800 VDC AI Data Centers: What Really Changes for Chillers, CDUs and Racks

November 05, 2026 | Blog

Part A — 800 VDC Cooling Infrastructure at a Glance

AI data centers are moving toward 800 VDC power distribution because rack power is climbing from tens of kilowatts toward hundreds of kilowatts and, on industry roadmaps, toward a megawatt per rack. At those levels, low-voltage AC and 54 V busbars run into hard limits on current, copper and conversion losses. Raising the distribution voltage cuts the current for a given power and frees space inside the rack.

What 800 VDC does not change is the heat. Every watt delivered to a rack still ends up as a watt of heat that the cooling plant must remove, all day, every day, including during utility outages, generator transfers and equipment failures. The cooling system — chillers or dry coolers, the facility water system, coolant distribution units (CDUs), the technology cooling loop, rack manifolds and cold plates — still has to be sized for heat load, coolant temperatures, flow, pressure and reliability.

So the practical question owners and design teams ask is: when the IT power moves to 800 VDC, what actually changes for the chillers, CDUs and racks?

Keentel Engineering's short answer:

800 VDC changes how power reaches the rack; it does not change how much heat must leave it. Chillers and CDUs can stay on AC power, and most will. What changes is the interface: rack-level DC protection and isolation next to liquid connections, the power architecture behind pumps and controls, the control link between cooling health and GPU workload, and — only where an owner chooses direct-DC mechanical equipment — new drive, precharge, discharge and insulation-monitoring requirements.

This guide works through that interface in US engineering terms. It covers the electrical options for chillers and CDUs, what really happens when a variable-frequency drive (VFD) is fed from an 800 V bus, how to size the technology cooling loop and CDU heat exchanger, how pump redundancy and power architecture interact, how to keep water away from 800 V conductors, and how to link cooling-loss detection to GPU power capping. Worked examples, tables and Keentel's own analysis make the points concrete.

Companion Keentel articles cover the 800 VDC UPS and battery architecture, DC fault protection with solid-state breakers, CDU fundamentals and digital-twin analysis of pump failures. This article focuses on what sits between them: the mechanical–electrical interface.

The installation used in this guide

Keentel Engineering uses one illustrative US installation throughout.

What Keentel's analysis shows

Figure 1. Left: technology cooling system flow needed for a 1 MW rack as a function of the share of heat captured by liquid, for water and PG25 at 8, 10 and 12 K temperature rise. Right: Keentel's model of a 10 MW hall after a chiller trip with a seven-minute restart, showing facility water supply temperature with only the normal system water volume and with a 60,000-gallon buffer tank, against a GPU power-cap threshold set 5 K above the supply setpoint.

Five results stand out:

• Inertia changes the path, not the settling point. Cases A and B settle at the same Value B, because settling frequency depends on droop and load damping, not inertia. But case B falls 2.5 times faster and reaches a nadir below the first load-shedding stage.
• Fast frequency response rescues the nadir. 1,000 MW of battery response, at full output within 0.5 s of a 59.7 Hz trigger, lifts the low-inertia nadir by 0.2 Hz — back above the load-shedding threshold — and raises Value B as well.
• Primary response leaves an offset; it does not restore 60 Hz. With 2,750 MW lost and about 8,300 MW/Hz of combined response, frequency settles about 0.36 Hz low (including the deadband effect). Only AGC removes that offset.
• Recovery happens while the shortfall persists. In every case the failed unit stays offline. Frequency returns to 60 Hz because other units' output has been raised by 2,750 MW — first by governors, then by AGC and reserve deployment.
• The clock is set by standards. In this model AGC restores frequency in under nine minutes, inside the 15-minute Contingency Event Recovery Period in NERC BAL-002-3. A balancing authority that lacks reserves cannot meet that clock, regardless of how good its governors are.

