Data Center Power & Cooling — Client Knowledge Base

A visual reference for electrical and HVAC equipment in modern data centers. Every component card follows the same shape: what it does, rule-of-thumb numbers, common solutions, major vendors, supply-chain status, key selling points, and the questions clients most often ask. Diagrams are interactive — click any component in the main one-line to jump to its section.

How this is organized

Two big sections — Power and Cooling — each with one card per equipment type. Supply-chain risk is flagged with a colored badge: RED means long industry lead times, AMBER means tight but workable, GREEN means generally available. The supply chain heatmap at the end shows SPM availability alongside industry context. Use the sidebar to jump anywhere.

DC fundamentals & the lingo customers use

Site sizing & the units that matter

Data center capacity is sold in megawatts of IT load — that is, the power the servers actually consume, not the utility feed. A "100 MW data center" means 100 MW of useful compute. The site itself draws more (cooling, UPS losses, lighting) — that ratio is PUE.

Term	What it means	Why it matters in a sales call
IT load	kW/MW consumed by servers, storage, networking	The number everyone sizes around
Critical load	IT + anything else on UPS (often network gear, sometimes cooling pumps)	Drives UPS sizing
Facility load	Everything: IT + cooling + losses + lighting	Drives utility feed and generator sizing
PUE	Facility load ÷ IT load. Lower is better.	Operating-cost story; sustainability story
Rack density	kW per rack. Legacy 5–10, modern 10–20, AI 30–150+	Everything downstream — cooling tech, busway sizing, floor loading — follows from this
N, N+1, 2N, 2N+1	Redundancy: N is just enough; N+1 one spare; 2N two independent paths	Tier classification & price; 2N roughly doubles equipment count
Tier I–IV	Uptime Institute classification (see standards section)	Customer asks "what tier is this for?" before specifying equipment
Concurrently maintainable	Any single component can be taken offline for service without dropping load	Tier III minimum requirement; affects topology
Fault tolerant	Any single failure does not affect load	Tier IV requirement

The numbers an engineer should have memorized

3,412

BTU/hr of heat per kW of IT load. Every watt in is a watt out as heat.

~285

CFM of cooling air per kW at a 12 °F (≈6.7 °C) ΔT — the classic air-cooling rule.

3.517

kW of heat removal per ton of refrigeration (1 ton = 12,000 BTU/hr).

1.25 – 1.5

PUE for a modern air-cooled facility. Best-in-class hyperscale hits 1.1.

5 min

Standard UPS battery autonomy. Long enough to start gensets and transfer, not run on battery.

10–20 s

Diesel genset time to reach rated voltage and frequency. Total ATS transfer ≤30 s typical.

Power — from the substation to the server

Power flows in one direction: utility → site substation → MV switchgear → step-down transformers → LV switchgear → UPS → PDU/busway → rack. Generators sit in parallel with the utility, switched in by ATS gear when the grid fails. Below is the canonical "one-line" you'll see in every customer drawing.

Interactive diagram. Click any component to jump to its section. Hover for a one-line description. Power flow is animated — A path (teal) and B path (red) are independent 2N systems. The cooling loop (blue) is shared. Heat (red dashed) returns from rack to atmosphere.

Four lanes, four flows. A path (teal) and B path (red) are independent power systems — each goes utility → substation xfmr → MV switchgear → unit xfmr → LV switchgear → ATS → UPS+battery → floor PDU → busway → rack. They stay electrically isolated all the way to the dual-corded server. Generators (amber) back-feed both MV buses through paralleling switchgear when utility fails. Cooling (blue) is a single loop shared by everything: cooling tower → chiller → pumps → branches to CRAH (air), CDU (liquid for chips), or RDHx (rack-door coil). Heat (red dashed) returns from rack to the tower. The animation shows direction of flow.

