Inside a Data Center Rack: Components, Power, Cooling & AI Infrastructure

Part A — The Data Center Rack at a Glance

The cloud is not a cloud. Every AI prompt, API call, card transaction and streamed video ends up as electrical current flowing through a power supply in a server, sitting in a rack, on a floor, inside a building that someone engineered. Kubernetes clusters, serverless functions and AI agents hide that hardware from application teams, but they do not remove it. When a production system becomes slow or unreliable, the investigation eventually reaches the physical layer: network paths, compute capacity, storage latency, power and cooling.

The standard equipment rack — a 19-inch cabinet, typically 42 rack units (42U) tall — is where all of these systems meet. It is where network cabling terminates, where servers and storage turn electrical energy into computation, and where every watt of that energy becomes heat that must be removed. For an electrical engineer, the rack is the last link in a power chain that may begin at a 345 kV transmission substation.

This guide from Keentel Engineering explains what is inside the rack and how each component works, then goes deeper into the engineering that keeps it running: the power chain from utility service to rack PDU, PDU and branch-circuit sizing, A/B redundancy, UPS and battery selection, short-circuit and arc-flash safety, grounding and bonding, heat removal, network and storage performance, and the codes and standards that govern each of them in the United States.

What lives in a typical rack

What Keentel's analysis shows

Keentel Engineering compared four generations of rack power density against the capacity of the rack PDUs and branch circuits that commonly feed them, and modeled A/B feed loading in normal and failover conditions.

Figure 1. Keentel analysis. Left: typical rack power by generation compared with the usable capacity of common rack feeds at 208 V and 415 V (three-phase, 80% continuous loading). Right: A/B feed loading for a 15 kW rack on 415 V 30 A feeds — normal operation (each side carries half the load) and failover (one side carries all of it), with the 80% continuous-load limit shown.

Five lessons

• The rack is an electrical system, not just furniture. It sits at the end of a power chain that may start at 138, 230 or 345 kV. Every decision upstream — voltage, redundancy, protection, grounding — shows up at the rack PDU.

• A/B redundancy only works if each side can carry the whole rack. With two feeds sharing the load, each must be loaded to no more than about 40% of its breaker rating in normal operation, so that it stays at or below 80% when the other side fails.

• Voltage matters. At the same power, a 415Y/240 V rack feed carries half the current of a 208Y/120 V feed and has one-quarter of the I²R loss in the same conductor. That is why high-density and AI racks have moved to 415 V.

• Every watt becomes heat. A 15 kW rack needs about 2,400 CFM of air at a 20 °F temperature rise. A 130 kW AI rack cannot be air-cooled practically and needs liquid cooling.

• Codes and standards are part of the design. NEC Articles 645, 250 and 240, NFPA 70E, NFPA 75 and 855, and UL 62368-1, 1778, 9540 and 9540A govern what can be installed and how people can safely work on it.

Engineering review of the common rack description

Part B — Inside the Data Center Rack: Power, Protection, Cooling and Connectivity Engineering

1. Why the physical layer still decides performance

Software abstractions have made computing look weightless. A developer can deploy a container to a cluster in another state without knowing which building, row or rack it lands in. That is the purpose of abstraction, and it works well until something goes wrong. Then the questions become physical:

• Is a link congested, or is a transceiver failing and retransmitting?

• Is the application slow because the storage array is saturated, or because the network path to it is oversubscribed?

• Did the servers throttle because the rack exceeded its power cap, or because inlet air was too hot?

• Did a whole rack go dark because one PDU tripped, and why did its partner not carry the load?

Each of these questions is answered at the rack. For an owner, developer or operator, understanding the rack is the difference between treating an outage as a mystery and treating it as an engineering problem with a measurable cause.

Keentel Engineering approaches the rack as the last node of an electrical system and the first node of a thermal system. The same discipline that sizes a 345 kV interconnection — load flow, short-circuit analysis, protection coordination, grounding and code compliance — applies to the 415 V branch circuit that feeds a GPU server.

2. Rack anatomy and dimensions

2.1 The 19-inch standard

Almost all data center equipment fits the EIA-310 standard rack: mounting rails spaced for 19-inch-wide equipment panels, with vertical height measured in rack units (U).

A 42U rack therefore has 73.5 inches of usable mounting height. Racks are commonly 24 inches (600 mm) wide, or 30 inches (750–800 mm) wide when extra space is needed for cable management and vertical PDUs. Depth ranges from about 42 to 48 inches; GPU servers and deep storage chassis increasingly need 48-inch racks.