Engineering review of common statements about frequency recovery

Five takeaways for owners and design teams

• Keep the cooling plant on AC unless there is a clear reason not to. Direct-DC mechanical equipment is emerging and needs new drive, protection and safety work.
• Size the liquid loop for the real fluid and the real capture fraction. Glycol adds about 5% to flow; a lower capture fraction can add several hundred kilowatts of air-side heat per megawatt rack.
• Put CDU pumps, controls and leak detection on uninterruptible power. Without flow, a megawatt-class rack has seconds, not minutes, before its GPUs must throttle.
• Design the outage timeline, not just the steady state. Chiller restart, generator transfer and pump changeover must fit inside the thermal ride-through that the water system provides — or be covered by GPU power capping.
• Separate water from 800 V by design, with drip paths, connector placement, insulation monitoring and leak-triggered isolation in the rack specification.
  

Part B — Technical Guide: The Mechanical–Electrical Interface of an 800 VDC AI Data Center

1. What 800 VDC changes — and what it does not

1.1 Current and power



For DC distribution the current is simply power divided by voltage:

The same rack fed at 54 V would draw about 18,500 A; at 415Y/240 V three-phase AC it would draw about 1,390 A per phase at unity power factor. The 800 V bus is attractive because it carries megawatt-class rack power with manageable conductors, fewer conversion stages and less copper. Keentel's companion article on 800 VDC UPS architecture covers the efficiency and storage arguments in depth.

1.2 Heat

The heat that must leave the rack is the electrical power that enters it:

Two details matter. First, the rack input already includes the losses of the rack's own DC/DC converters and power shelves. If those converters are 97.5% efficient, about 25 kW of every megawatt is converter heat — heat that lands on a cold plate only if the power shelf is liquid-cooled. Second, upstream conversion losses (rectifiers, solid-state transformers, busway) are released in electrical rooms, not in the rack, and need their own cooling. A higher distribution voltage reduces those upstream losses a little; it does not change the IT heat.

1.3 The three domains

It helps to separate an 800 VDC AI hall into three domains, each with different electrical exposure:

2. Chillers on an 800 VDC campus

2.1 Why most chillers stay on AC

A chiller does not care how the IT equipment is powered. It sees a heat load arriving in the facility water and a condenser rejecting heat to ambient air or a cooling tower. For nearly every project, the right answer is to keep chillers, dry coolers and facility water pumps on 480Y/277 V AC (or 4.16 kV or 13.8 kV for large centrifugal machines), backed by standby generators. The reasons are practical:

• AC chillers are mature, listed, serviceable by local technicians and available with short lead times.
• Their compressors, fans and pumps already run on VFDs with well-understood AC input protection.
• Chillers are not on the uninterruptible bus in most designs; they ride through outages using thermal storage and restart on generator power.
• A DC-fed chiller adds a new failure mode — the DC distribution — to the cooling path without improving cooling.

2.2 What must still be validated

Keeping the chiller on AC does not remove the engineering. For AI loads, the following need explicit checks.

2.3 Worked example 1 — chiller duty for one 1 MW rack

The heat delivered to the facility water by one rack's CDU is the liquid-captured heat plus the CDU pump heat:

In refrigeration tons (1 ton = 3.517 kW):

At a full-load efficiency of 0.6 kW/ton for an air-cooled chiller at high ambient, the compressor and fan input is about 163 kW per rack. With warm water and partial free cooling, an annualized figure near 0.35 kW/ton (about 95 kW per rack) is achievable in many US climates. Across a 10 MW hall, the difference is roughly 0.7 MW of mechanical power at peak — which matters for generator sizing, utility demand and power usage effectiveness (PUE).

2.4 Thermal ride-through when a chiller trips

When chillers stop, the heat keeps arriving. The facility water temperature rises at a rate set by the stored thermal mass:

For a 10 MW hall with about 5,000 gallons (18,800 kg) of water in its piping, the rise rate with no heat rejection is:

The supply temperature climbs 5 K (9 °F) in about 40 seconds. If the chillers need seven minutes to restart and ramp, that is not enough. The water volume needed to absorb seven minutes at 10 MW within 5 K is:

A useful rule of thumb from the same equation: each 10,000 gallons of water buys about 13 minutes per megawatt of load for a 5 K (9 °F) allowable rise. Figure 1 (right) shows the effect: with only system water, the GPU power-cap threshold is reached in under a minute; with a 60,000-gallon buffer tank, the hall rides through the restart and ramp just at the threshold.