Utility feed & site substation Long lead

The utility-owned step from transmission down to medium voltage at the property line. For a 100+ MW site you're looking at a 115 kV or 138 kV transmission tap, an onsite or adjacent substation, and a utility tariff that locks in capacity over 15–30 years.

Rule of thumb

A modern hyperscale campus needs 1.2–1.6 MVA of utility capacity per MW of IT load after accounting for cooling, losses, and PUE. AI campuses now routinely request 500–1,500 MW.

Common solutions

Dedicated substation on the customer's property is the norm above ~30 MW; below that, a utility-owned pad-mount feed is fine. Dual feeds from independent substations are required for Tier IV.

Supply chain reality

Interconnection queues are the #1 schedule killer. In ERCOT, PJM, and most western utilities the queue is 3–7 years. Customers will pay enormous premiums to get power earlier — that's why behind-the-meter gas generation is suddenly hot.

Key selling angles

If you're not selling the utility piece, sell around it: on-site generation packages, MV switchgear that can ride through utility events, BESS for peak shaving. Customers stuck in the queue will buy anything that reduces utility dependence.

Discovery questions for the customer

What's your target IT load capacity and the ramp schedule to first power?

Why this matters: Sets the sizing baseline for the entire site. Phased ramps often justify modular MV gear and split unit-substation builds. Listen for announced vs. likely realized load — over-provisioning is common.

Have you secured your utility interconnect, or are you still in queue? Which ISO/utility?

Why this matters: ERCOT, PJM, MISO, and most Western utilities now show 3–7 year queues. If they're not at front-of-queue, behind-the-meter prime generation becomes a real opportunity — and the conversation shifts to gas gensets, BESS, and air permitting.

What tier are you designing to — Tier III (concurrently maintainable) or Tier IV (fault-tolerant)?

Why this matters: Tier IV requires dual independent utility feeds. Tier III can run a single feed with redundant generation. Drives the entire utility-side conversation including substation strategy.

Who owns the utility relationship — your team, the EPC, or a developer?

Why this matters: Identifies the decision-maker for substation scope. If a developer or land partner owns it, your role may be to support their submittals rather than lead.

Are you considering behind-the-meter generation to bridge any interconnect delay?

Why this matters: Opens up natural-gas reciprocating engines, gas turbines, and BESS as adjacent sales. The "bridge year" can be the most lucrative single project component.

Transformers (substation & unit) Critical bottleneck

Two main types in a data center: generator step-up / substation transformers (typically 30–80 MVA, 138 kV → 34.5 kV) and unit substation / pad-mount transformers (typically 1.5–3 MVA, 34.5 kV or 13.8 kV → 480 V). Both are oil-filled or cast-resin dry-type.

A transformer is just two coils sharing a magnetic core. Many turns on the HV side, few on the LV side — the voltage ratio matches the turn ratio. Power in ≈ power out, so the LV side carries proportionally more current (thicker wire, the heavy bushings on top of an outdoor pad-mount).

Rule of thumb

Size at 125% of expected load to cover starting inrush, non-linear load harmonics, and future growth. For 1 MW of IT load expect a 2.0–2.5 MVA unit transformer once cooling and redundancy are layered in.

Common solutions

Pad-mounted oil-filled for outdoor 1.5–3 MVA. Cast-resin dry-type for indoor or environmentally sensitive sites. K-rated (K-13 or K-20) windings to handle harmonic loading from UPS rectifiers.

Supply chain reality

The hardest single piece of equipment to get in 2026. Lead times have stretched from 24–30 months pre-2020 to 2–4 years today, with some high-voltage units quoted at 128 weeks or longer. The choke point is grain-oriented electrical steel (GOES) — Cleveland-Cliffs is the only U.S. producer. Generator step-up transformer demand is up 274% since 2019.

Key selling angles

Schedule, schedule, schedule. Customers will pay 30–50% premiums to shave 6 months off delivery. Lead with allocated factory slots, secondary-market refurbished units for swing capacity, and modular skid-mounted designs that pre-integrate the transformer with switchgear (faster install offsets long lead).