2.2 Open Rack

The Open Compute Project's Open Rack specification takes a different approach. Equipment is 21 inches wide, the height unit (OU) is 48 mm, and power is distributed inside the rack on a DC busbar (about 48–54 V in Open Rack V3) fed by rack-mounted power shelves. Battery backup units (BBUs) can sit in the same rack. Many AI systems use Open Rack–style designs with liquid-cooling manifolds integrated into the rack.

2.3 Weight and placement

A fully loaded 42U enterprise rack typically weighs 1,500–2,500 lb. AI racks with integrated liquid cooling can exceed 3,000 lb. Placement rules that follow from this:

• Put the heaviest equipment — UPS, battery modules, storage enclosures — at the bottom.

• Load the rack from the bottom up, and use anti-tip stabilizers or bolt racks to the floor or to each other.

• Check rack static and dynamic load ratings, and check floor capacity (Section 14).

3. The network layer

3.1 Patch panels and structured cabling

A patch panel is a passive termination point. Permanent cabling — copper (Category 6A) or fiber (OM4/OM5 multimode or OS2 single-mode) — is terminated on the back of the panel, and short patch cords connect the front ports to switches or servers. The panel makes moves, adds and changes possible without disturbing the permanent cabling, and gives each link a documented, labeled endpoint.

The US structured cabling standards are the ANSI/TIA-568 series for cabling and components and ANSI/TIA-942 for data center infrastructure. TIA-942 defines the spaces a cable passes through:

Key channel limits engineers should know: a Category 6A copper channel supports 10GBASE-T to 100 m; passive direct-attach copper (DAC) cables for 100–400 Gb/s are limited to a few meters, which is why they are used only within a rack; multimode fiber supports 100–400 Gb/s over roughly 70–150 m depending on fiber grade and optic; single-mode fiber supports 500 m to 10 km and beyond.

3.2 Top-of-rack switches and the leaf–spine fabric

The switch in the rack is usually a top-of-rack (ToR) or "leaf" switch. Every server connects to it, often to two of them for redundancy. Each leaf connects to every spine switch in a leaf–spine (Clos) fabric. Any server can reach any other server in the fabric in at most three switch hops (leaf–spine–leaf), and adding spines adds bandwidth.

The key design parameter is oversubscription — the ratio of server-facing bandwidth to uplink bandwidth on the leaf.

Worked example 1 — leaf oversubscription

A leaf switch has 48 server ports at 25 Gb/s and 8 uplinks at 100 Gb/s.

An oversubscription of 1.5:1 is acceptable for general-purpose cloud workloads, where servers rarely transmit at full rate at the same time. AI training clusters are different: GPUs exchange gradients in synchronized bursts, so AI back-end networks are designed at 1:1 (non-blocking) with 400–800 Gb/s per GPU.

3.3 Latency — what the physics allows

Network latency inside a data center has four parts: propagation, serialization, switching and queuing.

Propagation in fiber is set by the refractive index of the glass (about 1.47):

Serialization is the time to clock a packet onto the wire:

Worked example 2 — latency budget across a fabric

A 1,500-byte packet travels from a server to another server in a different row: 3 switch hops, 300 m of total fiber, 100 Gb/s links.

The lesson is that in an uncongested fabric, physical distance and switch hops dominate and the total is a few microseconds. Under congestion, queuing can add hundreds of microseconds. When an application is slow, queuing — caused by oversubscription or a failing link — is a far more likely culprit than distance.

3.4 Where firewalls and load balancers really sit

A small server room may have a firewall and a load balancer in the same rack as the servers. Production data centers separate functions:

• Edge routers and perimeter firewalls sit in network racks near the meet-me room or MDA, where carrier circuits enter.

• Load balancers sit in network/services racks, or are deployed as software (virtual appliances, Kubernetes ingress controllers, service meshes).

• East–west security between workloads is enforced by distributed firewalls in the hypervisor, on smart NICs/DPUs, or by network policy in the container platform.

The inbound path for a typical web request is therefore: carrier circuit → edge router → perimeter firewall → load balancer → spine → leaf (ToR) → server. The patch panels and cross-connects are in the physical path at several points but play no role in the logical flow.

4. Compute and storage

4.1 Servers and power supplies

Server power has grown with every processor generation. A general-purpose two-socket 1U server draws about 300–800 W; a dense 2U server about 1–1.5 kW; an 8-GPU AI server 6–12 kW; and an integrated AI rack system 100–140 kW.

Server power supplies (PSUs) accept 100–240 V AC (some accept 277 V AC or DC), and are typically redundant: two PSUs per server, each connected to a different PDU (A and B). Modern PSUs have active power-factor correction, so the input power factor is above about 0.95 at moderate load and current harmonic distortion is low. Efficiency is highest on 200–240 V input; high-efficiency ("Titanium" class) units reach about 96% at half load.