What this means for owners: the question is not "can chillers stay on AC?" — they can. The question is "what holds the water temperature while they restart?" The answer is some combination of buffer storage, rapid-restart chillers, chillers on the uninterruptible bus, and controlled GPU power reduction.

2.5 Direct-800 VDC chillers: when and what changes

A few owners are evaluating chillers and dry coolers fed directly from the DC bus, usually to eliminate a rectifier stage or to put cooling on the same battery-backed bus as the IT load. Such equipment is rare and emerging in the US market. Where it is considered, the changes are substantial:

• compressor, fan and pump drives designed for the DC supply voltage;
• DC-rated disconnects, overcurrent protection, wiring and insulation;
• precharge circuits for DC-link capacitors;
• controlled discharge of stored energy for maintenance safety;
• ground-fault detection or insulation monitoring compatible with the DC grounding scheme;
• DC/DC or DC/AC supplies for controls, heaters, crankcase heaters and auxiliary fans;
• listing to an appropriate product standard for DC input, which many current chillers do not have.

The next sections explain why each item matters.

3. Why an existing VFD is not an 800 VDC load

3.1 The DC link inside a 480 V drive

A standard six-pulse VFD rectifies its AC input onto a DC link. With a stiff supply and light load, the DC link sits near the peak of the line-to-line voltage:

A 480 V drive's capacitors, insulated-gate bipolar transistors (IGBTs) and overvoltage trip are designed around a DC link of roughly 650–750 V. Its overvoltage trip commonly sits around 800 V or a little higher, because regenerative braking can raise the bus. Feeding such a drive with 800 V nominal — which may itself run at 820–840 V during battery float or charge — would put it at or above its trip level continuously, with no margin for transients.



3.2 Options for DC-fed motors

The manufacturer must confirm, in writing, that the drive is rated for continuous DC input at the actual bus voltage range, including float and equalize voltages, and that its listing covers that use.

4. Precharge, stored energy and discharge

4.1 Why precharge matters

A DC-input drive or power shelf presents a large uncharged capacitor bank to the bus. Connecting it directly would draw an inrush limited only by cable and source impedance — potentially thousands of amperes — welding contactors, stressing connectors and possibly tripping upstream solid-state breakers. A precharge circuit charges the capacitors through a resistor (or an active current-limited circuit) before the main contactor closes.

4.2 Worked example 2 — precharging a 10 mF DC-link bank

Consider a pump-drive or power-shelf input bank of 10 mF on an 800 V bus. The stored energy at full voltage is:

To limit the precharge current to 20 A:

The charging time constant is:

After five time constants (2 s) the bank is above 99% of bus voltage and the main contactor can close. The resistor dissipates the same 3.2 kJ that ends up in the capacitors, with a peak power of 800²/40 = 16 kW for an instant — so it must be rated for pulse energy, not just continuous watts, and the precharge sequence must be supervised so that a failed contactor cannot leave the resistor in circuit.

4.3 Discharge and maintenance safety

Stored energy does not disappear when a disconnect opens. NFPA 70E requires workers to verify the absence of voltage and treat capacitors as hazards until they are discharged, and product safety standards for drives and power conversion equipment (such as UL 61800-5-1 for adjustable-speed drives) require either automatic discharge to a safe voltage within a defined time or a warning marking with the required waiting time. The time to discharge from V0 to a safe voltage through a resistor is:

4.4 Worked example 3 — the cost of a permanent bleeder

To bring the 10 mF bank from 800 V to 50 V in 5 seconds:

A permanent 180 Ω bleeder would dissipate 800²/180 ≈ 3.5 kW continuously — unacceptable in a cooling system designed to remove heat. Relaxing the time to 60 seconds needs about 2.2 kΩ and still wastes about 300 W per bank. This is why DC-input equipment usually uses switched active discharge: a resistor or converter connected only when the input disconnect opens, plus a slower passive bleeder as a backup. The owner's specification should state the required discharge time, the safe voltage level, and the indication that tells a technician the bank is discharged.