Discovery questions for the customer

What's your incoming utility voltage and your final IT distribution voltage (480 V or 415 V)?

Why this matters: Determines whether you're sizing a substation transformer (138 → 34.5 kV) or unit transformers (34.5 → 480 V), or both. 415 V is the modern hyperscale default; 480 V is still common in U.S. enterprise.

What unit-substation MVA size do you need, and at what redundancy level?

Why this matters: 1.5–3 MVA pad-mounts are most common. N+1 or 2N drives the count. Customer's answer reveals the power-block architecture upstream.

Indoor cast-resin dry-type, or outdoor pad-mount oil-filled?

Why this matters: Indoor sites with strict environmental requirements (no oil) push toward cast-resin. Outdoor sites default to oil-filled. Pricing and lead-time differ materially.

What's your required delivery date — and have you secured an allocation slot?

Why this matters: Lead time is THE conversation for transformers. If they say "we need it in 12 months" without an allocation, your schedule advantage (if any) is the entire sale.

What K-factor / harmonic rating do you need to handle UPS rectifier loads?

Why this matters: K-13 or K-20 windings are common for data center duty. Tells you whether they've sized for harmonic loading — often missed by EPCs new to the space.

Medium-voltage switchgear (5 kV–38 kV) Long lead

Indoor metal-clad or metal-enclosed switchgear that distributes MV power from the substation to the unit transformers, and connects the generator paralleling bus. Includes vacuum or SF₆-free circuit breakers, protection relays (SEL is the de facto standard), and bus tie arrangements.

A "lineup" of MV switchgear is a row of vertical cubicles sharing a common bus. Each cubicle holds a draw-out vacuum circuit breaker plus protection relays. Incoming from utility, feeders out to unit transformers, a bus tie for redundancy, and a generator-bus tie. Sized in 15 kV / 27 kV / 38 kV voltage classes.

Rule of thumb

One MV lineup per ~20 MW of IT load is a common slicing. Vacuum breakers are the norm; SF₆-free is now the default specification at hyperscalers due to PFAS regulations.

Common solutions

ABB UniGear ZS2 / Digital, Schneider RM AirSeT (SF₆-free), Siemens NXAIR, Eaton VCP-W, GE Vernova Vacuum Replacement Breakers. 15 kV class is the workhorse; 25 kV/38 kV for large utility-tie applications.

Supply chain reality

MV switchgear is the second-longest pole in the tent after transformers — typically 12–24 month lead. Eaton has committed ~$1.5B to expand North American switchgear and transformer capacity; Siemens opened a $190M Fort Worth low-voltage plant. Most of the new capacity comes online mid-2025 to 2026.

Key selling angles

SF₆-free is the easy lead (regulatory tailwind + customer ESG goals). Digital switchgear with arc-flash mitigation and integrated protection cuts commissioning time. Pre-engineered "e-house" skids that ship complete with switchgear + transformer + relays already wired and tested.

Discovery questions for the customer

What voltage class are you specifying — 15 kV, 27 kV, or 38 kV?

Why this matters: 15 kV is the most common workhorse; 25/38 kV used on large-utility-tie applications. Determines breaker rating, cubicle size, and the entire vendor short-list.

Is SF₆-free vacuum mandatory in your spec, or are F-gas alternatives acceptable?

Why this matters: ESG-driven customers and EU/state-regulated sites usually require SF₆-free. If they haven't made it mandatory, raising it gives you an easy upgrade conversation.

How many feeders plus generator-tie and bus-tie cubicles do you need?

Why this matters: Gets you the cubicle count for quoting. Also reveals whether they've planned for a future feeder (spare cubicle = good practice).

What protection relay platform are you standardizing on — SEL, ABB REF, Siemens SIPROTEC?

Why this matters: SEL is the de facto standard in U.S. data centers. Customer preference affects integration with their DCIM and existing fleet.