Two practical consequences:

• Power factor is close to 1, so kW and kVA are similar at the rack — but not equal. Use kVA for circuit sizing.

• Load sharing between redundant PSUs means each PDU normally carries about half of each server's load. If one side fails, the surviving PSU carries all of it (Section 7).

4.2 Storage: NAS, SAN and object

Storage performance has three measures — IOPS (operations per second), throughput (MB/s) and latency — linked by block size and queue depth:

The second relationship is Little's law.

Worked example 3 — storage sizing

A database needs 200,000 random 4 KiB reads per second at no more than 0.5 ms latency.

A single NVMe SSD can sustain several hundred thousand random reads at sub-millisecond latency, so a small NVMe array meets the requirement with margin. The same workload on 7,200 rpm hard disks (about 100–200 IOPS each) would need more than 1,000 drives. This is why storage latency, not capacity, usually drives the storage design — and why storage racks have very different power and weight profiles from compute racks.

5. The power chain — from transmission substation to rack PDU

5.1 The voltage ladder

A large data center campus is a utility-scale load. A 100 MW campus is typically served at transmission voltage, steps down to medium voltage for distribution across the site, and then to low voltage at each data hall.

Two US low-voltage architectures dominate:

• 480 V → 208Y/120 V. The traditional design. PDUs contain 480 V delta to 208Y/120 V wye transformers. Rack outlets deliver 208 V phase-to-phase or 120 V phase-to-neutral.

• 480 V → 415Y/240 V. Increasingly common. Rack outlets deliver 240 V phase-to-neutral, which suits server power supplies rated 200–240 V, while the three-phase 415 V feed doubles the power per conductor compared with 208 V. Some facilities use 415 V UPS outputs and eliminate a transformer stage.

Keentel's companion articles on data center load calculations and tier classifications cover how the utility service, generation and redundancy levels are sized for a whole campus. This guide focuses on the last few hundred feet: from the UPS output to the server power supply.

5.2 Three-phase power at the rack

For a balanced three-phase circuit:

where S is apparent power (VA), P is real power (W), VLL is line-to-line voltage, I is line current and PF is power factor.

For single-phase circuits:

5.3 The 80% continuous-load rule

IT loads are continuous loads — they run for three hours or more at a time. The NEC requires branch-circuit and feeder overcurrent devices to be rated at least 125% of the continuous load unless the device and its assembly are listed for 100% continuous operation (NEC 210.20(A) and 215.3). NEC 645.5(A) applies the same 125% factor to branch-circuit conductors supplying IT equipment. Stated the other way round, a standard breaker may carry continuously no more than 80% of its rating:

Worked example 4 — usable capacity of common rack feeds

At the same breaker size, a 415 V circuit delivers twice the power of a 208 V circuit. At the same power, it carries half the current — and since conductor losses scale with current squared, one-quarter of the I²R loss in the same conductor:

5.4 Phase balance

A three-phase rack PDU supplies single-phase outlets: line-to-line at 208 V, or line-to-neutral at 240 V on a 415Y/240 V PDU. Outlets are grouped by phase or phase pair. If servers are plugged in without regard to phase, one phase can reach its limit while others are lightly loaded. The usable capacity of the PDU is then set by the most heavily loaded phase.

Worked example 5 — phase imbalance on a 415Y/240 V PDU

A 415 V 30 A PDU feeds 240 V outlets on three phases. Loads: L1 = 22 A, L2 = 14 A, L3 = 12 A.

The PDU is limited to 24 A (80% of 30 A) per phase, so L1 is at 92% of its continuous limit while total load is only 48 A of a possible 72 A — 67% of capacity. Rebalancing to 16 A per phase leaves headroom of 8 A per phase. Metered PDUs report current per phase; commissioning should include a phase-balance check, and change procedures should specify which outlet group each new device uses.

Note that on a 415Y/240 V or 208Y/120 V PDU with line-to-neutral loads, imbalance also produces neutral current. Modern power supplies with power-factor correction produce little third-harmonic current, but neutral conductors in legacy or mixed installations should be checked for heating.

6. Sizing a rack PDU and its branch circuit

6.1 The procedure

• Establish the design load per rack — not nameplate. Server nameplates often state the PSU rating, which may be two to three times the actual maximum draw. Use measured data, vendor power calculators or the configured maximum.

• Convert to kVA using the power factor (about 0.95–0.99 for modern IT).

• Apply redundancy: with A/B feeds, each side must carry the full rack load on its own.

• Select a PDU and branch circuit whose 80% continuous capacity is at least the full rack load.

• Check outlet types and counts (C13/C19 at 208–240 V; C13 rated about 10–15 A, C19 about 16–20 A depending on listing), per-outlet-group breaker ratings and phase balance.

