because there is almost no stored coolant volume buffering the chip. Lose the pump or the flow and a 100 kW rack can drive components from safe to critically hot in tens of seconds, with thermal shutdown inside about a minute.

Compare that to a legacy air room, where the thermal mass of the air and the slab bought you minutes to react to a cooling hit. That cushion is gone. The cooling cannot drop, even briefly, so the cooling becomes an uptime-critical load that rides the UPS alongside the IT gear, not a mechanical system you can let coast through a transfer. The pumps, the CDU, and the controls have to stay running through a power event, because the racks will not wait for the generator to come up.

The practical answer the industry has landed on is real redundancy on cooling and power, commonly 2N or N+1 on the critical path, and the willingness to accept stranded capacity as the price of not losing a hall of GPUs. The CPU sets a junction-temperature limit, often in the 85 to 92 degree C range, where it throttles to protect itself, and throttling a training cluster is a real cost before you ever reach hardware damage. Design so the cooling never gets close to that line.

PUE and heat reuse

Liquid cooling at high density is more efficient than air, and it opens a door air never could: usable heat recovery. Direct-to-chip loops let the facility water run warm, in the warmer ASHRAE W-classes, which cuts or eliminates mechanical chilling hours and pulls the facility PUE down toward the low end. The same warm water is hot enough to be worth capturing.

Heat reuse is where the warm-water classes pay off. ASHRAE TC 9.9 flags the higher facility-water supply classes, W32 and above, as candidates for district heating, because water leaving the racks at those temperatures is useful to a heating loop next door instead of being thrown to the sky. Capturing it takes a heat-recovery design, a host for the heat, and a plant that can run warm, so it is a design decision made early, not a bolt-on.

Set the water temperature target with the cold-plate rating, the PUE goal, and any heat-reuse host together, because they pull in the same direction. Warmer water means better PUE and more reusable heat, up to the limit the cold plates and the chips will tolerate. The manufacturer's coolant-temperature spec is the ceiling on how warm you can run, so design to it and confirm the plant can hold it at the worst-case outdoor day.

Phased deployment and the readiness gate

AI capacity usually lands in phases, a few rows at a time, and each phase needs its own readiness gate before its racks ship. The gate is a hold point: the GPUs do not get scheduled until the power, the cooling, the floor, the network, and the leak detection for that zone are confirmed and signed against the design. The phasing lets the plant and the busway grow with the load instead of all at once, but only if each phase is fully ready before it loads.

Phasing also protects the schedule from the long-lead items. The chiller plant, the busway, and the structural floor work all run ahead of the racks, so the phase plan has to sequence those upgrades to be done and commissioned before the racks they serve arrive. A phase that ships racks into a zone whose plant capacity is not yet online is the classic AI deployment failure: expensive hardware sitting idle, or worse, running throttled or at risk.

Keep the gate concrete and per-zone. For each phase, confirm the rack power feed and redundancy, the liquid loop commissioned and leak-proven, the floor rated and reinforced, the fabric and fiber in, the cooling capacity online, and the monitoring live. When that is true and recorded, the zone is ready and the hardware can land. Until it is, the gate holds.

Monitoring: inlet, power, and leaks

Monitoring for AI density has to be live before the racks load, because the failure window is too short to catch by walking the floor. Three signals carry the weight: the rack inlet and chip-supply temperatures, the per-rack and per-feeder power, and the leak detection on the liquid loop. All three need to alarm fast and route to someone who can act in the seconds the density allows.

Tie the signals into the building's platforms rather than running them as islands. A DCIM platform tracks rack-level power and thermal against capacity, an electrical power monitoring system watches the feeders, busway, and UPS for the real load and the harmonics the GPU supplies inject, and the leak-detection system shrouds the CDU, the manifolds, and the under-floor or overhead piping. Those monitoring topics are covered elsewhere; the readiness point is that they are commissioned and proven, not just wired.

Test the alarms, do not assume them. Trip a leak sensor with a wet probe, drive a thermal alarm with a load bank, and confirm the path from sensor to operator actually fires and actually reaches a human. An unproven leak sensor is worse than none, because the room runs trusting an alarm that was never confirmed to work.