What arc-flash mitigation are you specifying — remote racking, IR windows, arc-resistant construction?

Why this matters: Safety story sells. Arc-resistant switchgear and remote-racking robots are easy add-ons that customers care about more than they admit.

LV switchgear, switchboards & paralleling gear Tight

480 V (US) or 415 V (EU/global) gear that takes the secondary of unit transformers and distributes to UPS inputs and mechanical loads. Generator paralleling switchgear (digital governors + load-share controls) sits at the same voltage level.

Rule of thumb

Sized at 2,000 A–5,000 A main bus per ~2 MW block. Selective coordination (cascading trip curves) is mandatory for Tier III/IV. 415 V is replacing 480 V in new hyperscale builds because 240 V rack outlets become available without rack-level step-down transformers.

Common solutions

ANSI-rated metal-enclosed switchgear, draw-out air circuit breakers (Eaton Magnum, Schneider Masterpact, ABB Emax 2, Siemens 3WL). Paralleling gear from ASCO 7000, Russelectric, Cummins PowerCommand, Caterpillar EMCP.

Supply chain reality

Better than MV — typically 6–12 months. Air circuit breakers themselves have been pinch points. Domestic content for IRA tax credits is a new wedge: customers ask for U.S. assembly point of origin.

Key selling angles

Arc-flash safety (remote racking, infrared windows, predictive maintenance via IoT sensors). Integrated power monitoring (Schneider PowerLogic, Eaton Foreseer) — every customer wants per-circuit metering. Factory-tested paralleling gear with full FAT documentation accelerates commissioning.

Discovery questions for the customer

480 V or 415 V system voltage?

Why this matters: 415 V is the modern hyperscale direction (eliminates rack-row step-down transformers and improves PUE 1–2 points). 480 V is still common in U.S. enterprise. If they're still 480 V, there's a conversation about migrating.

What's your main bus amperage rating, and minimum short-circuit AIC requirement?

Why this matters: Drives the breaker selection (Magnum, Masterpact, Emax 2, 3WL). AIC rating reflects the upstream system fault-current — often higher than EPCs initially estimate.

Is selective coordination a hard requirement?

Why this matters: Required for Tier III/IV and NEC emergency systems. If they don't have a coordination study, that's a service-revenue gap and a chance to add value beyond the metal.

Do you need integrated paralleling controls for the generator bus, or is that handled separately?

Why this matters: Integrated paralleling (ASCO 7000, Russelectric, EMCP) bundled with LV switchgear simplifies commissioning. If they're treating these as separate scopes, propose the bundle.

What per-circuit metering and DCIM platform are you standardizing on?

Why this matters: Every customer wants metering now. Identifies whether you can attach Schneider PowerLogic / Eaton Foreseer / Vertiv monitoring or need to integrate with an existing third-party DCIM.

UPS systems Tight, improving

The bridge between utility/genset and the IT load. In a data center context, "UPS" essentially always means online double-conversion — AC in → DC rectifier → battery bus → DC-AC inverter → AC out. The load is always fed from the inverter, so transfer time on utility loss is zero.

N+1 shares a common output bus — cheaper but single-bus failure mode. 2N has two electrically isolated paths feeding both PSUs of every server — the modern hyperscale default.

Rule of thumb

Size UPS at ~1.1× expected IT load; load each UPS module to 40–60% of nameplate in N+1 (so any one failure leaves the rest below 80%). Module sizes: 250 kW, 500 kW, 750 kW, 1.0 MW, 1.25 MW are the common increments. Hyperscale prefers 1.25 MW modules in 2N pairs.

Common solutions

Online double-conversion with IGBT rectifiers and transformerless design — 96–97% efficiency at half load, 98%+ in eco/ECOnversion modes. Lithium-ion batteries are now the default for new builds (3–5× cycle life, ¼ the floor space vs VRLA).