• Check voltage drop and short-circuit current rating (SCCR) at the PDU location.

Worked example 6 — sizing feeds for a 15 kW rack

A rack of 1U and 2U servers draws a maximum of 15 kW at PF 0.98, with A/B redundancy.

Each feed must carry 15.3 kVA alone.

• 208 V 3-phase 60 A: usable 17.3 kVA — adequate (88% of usable on failover).

• 415 V 3-phase 30 A: usable 17.3 kVA — adequate, with half the conductor size and current of the 208 V option.

• 208 V 3-phase 30 A: usable 8.6 kVA — inadequate; would need two circuits per side.

Line current on the 415 V feed in failover:

That is 71% of the 30 A breaker — within the 80% limit (24 A).

6.2 Metering and switching

Branch circuit monitoring at the RPP and per-phase metering at the rack PDU should feed a monitoring system (DCIM or BMS) with alarms set below the 80% continuous limit — typically a warning at about 40% per feed in normal A/B operation and a critical alarm at about 45–50%.

7. A/B redundancy and failover loading

7.1 The 40% rule

In a 2N or dual-corded design, each server has two power supplies, one plugged into PDU A and one into PDU B, fed from independent upstream paths. In normal operation, the two supplies share the load roughly equally. If feed A fails, feed B instantly carries the entire load.

For feed B to stay within the 80% continuous limit after failover, the combined load must be no more than 80% of one feed's rating, so each feed must be loaded to no more than 40% in normal operation:

This is one of the most frequently violated rules in data center operations. Facilities add servers one at a time, each PDU reading looks comfortable at 50–60%, and the problem appears only when a maintenance outage or a PDU failure puts the whole load on one side — which then trips.

Worked example 7 — the failover trip

A rack is fed by two 208 V 3-phase 30 A circuits. Each PDU reads 14 A per phase in normal operation — 47% of rating, which "looks fine."

On failover, the surviving PDU carries 28 A per phase — 93% of rating and above the 24 A continuous limit. A standard thermal-magnetic breaker carrying 93% of rating will not necessarily trip immediately, but it will run hot, and any additional load (a server boot surge, a fan ramp on a hot day) can trip it. With the breaker open, the rack is dark — the redundancy design produced a total outage.

The fix is to cap normal per-feed loading at 40% (12 A on a 30 A circuit) and move load to additional circuits or a higher-capacity feed.

7.2 Single-corded equipment

Some devices have only one power supply. They should be connected through a rack-mounted automatic transfer switch (ATS), which transfers between feeds A and B in a few milliseconds — fast enough for most power supplies to ride through. The ATS becomes a single point of failure for those devices, so critical single-corded equipment should be minimized.

7.3 Unequal sharing

Some power supply pairs do not share equally; one may carry 60% or more. Others operate in active–standby mode, in which one PSU carries all the load. Measure per-PDU current, not just the total, and design for the worst case: the full rack load on either side.

8. Voltage drop and conductor sizing

The NEC informational notes recommend limiting voltage drop to 3% on a branch circuit and 5% total across feeder and branch circuit. Data center designers often hold tighter limits because server power supplies regulate their output but draw more current at lower voltage.

where L is the one-way length and R and X are conductor resistance and reactance per unit length.

Worked example 8 — voltage drop on a 208 V 60 A rack feed

A 208 V 3-phase feed carries 48 A (80% of 60 A) over 100 ft of 4 AWG copper in PVC conduit. From NEC Chapter 9, Table 9: R = 0.31 Ω/1,000 ft, X = 0.048 Ω/1,000 ft. With PF = 0.95 (sin θ = 0.31):

The same rack on a 415 V 30 A feed carries half the current, and the percentage drop on the same conductor is one-quarter as large — another reason higher-density racks favor 415 V.

9. UPS systems and batteries

9.1 What the UPS really does

A UPS performs three functions: it conditions power (filtering sags, swells, transients and frequency variation); it carries the load on battery for a defined time when utility power fails; and it bridges the load until standby generators start and take over. NFPA 110 classifies a generator system that restores power within 10 seconds as Type 10, so UPS runtime is designed to cover generator starting, transfer and several retry attempts — typically 5 to 15 minutes for centralized UPS, and 1.5 to 5 minutes for rack-level battery backup units in some hyperscale designs.

In the United States, UPS equipment is listed to UL 1778. The UPS shown in many rack diagrams — a 2U or 3U unit at the bottom of the rack — is a line-interactive or small double-conversion unit, typically 1–10 kVA. It is the right solution for an edge site, a branch office or a small server room. It is not how large facilities protect their racks: they use centralized UPS systems in dedicated electrical rooms, or rack-level BBUs integrated into the rack power shelf.