What to document per rack

The readiness record is what lets the owner sign the gate and what answers the question later when a rack runs hot or trips. Capture it per rack, keyed to the rack's position, so the power feed, the cooling assignment, the floor check, and the network all tie back to one cabinet.

For each rack, record the model and ID, the design and measured power draw with the feed details, the cooling type and loop assignment, the weight against the floor rating at that location, the power feed and redundancy, and the network fabric and fiber count. Record who verified each item and against which design document, because the next person needs to know the floor was checked at this spot, not in general.

Common mistakes

• Assuming air, or a little extra air, can cool a 40 kW or higher rack instead of designing for liquid.
• Sizing the facility water or chilled water plant short, so the CDUs cannot reject the rack heat at the worst-case day.
• Rolling or standing a 3,000 lb rack on a floor rated for 2,000 to 2,500 lb without checking the point and rolling loads.
• Sizing the feeder, busway, or UPS to nameplate instead of measured usable capacity, and missing stranded capacity.
• Treating the liquid loop's leak detection as installed rather than commissioned and proven before the racks load.
• Skipping the readiness gate and shipping GPUs into a zone whose plant or floor upgrade is not yet online.
• Forgetting the harmonics and oversized neutral that GPU power supplies demand on the busway and metering.

Field readiness checklist

Standards and references

ASHRAE TC 9.9 is the thermal authority for this work. Its thermal guidelines, now in the fifth edition with an expanded liquid-cooling chapter, define the facility water temperature W-classes, the water quality classes, and the surface temperature classes that the CDU, the cold plates, and the plant all reference. Use the W-class your equipment is rated for, and let the manufacturer's coolant and temperature spec set the ceiling.

The Open Compute Project covers the high-density and liquid side by topic. The OCP Cooling Environments project and its Cold Plate sub-project drive standardization of the technology cooling system from the cold plate to the CDU, and the door heat exchanger work standardizes rear-door heat exchangers within the ORv3 rack framework, including the blind-mate manifold interfaces. Where a deployment follows OCP rack and cooling conventions, those documents define the interfaces.

For the facility tier and topology, the Uptime Institute Tier classification and TIA-942 cover redundancy and infrastructure expectations by topic, and they frame the 2N or N+1 decisions on the critical cooling and power path. Above all of these sits the manufacturer's site planning and rack specification and the project specification. Where the manufacturer or the project calls a tighter number, that number governs. Standards and editions change, so confirm the current edition and any project amendments before citing a specific class or requirement on a submittal.

Units, terms, and conversions

AI readiness mixes electrical, mechanical, and structural units across power, cooling, and weight, so the same load reads differently on a power one-line, a mechanical schedule, and a structural drawing.

Rack power is in kW per rack, and chip power is in watts of TDP. Cooling capacity is in tons or kW, where one ton of refrigeration is about 3.517 kW (12,000 BTU per hour), so a 130 kW rack is roughly 37 tons. Weight crosses metric and imperial: one metric ton is about 2,205 lb, and floor capacity is rated in lb per square foot or kg per square meter as a point or distributed load. Facility water supply temperature follows the ASHRAE W-class in degrees Celsius.

kW per rack

Power drawn by a single cabinet, the design number for feeder, busway, and cooling sizing

TDP

Thermal design power, the watts a chip is rated to dissipate, which the cooling has to remove

DLC / direct-to-chip

Direct liquid cooling, coolant through cold plates clamped to the CPUs and GPUs

CDU

Coolant distribution unit, isolates the clean rack loop from facility water and sets its flow, pressure, and temperature

W-class

ASHRAE TC 9.9 facility water supply temperature class, W17 through W+, the number being the max supply temperature in degrees Celsius

PUE

Power usage effectiveness, total facility power divided by IT power, lower is more efficient

Ton of cooling

About 3.517 kW or 12,000 BTU per hour of heat rejection

2N / N+1

Redundancy levels on the critical power and cooling path, full duplicate versus one spare unit

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• What is direct-to-chip liquid cooling?
• What is rack readiness in a data center?
• What is a CDU in a data center?
• What is direct-to-chip liquid cooling?
• Why do data centers need liquid-cooling leak detection?