Supply chain reality

6–12 months for large modular UPS — has improved from 2022–2023 peak. Vertiv reports a $9.5B backlog. Lithium battery supply is tied to the same constrained EV cells; nickel-zinc (Vertiv-ZincFive partnership, announced April 2025) is the emerging alternative for cabinet batteries. ABB launched MegaFlex UL 415V in June 2025 for AI-scale loads.

Key selling angles

Efficiency: 97.5%+ double-conversion, 99% in ECOnversion mode. On a 1 MW UPS each percentage point saves ~$10K/year in electricity.
Lithium-ion batteries: 10-year warranty, smaller footprint, telemetry built in. Lifecycle cost beats VRLA in 4 years.
Modular architecture: pay-as-you-grow, hot-swap power modules, no downtime for capacity adds.
Grid-interactive (UPSaaS): ABB, Schneider, Eaton all now offer battery-to-grid services that monetize the UPS battery during normal operation. Real revenue, not just savings.
AI-ready: tolerates the dynamic load steps from GPU training cycles without nuisance trips — a real differentiation point in 2026.

Discovery questions for the customer

What's your critical IT load size, and the redundancy topology — N+1, 2N, or 2(N+1)?

Why this matters: Drives module count and total UPS megawatts. 2N doubles equipment count vs. N+1. Hyperscale defaults to 2N; enterprise often runs N+1.

What module size do you prefer — 250 kW, 500 kW, 1 MW, or 1.25 MW?

Why this matters: Larger modules = lower per-kW capex; smaller modules = better part-load efficiency. Hyperscale leans 1–1.25 MW; smaller deployments often want 250–500 kW for granularity.

Are you ready for lithium-ion (LFP) batteries, or does policy still require VRLA?

Why this matters: Li-ion is the standard for new builds but some risk-averse enterprises still require VRLA. If they want VRLA, raise the 10-year TCO argument and the floor-space story.

Are you running AI training workloads with dynamic GPU load steps?

Why this matters: AI workloads stress older UPS control loops badly. Newer platforms (ABB MegaFlex, Schneider Galaxy VXL, Vertiv Trinergy) handle the transients. This is a real differentiation point in 2026.

What's your day-1 expected utilization vs. design capacity?

Why this matters: Day-1 load is often 20–40% of design. Modular UPS architectures let modules be turned off at low load to keep operating in the efficiency sweet spot. Drives the part-load efficiency conversation.

Any interest in UPSaaS / grid-interactive battery services?

Why this matters: Battery-to-grid revenue is real in NYISO, PJM, ERCOT. A 1 MWh UPS battery can earn $30–80K/year while still meeting backup duty. Opens an ongoing service relationship.

UPS batteries (VRLA, Li-ion, NiZn) Tight on Li-ion

The energy storage that bridges the gap between utility loss and generator pickup. Sized for runtime, but in modern designs runtime is short by design — the goal is to start the gensets, not run on battery.

Same 1 MW UPS load, same 5-minute runtime. VRLA needs racks of jars in a dedicated battery room with H₂ ventilation and spill containment. Modern Li-ion (LFP) cabinets sit next to the UPS, take ~25% of the floor space, last 2–3× longer, and report cell health continuously.

Chemistry	Energy density	Cycle life	Cost	Best for
VRLA (lead-acid)	Low	200–500 cycles, 5–8 yr	$	Retrofits, low cycle count, budget-driven
Li-ion (LFP)	High	3,000+ cycles, 10–15 yr	$$	Default for new hyperscale & AI builds
Li-ion (NMC)	Highest	1,500+ cycles, 10 yr	$$	Where density is critical (less common in DC due to thermal runaway)
Nickel-Zinc	Medium	2,000+ cycles, 10 yr	$$	High-power short-duration; emerging hyperscale option
Flywheel	n/a	20+ yr mechanical	$$$	Sub-30s ride-through, no battery maintenance

Rule of thumb

5 minutes at full load is the industry-standard runtime. Anything more is wasted capex unless the customer has a specific reason (e.g., 15-min runtime for sites without gensets, or 30-min for grid-services arbitrage).