5. DC grounding, insulation monitoring and ground faults

5.1 Grounding schemes

How the 800 V bus is referenced to ground determines what a ground fault looks like:

Industry 800 V proposals generally favor a midpoint reference, so the maximum voltage to ground is 400 V, combined with continuous insulation monitoring. Whatever scheme is chosen, every piece of equipment on the bus — including any DC-fed CDU or chiller — must be compatible with it. A drive designed for a grounded-negative system can misbehave on a midpoint-grounded bus.



5.2 Insulation monitoring devices

An insulation monitoring device (IMD) injects a small measuring signal between the system and ground and reports the insulation resistance. Thresholds are commonly set as resistance per volt of system voltage. A widely used practice is to warn at a few hundred ohms per volt and alarm at around 100 Ω/V:

At that resistance, the leakage current at full pole-to-pole voltage would be only 10 mA — far below a short circuit, but a clear sign that something, very often water, is bridging insulation. In a liquid-cooled rack, a falling insulation reading is one of the earliest indicators of a coolant leak onto energized parts. The threshold actually used should follow the equipment manufacturer's guidance and the system's grounding design.

5.3 Code framework in the US

DC distribution at these voltages touches several parts of the National Electrical Code (NFPA 70). In the 2023 edition, Article 712 covers DC microgrids, Article 706 covers energy storage systems, Article 480 covers stationary batteries and Article 645 covers information technology equipment rooms. Disconnects, overcurrent devices, conductors and connectors must be listed and rated for DC at the system voltage; AC ratings do not apply. Later code editions are expanding requirements for DC systems above 1,000 V; owners should confirm with the authority having jurisdiction which edition is adopted and how it treats the specific architecture. Keentel's companion article on 800 VDC fault protection covers breakers, fuses and solid-state protection in detail.

6. Facility water and technology cooling loops



6.1 Two loops, two jobs

Liquid-cooled AI halls almost always use two hydraulically separate loops:

• Facility water system (FWS): the plant-side loop between chillers or dry coolers and the CDUs. It carries large flows, may contain glycol for freeze protection, and its water quality is managed at the plant.
• Technology cooling system (TCS): the clean, tightly filtered loop between the CDU and the rack cold plates. Its fluid, materials and filtration are controlled to protect micro-channel cold plates.

The CDU heat exchanger is the boundary between them. Keeping the loops separate protects the IT equipment from plant-side contamination and lets each loop run at the pressure and temperature it needs.

6.2 Temperature classes

ASHRAE TC 9.9's liquid-cooling guidance classifies facility water supply temperatures into classes such as W17, W27, W32, W40, W45 and W+ — the number being the maximum facility supply temperature in °C. Warmer classes allow more hours of compressor-free cooling with dry coolers. AI cold plates are increasingly designed for warm coolant, which is why many new halls target W32 to W45 facility water and use chillers only for trim and peak conditions. The TCS supply temperature is then the facility supply plus the CDU approach.

6.3 Worked example 4 — TCS flow for a 1 MW rack

The required mass flow follows from the heat balance:

For 950 kW of liquid heat, a 10 K (18 °F) rise and pure water (cp = 4.18 kJ/kg·K, density about 995 kg/m³ at 35–45 °C):

For PG25 (cp about 3.9 kJ/kg·K, density about 1,020 kg/m³ at operating temperature):

The often-quoted 1,360 L/min is therefore a pure-water figure with a density of 1,000 kg/m³. With glycol, the loop needs about 5% more flow — and glycol's higher viscosity also raises pressure drop, especially when cold.