9.2 Battery runtime and sizing

The energy the battery must deliver is:

But battery capacity depends strongly on discharge rate. A lead-acid battery discharged in 5–10 minutes delivers far less than its nominal (20-hour) capacity. Size batteries from the manufacturer's constant-power discharge tables at the required runtime and end voltage, and apply an aging factor (typically 1.25, because batteries are replaced at 80% of rated capacity) and a temperature correction, following the methods in IEEE 485 (lead-acid sizing) and IEEE 1184 (batteries for UPS).

Worked example 9 — rack UPS runtime

A 6 kVA rack UPS supports a 4.5 kW load for 8 minutes. Inverter efficiency is 94%.

With an aging factor of 1.25, the battery must deliver 0.80 kWh at the 8-minute rate at end of life. For VRLA batteries, which deliver roughly half their nominal energy at an 8-minute rate, that means about 1.6 kWh of nominal capacity — typically the internal battery plus one external battery pack. Lithium-ion delivers about 90% of its nominal energy at this rate, so about 0.9 kWh nominal suffices, at much lower weight.

Worked example 10 — centralized UPS battery for a 1 MW block

A 1 MW UPS module (inverter efficiency 96%) needs 5 minutes of runtime at full load.

With a 1.25 aging factor and a 10% design margin, the battery must deliver about 119 kWh at a 5-minute rate. At this discharge rate, the battery string is sized by power, not energy: about 1,040 kW of DC power for five minutes.

9.3 VRLA versus lithium-ion

Lithium-ion UPS batteries should be listed to UL 1973 (stationary batteries), with the system listed to UL 1778, and supported by UL 9540A thermal-runaway test data when the installation is large enough to fall under NFPA 855 or the fire code's energy storage provisions. NFPA 855 sets aggregate thresholds by technology — for lithium-ion, 20 kWh in a room or fire area — above which energy storage requirements apply. How those provisions apply to UPS batteries depends on the edition adopted and the authority having jurisdiction (AHJ), so confirm early in design.

9.4 Rack-level battery backup

Open Rack–style AI and hyperscale racks often use BBU shelves on the rack's 48–54 V DC bus. They provide 90 seconds to a few minutes of ride-through — enough to bridge generator start — without a centralized UPS. Advantages include higher efficiency (no double conversion) and fault isolation per rack. Engineering issues include fire separation and detection in the white space, UL 9540A data at the rack level, recharge power demand after an outage, and coordination of BBU discharge with generator transfer.

10. Short-circuit current, SCCR and selective coordination

10.1 How much fault current reaches the rack?

Every piece of equipment in the power chain must withstand or interrupt the fault current available at its terminals. NEC 110.9 requires interrupting ratings at least equal to the available fault current, and NEC 110.10 requires that circuit impedance, short-circuit current ratings and overcurrent device characteristics be coordinated so that a fault is cleared without extensive damage. Panelboards, RPPs and many rack PDUs carry a short-circuit current rating (SCCR) that must be checked against the calculated available current.

For a transformer fed from a strong (infinite) source, the maximum three-phase fault current at its secondary is approximately:

Worked example 11 — available fault current through the chain

These are upper-bound estimates. A full short-circuit study (to IEEE 551 methods, in commercial software) includes source impedance, cable impedance, motor contribution and the current-limiting behavior of inverters and UPS. But the screening calculation already shows something important: rack PDUs installed close to large PDU transformers can see fault currents above their SCCR (commonly 5 kA or 10 kA). Where that happens, use upstream current-limiting fuses or breakers with tested series ratings, or select PDUs with higher ratings.

10.2 Selective coordination

Selective coordination means that for any fault, only the overcurrent device nearest the fault opens. The NEC requires it for emergency systems (700.32), legally required standby systems (701.32) and critical operations power systems (708.54). It is not generally mandated for data center IT distribution, but it is one of the most valuable design choices an owner can make. A fault in one server's power cord should trip a 20 A outlet breaker in the rack PDU, not the 225 A RPP main or the 800 A UPS output breaker serving 40 racks.

Selectivity at high fault currents cannot be judged from time–current curves alone, because molded-case breakers in the instantaneous region can open together. Designers must use manufacturer-tested selectivity tables for breaker pairs, or fuse ratio tables (for many current-limiting fuse classes, a 2:1 ampere ratio between upstream and downstream fuses is selective). Keentel's protection studies combine time–current coordination with these tables to show selectivity down to the rack PDU.

11. Arc-flash safety

Data center electrical rooms contain some of the highest-energy low-voltage equipment in any building. Work on energized equipment must follow NFPA 70E, which requires an arc-flash risk assessment, equipment labeling with incident energy or PPE category, and an electrically safe work condition wherever possible. Incident energy is calculated with IEEE 1584-2018, which models electrode configuration, enclosure size, gap and clearing time.