Supply chain reality

Li-ion cells compete with EV demand — pricing is volatile, lead times 4–9 months. VRLA is more available but losing share fast. NiZn is supply-constrained because production volume is still small.

Key selling angles

Total cost of ownership over 10 years — Li-ion wins on space, replacement cycles, and monitoring. Fire/safety story — LFP chemistry is thermally stable, NiZn is non-flammable. Grid services revenue is the new ROI hook: 1 MWh of battery in NYISO ICAP can earn $30–80K/year while still meeting UPS duty.

Discovery questions for the customer

What battery runtime are you specifying — 5 minutes (standard) or longer?

Why this matters: If longer than 5 min, ask why. Could be a legitimate grid-services case or just over-spec. Either way, runtime drives capex significantly.

Is footprint / floor space a critical constraint at your site?

Why this matters: Li-ion takes ~25% of VRLA's footprint. Tight sites are pre-sold on Li-ion before you even start the TCO argument.

Does your corporate or insurance policy place any restriction on Li-ion in a UPS room?

Why this matters: Some risk-averse enterprises still ban Li-ion or require specific NFPA 855 / IFC 1207 measures. Identifies whether you need to bring fire-suppression and compliance materials to the table.

Any interest in monetizing the battery via grid services / UPSaaS?

Why this matters: Real revenue: 1 MWh of UPS battery in NYISO ICAP earns $30–80K/year. Opens a recurring service and an ROI conversation that justifies Li-ion upgrade.

What BMS telemetry / integration do you need with your DCIM?

Why this matters: Modern Li-ion cabinets stream cell-level data. If they're already running Schneider EcoStruxure, Vertiv iCOM, or a third-party DCIM, integration scope is part of the sale.

Standby & prime generators 12–24 mo lead

Diesel reciprocating engines remain the workhorse. Large data centers use 2–3 MW units paralleled in N+1 banks. Natural gas reciprocating and turbines are gaining ground for "prime power" applications where the customer can't wait for utility interconnection.

A standby generator "package" is an engine + alternator on a common skid, with a radiator for engine cooling, a sub-base fuel tank for autonomy, an exhaust silencer / SCR for emissions, and a digital control panel. The whole thing sits inside a sound-attenuated walk-in enclosure.

Rule of thumb

Size at 110–125% of facility load (IT + cooling + losses) to cover transient steps and altitude/temp derating. 2.5 MW is the most popular block size — large enough to be efficient, small enough to maintain N+1 economics. Total genset capacity typically equals 1.3–1.5× IT load.

Common solutions

Tier 4 emissions-compliant diesel (Cummins QSK60/QSK95, Caterpillar 3516E/C175-20, MTU Series 4000, Mitsubishi MGS), enclosed in walk-in sound-attenuated housings with belly tank or sub-base fuel storage (24–72 hour autonomy). Natural-gas reciprocating engines (Cummins HSK78G, Cat G3520, INNIO Jenbacher) for behind-the-meter prime power.

Supply chain reality

One of the longest-lead items behind transformers and MV gear. Cummins QSK95 quoted at ~18 months and they announced a $150M Fridley, MN expansion in Feb 2026 to add 30% more output. Caterpillar grew data-center engine capacity 125% in two years and bought Weichai Baudouin's European distribution in Nov 2025. Global data center generator market: $9.5B (2025) → $17.0B (2035).

Key selling angles

Schedule: if you have allocated factory slots, that is your sale.
Fuel flexibility: dual-fuel (diesel + natural gas) or 100% NG for sites that need to run prime to bypass utility queues.
HVO / renewable diesel: drop-in renewable fuel meets corporate ESG without engine modifications.
Emissions packages: SCR + DPF + DOC for Tier 4 final, Bay Area / Texas air-district permitting expertise.
Black-start ability + paralleling controls: integrated solution sells better than a bare engine.