A smaller temperature rise means more flow, larger pipes and more pump power; a larger rise means hotter return coolant and a hotter worst-case cold plate. The rise is set by the cold-plate manufacturer's inlet temperature and flow requirements, not chosen freely by the facility designer.

6.4 Worked example 5 — sensitivity to the liquid capture fraction

The share of rack heat that reaches the liquid depends on what is on cold plates. Processors and accelerators are always liquid-cooled; memory, voltage regulators, network interfaces, storage and power shelves may or may not be. For a 1 MW rack:

The airflow uses the standard sensible-heat relation in US units:

At 95% capture, 50 kW (about 170,600 BTU/h) needs about 7,900 CFM at a 20 °F air rise — roughly what a whole row of legacy enterprise racks needed. At 80% capture, the same rack needs about 31,600 CFM. The capture fraction must come from the rack manufacturer's data for the actual configuration, and the air-side system — rear-door heat exchangers, in-row coolers or room air handlers — must be sized for it.

7. Sizing the CDU

7.1 Approach temperature

The CDU approach is the difference between the TCS supply temperature leaving the CDU and the facility water supply entering it:

A smaller approach lets the facility water run warmer (more free cooling) for the same TCS supply, but requires a larger heat exchanger. Typical CDU approaches are about 2–5 K (4–9 °F). The approach must always be quoted together with the heat load, both flows and both fluids; an approach quoted for water on both sides does not hold for glycol.

7.2 Worked example 6 — heat-exchanger duty and LMTD

Take a CDU serving one 1 MW rack with 950 kW of liquid heat: TCS supply 35 °C and return 45 °C; facility water supply 32 °C (a 3 K approach). If the facility water flow is set for a 10 K rise, it returns at 42 °C. In a counterflow plate heat exchanger the two end temperature differences are both 3 K, so the log-mean temperature difference (LMTD) is 3 K:

If the facility side runs a smaller 8 K rise (returning at 40 °C), the end differences become 5 K and 3 K, the LMTD rises to about 3.9 K and the required UA falls to about 243 kW/K — a smaller heat exchanger, at the cost of more facility water flow (about 1,380 L/min or 364 gpm of water at a 10 K rise versus about 1,720 L/min at 8 K). This trade-off between heat-exchanger size and facility flow is one of the main CDU selection decisions.

7.3 Pumps, pressure and the system curve

The TCS pump must deliver the design flow against the total pressure drop of the loop: the CDU heat exchanger and filter, supply and return piping, row manifolds, hoses, quick disconnects, rack manifolds and the cold plates themselves. The operating point is where the pump curve crosses the system curve, and the system curve rises roughly with the square of flow:

7.4 Worked example 7 — pump power



For 1,433 L/min (0.0239 m³/s) of PG25 against 250 kPa (about 36 psi, or 83 ft of head):

With 70% pump efficiency and 94% motor efficiency, the electrical input is about 9.1 kW per megawatt rack — roughly 0.9% of the IT load. Pumps are a small share of power but the most critical share for continuity: without them, the cold plates have no flow.

7.5 Filtration, materials and coolant chemistry

Micro-channel cold plates can be blocked by particles and corrosion products. TCS loops typically use fine filtration (commonly in the tens of microns, as specified by the cold-plate manufacturer), wetted-material lists that exclude incompatible metals, inhibited glycol or treated water, and regular fluid sampling. These are not 800 VDC issues, but they become reliability issues for the electrical system: a loop that clogs or leaks puts water and debris next to high-current DC conductors.

8. Pump redundancy and the power behind it

8.1 Redundancy in the hydraulics

CDUs commonly use N+1 pumps, and halls commonly use N+1 or row-level redundancy across CDUs. Redundancy only works if the standby pump starts automatically and quickly, and if the surviving pumps can deliver enough flow to the worst-served rack. Keentel's digital-twin article on pump failure shows how flow falls to about 75% when one of two duty pumps trips and how the worst rack reaches throttling first.