Three points are particular to data centers:

• Clearing time dominates. Incident energy is roughly proportional to arc duration. A 480 V switchboard breaker set with a 0.3 s short-time delay (for coordination) might produce about 30 cal/cm² at the working distance; the same fault cleared in 0.05 s produces about one-sixth of that. NEC 240.87 requires an arc energy reduction method — such as an energy-reducing maintenance switch, zone-selective interlocking or differential relaying — for breakers rated or adjustable to 1,200 A or more.

• UPS and inverter sources limit fault current when operating on battery, which can lengthen clearing time and increase incident energy at downstream devices. The arc-flash study must consider both utility and battery operation.

• 208 V and 240 V equipment is not automatically low-risk. IEEE 1584-2018 removed the earlier 125 kVA transformer exemption. It notes that sustained arcs are possible but less likely on three-phase systems of 240 V or less with available fault current below 2,000 A. Equipment above that threshold must be evaluated.

12. Grounding and bonding

Grounding and bonding do three jobs in a data center: they provide a low-impedance fault-return path so overcurrent devices operate (NEC Article 250); they limit touch voltages; and they create a common reference that minimizes noise and potential differences between interconnected equipment.

Common faults that Keentel finds in field investigations include missing system bonding jumpers on PDU transformers, neutral-to-ground bonds in more than one place (creating objectionable current on grounding conductors), racks bonded only through painted surfaces, and isolated-ground receptacles misused as a substitute for proper bonding. These problems can cause both shock hazards and network errors — elevated ground noise shows up as CRC errors and link flaps on copper interfaces.

13. Heat removal — every watt becomes heat

13.1 The heat balance

Essentially all the electrical power delivered to IT equipment becomes heat that the cooling system must remove:

For air in US units, at standard conditions:

For water in US units:

Worked example 12 — airflow for a 15 kW rack

With a 20 °F (11 °C) air temperature rise across the servers:

That is the airflow of several large floor tiles or an in-row cooler. It is achievable with hot- or cold-aisle containment, but at 30–50 kW per rack the airflow becomes impractical and rear-door heat exchangers or liquid cooling take over.

Worked example 13 — liquid cooling for a 130 kW AI rack

Assume 85% of the heat goes to direct-to-chip liquid cooling, with an 18 °F (10 °C) coolant temperature rise.

The remaining 19.5 kW still has to be removed by air (about 3,100 CFM at 20 °F rise). Keentel's article on coolant distribution units (CDUs) covers the liquid loop in depth.

13.2 Thermal envelope

ASHRAE Technical Committee 9.9 publishes the thermal guidelines that most US facilities use. The recommended server inlet range is 18–27 °C (64.4–80.6 °F); allowable ranges for equipment classes A1–A4 extend wider for limited periods. Liquid-cooled equipment is classified by facility water supply temperature (W-classes). Key practices:

• Measure inlet temperature at the front of the rack, at the top, middle and bottom.

• Fill unused rack spaces with blanking panels to stop recirculation of hot exhaust.

• Seal cable openings in raised floors and containment.

• Keep the power and cooling capacity of each rack in step: a rack that can draw 30 kW but sits in an aisle cooled for 10 kW per rack will overheat.

14. Structural loading

A 3,000 lb AI rack on a 24 × 48 in footprint imposes an average load of:

That exceeds the uniform load rating of many raised floors and some elevated slabs. Engineers must check concentrated loads at casters or leveling feet, rolling loads during installation, and seismic anchorage. In seismic regions, racks and their anchorage are designed to ASCE 7 nonstructural component provisions, and tall racks need overhead bracing or floor anchorage. Many new AI halls use slab-on-grade with no raised floor for this reason.

15. Codes and standards that govern the rack

16. Monitoring, power capping and efficiency

Modern racks are instrumented. Intelligent PDUs report per-phase and per-outlet power; servers report their own power through baseboard management controllers; and data center infrastructure management (DCIM) systems combine both with temperature, humidity and leak data.

Power capping is a control layer on top of this measurement. The server or cluster management system limits processor and GPU power so that a rack never exceeds a set budget. It allows owners to oversubscribe electrical capacity safely — provisioning more servers than the nameplate sum would allow, because they rarely all peak at once — while guaranteeing that a feed never exceeds its continuous limit. The budget must account for failover: the cap should hold the rack within one feed's 80% capacity.

Facility efficiency is commonly reported as Power Usage Effectiveness:

Rack-level design affects PUE directly: higher distribution voltage reduces I²R losses, eliminating a transformer stage saves about 1–2%, high-efficiency power supplies save several percent, and warmer inlet temperatures (within the ASHRAE range) reduce chiller energy.