Discovery questions for the customer

What total backup capacity do you need, and what block size — 2 MW, 2.5 MW, or 3 MW units?

Why this matters: Drives count and N+1 economics. 2.5 MW is the most common sweet spot. Smaller blocks = more redundancy granularity; larger blocks = fewer units to maintain.

Diesel, natural gas, or dual-fuel?

Why this matters: Natural gas is rising for prime power. Dual-fuel hedges fuel-availability risk. Diesel is still default for pure standby. Question reveals their fuel strategy.

Standby duty only, or prime / continuous duty (behind-the-meter)?

Why this matters: Prime rating derates capacity ~15–20% vs. standby. Customers in utility queues increasingly need prime — and pay a premium for it.

What sound attenuation level at the property line — 65, 70, 75 dBA?

Why this matters: Drives enclosure scope. Urban / suburban sites usually have local ordinances. If they don't know, they probably have a permitting risk you can flag.

Emissions tier — Tier 4 final with SCR/DPF, or is the site exempt?

Why this matters: Tier 4 adds significant cost (urea/AdBlue infrastructure included). California, Texas air districts, and northeastern states are strictest. Reveals permitting maturity.

Required fuel autonomy — 24, 48, or 72 hours?

Why this matters: Drives sub-base tank size or external bulk storage scope. 24h is common; 72h for sites in disaster-prone regions or with strict resilience requirements.

Are HVO / renewable diesel commitments part of your ESG roadmap?

Why this matters: Renewable diesel drops in without engine mods and yields ~90% carbon reduction. Easy ESG win — flag it even if not asked.

Automatic Transfer Switches (ATS) Tight

The gear that transfers load from utility to generator (and back). Modern hyperscale uses closed-transition (make-before-break) ATS in paralleling switchgear, but mechanical ATS still dominates 480 V branch transfers.

Open transition leaves a brief power gap on every transfer — that's why a UPS is always upstream. Closed transition adds the ability to retransfer back to utility with no gap, useful for monthly testing or grid services. On a true utility failure both types still see the unavoidable utility-loss gap.

Common solutions

ASCO 7000 series (Schneider), Russelectric (Siemens), Cummins PowerCommand, Caterpillar CTS/ATC, Kohler KEP/KCT, Eaton ATC. 4-pole switching is standard for separately derived systems (neutral transfer).

Selling angles

Closed-transition (no momentary outage during retransfer test), digital ATS with grid-monitoring, integrated with paralleling gear. Service contracts — ATS testing is a recurring revenue engine.

Discovery questions for the customer

Closed-transition (no power gap on retransfer) or open-transition acceptable?

Why this matters: Closed-transition lets the team test gensets monthly under live load with no UPS event. Customers who run frequent maintenance always prefer it. Open-transition is fine for sites with disciplined UPS coverage.

Is 4-pole switching (including neutral) required for separately derived systems?

Why this matters: Standard in U.S. data centers with separately derived generator systems. Confirms NEC compliance scope.

Standalone branch-level ATS, or integrated within the paralleling switchgear lineup?

Why this matters: Modern hyperscale designs put transfer logic into the paralleling gear and skip discrete ATS. Smaller deployments still use ASCO 7000 / Russelectric / Cummins branch units.

Are you planning a service contract for annual ATS testing?

Why this matters: Recurring service revenue. ATS testing is a regulated annual activity; bundle with the equipment or sell separately.

PDUs, RPPs, and overhead busway Tight on busway

"PDU" means two different things. In the white space, a PDU (power distribution unit) is a floor-mount cabinet that takes 480 V or 415 V and breaks it down to rack-level circuits via panelboards. RPP (remote power panel) is the same idea further downstream. Rack PDU is the strip in the back of the cabinet that the server cords plug into.