8.2 Redundancy in the power

Two pumps fed from the same panel are not redundant. Recommended practice for megawatt-class racks:

• feed duty and standby pumps from different power sources (A and B), each on uninterruptible power;
• put CDU controls, valves, leak detection and the network link to the building management system on uninterruptible power;
• make sure pump VFDs ride through voltage sags and the generator transfer without tripping, or restart automatically within seconds;
• if pumps are fed from the DC bus, give them the same A/B diversity and the same DC protection selectivity as the IT load.

8.3 Worked example 8 — seconds without flow

If all TCS flow stops, the cold plates lose their heat path almost immediately. Even with flow continuing but no heat rejection (CDU heat exchanger lost its facility water), the rack loop's own fluid heats quickly. Take a TCS loop for one rack containing 400 L (about 106 gal) of PG25:

The loop gains 5 K in about 8 seconds. With no flow at all, GPU junction temperatures reach their throttling limits even faster. Cooling continuity for megawatt racks is measured in seconds, which is why pump power must be uninterruptible and why cooling-loss detection must be wired to automatic GPU power reduction.

9. Keeping water away from 800 V

9.1 Physical separation

The diagram many teams draw — DC protection and DC/DC conversion at the top of the rack, liquid manifolds down the back — reflects a sound principle: put energized parts above or beside, never below, liquid connections, and give any leak a defined drip path away from conductors. Specific practices include:

• placing quick disconnects and manifold joints below and away from busbars and DC connectors;
• drip trays and drip shields with drains or leak sensors;
• enclosures and connectors with ingress protection appropriate to the location;
• non-conductive, compatible coolants where the design allows, while recognizing that inhibited glycol mixtures are still conductive;
• clear segregation in cable trays and overhead busway from coolant piping;
• insulation monitoring that trends, not just alarms.

9.2 Leak detection that acts

Leak detection is only useful if it triggers a defined action. A practical response ladder:

9.3 Rack requirements summary

10. Linking cooling loss to workload reduction

10.1 Why the link is essential

A megawatt rack cannot wait for an operator. When flow drops, coolant temperature rises or a leak is detected, the rack must reduce power automatically and predictably. Uncontrolled GPU thermal throttling degrades performance unevenly across a training job and gives operators no control; a planned power cap holds every GPU to a known level and keeps the job running.

10.2 A typical response timeline

The thresholds and times must come from the cold-plate and GPU thermal data and from a model of the actual loop — ideally a validated digital twin — and they must be tested during commissioning.

11. The power chain for cooling and its effect on PUE



The electrical load of the cooling system is part of the facility's total and therefore part of PUE:

For the 10 MW example hall:

These are illustrative. The point is that 800 VDC improves the electrical-loss term, while warm-water design, free cooling and a high liquid capture fraction improve the cooling term. Both matter.

12. Commissioning the interface

Commissioning must test the electrical and mechanical systems together, not separately:

• Pump failover under load: trip a duty pump at full rack power and record flow, temperatures and GPU power.
• Utility outage and generator transfer: confirm that pumps, CDU controls and leak detection stay up, and that chillers restart within the ride-through time.
• Chiller trip: record facility water temperature against the buffer-tank prediction.
• Leak simulation: confirm that a leak signal and a simulated insulation fault produce the specified power reduction and isolation.
• DC precharge and discharge: verify inrush, contactor sequencing and discharge time on DC-fed equipment.

13. A practical specification checklist

• State the rack power, DC voltage range and grounding scheme in every mechanical specification that touches the rack.
• Specify CDU capacity with heat load, both fluids, both flows, approach and maximum pressure drop.
• Use the manufacturer's liquid capture fraction, and size the air-side system for the remainder.
• Put CDU pumps, controls and leak detection on A/B uninterruptible power.
• Define the chiller restart time and size thermal storage or GPU capping to cover it.
• Keep chillers on AC unless a DC-input product is listed for the actual DC system; for any DC-fed equipment, specify DC ratings, precharge, discharge time, insulation compatibility and listing.
• Define the cooling-loss response ladder and test it at commissioning.
  