17. The AI rack — what changes

AI training racks combine all of the challenges above at once:

• Power. 100–140 kW per rack today, with roadmaps beyond 200 kW. At 415 V, 130 kW is about 185 A per rack — fed by multiple 60 A circuits or busway tap-offs per side.

• Power conversion inside the rack. AC feeds supply power shelves that convert to a 48–54 V DC busbar. At 54 V, 130 kW is about 2,400 A on the busbar, which is why designers are moving toward higher DC voltages; Keentel's article on 800 VDC rack fault protection covers that transition.

• Load dynamics. Synchronized GPU training produces fast, large power swings — tens of percent within seconds across thousands of GPUs. These swings propagate upstream to the UPS, generators and utility interconnection. Keentel's article on data center load calculations addresses how these swings affect utility studies.

• Cooling. Direct-to-chip liquid cooling is mandatory, with coolant distribution units, manifolds and leak detection in or beside the rack.

• Weight. Often more than 3,000 lb per rack, with floor and seismic implications.

• Network. Separate front-end and back-end (GPU) fabrics, with hundreds of high-speed optical links per rack.

18. Commissioning checklist for a new rack row

19. Common misconceptions corrected

20. Key takeaways

• The rack is the end of an electrical chain that starts at 138–345 kV. Design it with the same rigor.

• Size each feed to carry the whole rack alone, and keep normal per-feed loading at or below about 40%.

• Prefer 415Y/240 V for racks above about 10 kW: double the capacity per circuit and one-quarter the conductor losses.

• Check SCCR, selective coordination and arc-flash energy down to the rack PDU.

• Bond every rack to a proper mesh bonding network; grounding problems show up as network problems.

• Match cooling to power, rack by rack. Above about 30–50 kW, plan for liquid.

• Treat AI racks as a different class of load: higher power, heavier, liquid-cooled and dynamically variable.

21. How Keentel Engineering helps

Keentel Engineering provides electrical power systems engineering for data centers from the utility interconnection to the rack:

• Utility interconnection and substations. Load interconnection studies, 69–345 kV substation design, and coordination with utilities and ISOs/RTOs for large data center loads.

• Power distribution design. One-line diagrams, MV and LV distribution, UPS and PDU architecture (208 V or 415 V), busway and RPP layouts, and rack power density planning.

• Power system studies. Load flow, short-circuit and equipment duty evaluation, protective device coordination and selectivity, and arc-flash studies to IEEE 1584-2018 and NFPA 70E, with labels.

• UPS and battery engineering. Topology selection, battery sizing, VRLA-to-lithium-ion conversions, and NFPA 855 / UL 9540A code reviews.

• Grounding and bonding. Design and field verification to NEC Article 250 and TIA-607, and investigation of noise and bonding problems.

• AI and high-density readiness. Assessments of existing halls for 30–140 kW racks: power, cooling interface, floor loading and dynamic-load effects on UPS and utility.

• Commissioning support and owner's engineering. Design reviews, failover testing, and acceptance testing.

• EMT and dynamic modeling of large data center loads for utility studies.

Why clients choose Keentel Engineering

• Utility-to-rack perspective. Our engineers work across transmission interconnection, substations, power system studies and building distribution, so recommendations at the rack reflect the whole power chain.

• Study-driven design. Every recommendation is backed by calculations — load flow, short circuit, coordination, arc flash — not rules of thumb alone.

• US codes and standards expertise. NEC, NFPA 70E, NFPA 75/855/110, UL, IEEE and TIA, applied in practice.

• Licensed professional engineers with experience from 4 kV to 765 kV.

• Responsive, practical delivery for owners, developers, colocation providers and contractors.

Planning a high-density or AI rack deployment, or checking whether your existing feeds can survive a failover? Talk to Keentel Engineering at (813) 389-7871, contact@keentelengineering.com, or book a 15-minute scoping call at calendly.com/keentel-engineering/15min.

Part C — Case Studies

The following case studies are illustrative composites based on conditions Keentel Engineering commonly encounters. They do not describe any specific client or facility.

Case Study 1 — The redundant rack that went dark

Situation. An enterprise owner in the Southeast operated a 1.2 MW data hall with 208 V three-phase 30 A A/B feeds to each rack. Over three years, racks had grown from 5 kW to 10–12 kW as servers were refreshed. During a planned maintenance outage of the B-side UPS, six racks lost power within twenty minutes when their A-side PDU breakers tripped. The owner had believed the hall was fully redundant.

Analysis. Keentel reviewed two years of PDU metering data. Normal per-feed loading on the affected racks averaged 13–15 A per phase — 43–50% of the 30 A breakers. Each PDU appeared comfortable on the monitoring dashboard, which alarmed at 80%. In failover, the surviving feeds carried 26–30 A per phase, above the 24 A continuous limit. Phase imbalance made it worse: on several PDUs one phase carried 40% more current than the others. The monitoring system had never been configured with the failover condition in mind.

Engineered solution. Keentel recommended re-feeding the twelve densest racks with 415Y/240 V three-phase 30 A circuits (17.3 kVA usable per feed) from new PDUs with 480–415Y/240 V transformers, re-setting PDU alarms to warn at 40% and alarm at 45% of rating per feed, rebalancing phases on all remaining 208 V racks, and adopting a capacity rule in the change process that any new device must fit within one feed's 80% limit. A short-circuit check confirmed the new PDUs' 10 kA SCCR exceeded available fault current, and the arc-flash study and labels were updated.

Outcome. A repeat pull-the-plug failover test on every rack row passed with no trips; maximum single-feed loading in failover was 74% of breaker rating. The 415 V conversion also freed capacity for the next server refresh without new branch circuits, and the owner's operations team now reviews failover loading in its monthly capacity report.

Case Study 2 — Preparing an existing hall for 120 kW AI racks

Situation. A colocation provider in the central United States, served at 230 kV with 34.5 kV campus distribution, wanted to convert one 4 MW data hall from 8 kW air-cooled racks to 32 AI racks of about 120 kW each with direct-to-chip liquid cooling. The hall had 415 V distribution, a raised floor rated at 250 lb/ft², and centralized double-conversion UPS with VRLA batteries.

Analysis. Keentel's assessment found five constraints. First, 32 racks at 120 kW needed 3.84 MW of IT power plus about 0.5 MW for CDUs and residual air cooling — within the hall's UPS capacity only if redundancy moved from 2N to distributed redundancy. Second, existing RPP breakers could not feed 170 A per rack; 415 V overhead busway with 100 A tap-offs was needed. Third, the raised floor could not carry 3,200 lb racks; the rack row needed a slab-level steel support frame. Fourth, the UPS batteries were sized for 10 minutes at 4 MW; the AI load's fast power swings were within UPS ratings but would cause noticeable battery cycling during utility disturbances. Fifth, arc-flash energy at the new busway tap-off units exceeded 8 cal/cm² with existing upstream settings.

Engineered solution. Keentel designed dual 415 V busways per row (A and B), each sized so that one busway carries the full row load within its continuous rating; specified tap-off units with current-limiting fuses selectively coordinated with the busway plug-in breakers; added an energy-reducing maintenance switch scheme at the upstream switchboard; replaced the VRLA strings with lithium-ion modules supported by UL 9540A data and an NFPA 855 code review with the AHJ; and coordinated with the owner's mechanical engineer on CDU power feeds and leak-detection interlocks. Dynamic load profiles were supplied to the utility for its large-load study.

Outcome. The hall was converted in two phases with no outage to adjacent halls. Arc-flash energy at the tap-off units fell to below 4 cal/cm² with maintenance mode active. The lithium-ion conversion freed about 40% of the battery room floor area, which was reused for CDU equipment.

Case Study 3 — Network errors traced to a grounding problem

Situation. A regional services company operated an edge data room with eight racks at a branch campus in the Mid-Atlantic. Each rack had its own 3 kVA line-interactive rack UPS. Users reported intermittent slowness; network monitoring showed bursts of CRC errors and link resets on copper connections between two rows of racks, and two rack UPS units had failed their self-tests.

Analysis. Keentel's field investigation found that the room's PDU transformer had been replaced two years earlier without reinstalling the system bonding jumper, so the 208Y/120 V neutral was bonded to ground only at the building service, about 300 ft away. Neutral current was flowing on equipment grounding conductors, and voltage between the two rack rows measured 1.8 V RMS with significant high-frequency content. Several racks were bonded only through painted rails. Separately, the rack UPS units contained VRLA batteries that were five years old in a room averaging 30 °C (86 °F), well beyond their expected life at that temperature.

Engineered solution. Keentel restored the system bonding jumper and grounding electrode conductor at the PDU transformer per NEC 250.30, removed an improper downstream neutral-to-ground bond at a panelboard, installed a TIA-607-compliant secondary bonding busbar and bonded each rack with 6 AWG conductors and paint-piercing hardware, and replaced the eight rack UPS units with two lithium-ion rack UPS systems feeding A/B PDUs in a consolidated design, with battery monitoring tied into the facility BMS. Cooling setpoints were adjusted to hold rack inlet at 24 °C (75 °F).

Outcome. Rack-to-rack voltage fell to below 0.1 V, CRC errors stopped, and the network team closed a ticket that had been open for months. The new UPS design gave the room true A/B redundancy and a battery service life expected to exceed ten years.