Part C — Three Anonymous Case Studies

These composite case studies are based on the kinds of issues Keentel Engineering sees on US projects. Names, locations and details have been changed or combined.

Case study 1 — A "compatible" pump drive on the DC bus

Situation. A developer in the Southwest designed a 20 MW AI hall with 800 VDC rack distribution and wanted CDU pumps on the same battery-backed DC bus as the racks, to avoid a separate AC uninterruptible system. The mechanical contractor submitted the CDU's standard 480 V pump drives, noting that they "already have a DC link."

Analysis. Keentel reviewed the drive data. The drives' DC links were designed for about 650 V nominal, with an overvoltage trip at about 810 V. The DC bus would run at 800 V nominal and up to about 830 V during battery charging. The drives would have tripped as soon as the batteries went to float — and the precharge circuit, designed for an AC-fed rectifier, was not active on a DC input. There was also no specified discharge time for the drive capacitors once DC-fed, and the drive listing did not cover DC input.

Engineered solution. Keentel compared three options: 690 V-class drives listed for DC input, DC/DC converters feeding standard drives, and keeping pumps on AC with a small dedicated AC uninterruptible supply. The owner chose the AC option for the first phase, because listed DC-input pump drives in the needed size had long lead times, and kept a DC/DC option for later phases. Pump feeders were split A/B, and the pump VFDs were set to ride through generator transfer.

Outcome. Commissioning tests confirmed that pumps stayed running through a full utility outage and transfer, with no change in TCS flow. The owner avoided a design that would have tripped every pump the first time the batteries charged.

Case study 2 — The chillers were fine; the minutes were not

Situation. A Mid-Atlantic AI campus converted one hall to liquid-cooled racks at roughly 130 kW each, with a roadmap toward much higher densities on an 800 VDC architecture. The air-cooled chillers had ample capacity on paper. During a utility outage, the chillers took about eight minutes to return to full capacity on generator power, and the GPUs throttled heavily across the hall.

Analysis. Keentel built a thermal model of the facility water system. The hall carried about 6 MW of liquid heat, but the system water volume was only about 4,000 gallons. With no heat rejection, the facility supply temperature rose about 5 K in roughly a minute, far shorter than the chiller restart time. The CDU pumps were on uninterruptible power and kept circulating — which delivered the heat efficiently to a loop that could not reject it.

Engineered solution. The solution combined three measures: a 40,000-gallon chilled-water buffer tank sized with the equation in Section 2.4, rapid-restart settings on the chillers, and an automatic hall-level GPU power cap triggered by facility water supply temperature, stepping to 70% before the throttling limit. The controls were tied to the generator transfer signal so the cap could anticipate the event instead of waiting for temperature.

Outcome. A repeat outage test showed the facility water rising less than 4 K, the power cap engaging briefly and smoothly, and no uncontrolled GPU throttling. The model was retained for planning the 800 VDC phase, where higher rack densities will need a larger tank or chillers on the uninterruptible bus.

Case study 3 — Water, glycol and 800 V in the same rack

Situation. A Texas operator was qualifying a rack design for a future 800 VDC deployment. The prototype rack placed the DC input connectors and the rack's DC/DC shelf at the bottom, with liquid manifolds running vertically and quick disconnects above the power shelf.

Analysis. Keentel's review identified three risks: any quick-disconnect drip would fall onto the DC shelf; the inhibited PG25 coolant was conductive enough to create a leakage path; and the rack had no insulation monitoring input or leak sensor tied to its power controls. The flow sizing was also based on water, not PG25, and on 95% liquid capture, whereas the vendor's data showed about 85% for the planned configuration — leaving an extra 100 kW per rack for the air-side system.

Engineered solution. The rack layout was revised to put the DC input and power shelf at the top with drip shields, move quick disconnects below the compute trays with a drain tray and leak sensor, add an insulation monitoring channel per rack, and link leak and insulation alarms to a rack power cap and DC isolation sequence. Flow and CDU sizing were recalculated for PG25 at 85% capture, and the row's rear-door heat exchangers were upsized for the air-side load.

Outcome. The revised rack passed the operator's leak and fault tests, and the updated CDU and air-side design avoided a hidden 10% cooling shortfall per rack.

Part D — 20-Question Technical FAQ

Keentel Services, References and Contact

How Keentel Engineering Helps

Keentel Engineering supports AI data center owners, developers, design-build teams and operators at the point where power and cooling meet:

• 800 VDC and AC power architecture: single-line design from 138 kV service through 34.5 kV and 13.8 kV distribution to 480Y/277 V mechanical power and 800 VDC rack distribution.
• Mechanical load power design: power for chillers, dry coolers, CDUs and pumps, including uninterruptible power for critical cooling, A/B diversity and ride-through settings.
• DC equipment interface reviews: drive DC-link compatibility, precharge, discharge, insulation monitoring and protection coordination for any DC-fed equipment.
• Thermal–electrical outage studies: chiller restart, buffer-tank sizing and GPU power-cap timelines, including digital-twin models of the cooling loops.
• Generator, utility and large-load interconnection studies: load ramps, transfer behavior and utility requirements for large data center loads.
• Commissioning support: integrated tests of pump failover, outage transfer, leak response and DC precharge.

Why Clients Choose Keentel

• Power and cooling in one view. We design the electrical systems that keep cooling running, so the interface is engineered, not assumed.
• From the substation to the rack. Power system studies from 4 kV to 765 kV, plus DC distribution and EMT modeling.
• Numbers, not slogans. Every recommendation is backed by worked calculations the owner's team can check.
• US codes and standards. NEC, NFPA 70E, IEEE and utility requirements applied correctly.
• Responsive senior engineers who stay with the project from concept through commissioning.

Talk to Keentel Engineering

Planning an 800 VDC AI hall, or upgrading a liquid-cooled facility for higher rack densities? Keentel Engineering can help you design the power behind the cooling — and the cooling behind the power.

• Phone: (813) 389-7871
• Email: contact@keentelengineering.com
• Book a 15-minute call: calendly.com/keentel-engineering/15min
• Web: keentelengineering.com

References

• NFPA 70, National Electrical Code, 2023 edition (Articles 480, 645, 706 and 712, among others).
• NFPA 70E, Standard for Electrical Safety in the Workplace, 2024 edition.
• NFPA 75, Standard for the Fire Protection of Information Technology Equipment.
• UL 61800-5-1, Adjustable Speed Electrical Power Drive Systems — Safety Requirements.
• UL 1778, Uninterruptible Power Systems.
• UL 9540 and UL 9540A, Energy Storage Systems and Thermal Runaway Fire Propagation Test Method.
• ASHRAE TC 9.9, Liquid Cooling Guidelines for Datacom Equipment Centers, and Thermal Guidelines for Data Processing Environments.
• ASHRAE Handbook — HVAC Systems and Equipment, chapters on chillers, pumps and heat exchangers.
• IEC 61557-8, Insulation monitoring devices for IT systems (as applied by equipment manufacturers).
• Open Compute Project, published guidance on liquid cooling and rack power distribution.
• Incropera, F. P., et al., Fundamentals of Heat and Mass Transfer, Wiley.

Disclaimer

This article is provided by Keentel Engineering for general technical education. It is not engineering advice for any specific project. The installation, densities, fluid properties, efficiencies, volumes, thresholds and case studies are illustrative or composite. Actual designs depend on the equipment manufacturers' data, the adopted codes, the authority having jurisdiction and site conditions. Standards and products change; always consult the current editions and the equipment manufacturers. Keentel Engineering accepts no liability for decisions made on the basis of this article without a project-specific engineering review.

© 2026 Keentel Engineering. All rights reserved. Written by Keentel Engineering.