Two parallel busway runs (A and B) at the ceiling, fed from independent UPS systems. Each rack gets a pair of tap-off boxes that drop cables down to the two vertical rack PDUs (the strips servers plug into). To add a new rack: snap a new tap-off onto the busway — no conduit, no electrical work, minutes not days.

Rule of thumb

One PDU per 200–500 kW of rack load. Modern designs run 415 V → 240 V single-phase rack PDUs, eliminating the in-row step-down transformer and saving 1–2% on PUE. Busway is replacing conduit for runs over 200 A — faster install, tap-off plugs anywhere, easy to reconfigure for rack-density growth.

Common solutions

Floor PDUs from Schneider, Vertiv, Eaton (with main breaker, surge suppression, sub-feed breakers, monitoring). Rack PDUs from Server Technology (Legrand), Raritan (Legrand), Vertiv Geist, APC AP series, Enlogic, Panduit. Busway from Starline (Legrand), Universal Electric, Schneider Canalis, Eaton Pow-R-Way.

Supply chain reality

Floor PDUs: 12–20 weeks. Rack PDUs: 8–16 weeks. Busway has been particularly tight (copper bus + custom lengths) — 20–30 weeks. AI deployments are buying high-amperage 60–100 A rack PDUs at scale, which has tightened that segment.

Key selling angles

Outlet-level metering (per-server kWh visibility), remote outlet switching for reboots, environmental sensors built into the PDU strip, color-coded A/B cord PDUs, high-density 3-phase 415 V rack PDUs for 30 kW+ cabinets. Busway: speed of deployment is the killer feature — drop a tap-off anywhere along a run in 15 minutes.

Discovery questions for the customer

What's your current rack density, and what are you projecting in 24 months?

Why this matters: Drives PDU/busway amperage. 5–10 kW racks just need standard 30 A; 30+ kW AI racks need 60–100 A rack PDUs. Future density tells you whether to oversize now.

Overhead busway, or hardwired conduit feeders?

Why this matters: Busway is the modern hyperscale default for flexibility. Older sites still spec conduit. If they're new to busway, this is an educational sell.

Are you running 415 V → 240 V at the rack, or staying with 480 V → 208 V?

Why this matters: 415 V eliminates the in-row step-down transformer and saves PUE points. Determines rack PDU spec entirely.

Do you require outlet-level metering and remote outlet switching?

Why this matters: Modern intelligent rack PDUs (Server Tech, Raritan, Vertiv Geist, APC) offer per-outlet kWh + remote reboot. Standard for hyperscale; valuable add-on for enterprise.

Color-coded A/B cord rack PDUs?

Why this matters: Hyperscale ops teams want clear visual A/B distinction to prevent miswiring during installs and maintenance. Cheap upgrade with big operational value.

Power section test (archived)

1A customer has 30 MW of IT load. What's a reasonable estimate of the utility feed capacity they need to request?

Answer: 36–48 MVA. Rule of thumb is 1.2–1.6 MVA of utility per MW of IT to account for PUE (1.3–1.5 typical), losses, and headroom. At PUE 1.4 with a 90% power factor and 10% headroom: 30 × 1.4 / 0.9 × 1.1 ≈ 51 MVA. Conservative answer is ~45 MVA.

2Why is grain-oriented electrical steel (GOES) suddenly a sales concern?

Answer: It's the magnetic core material for power transformers. Cleveland-Cliffs is the only U.S. producer. Demand is up 119–274% since 2019, lead times have stretched to 2–4 years, and there's no quick substitute. Any conversation about transformer schedules eventually lands here.

3What does "concurrently maintainable" mean and what tier requires it?

Answer: Any single capacity component or distribution element can be removed from service for maintenance without taking the load offline. That's the Tier III definition. Tier IV adds fault tolerance — any single failure (not just planned maintenance) doesn't drop load. Tier III drives N+1; Tier IV drives 2N or 2(N+1).

4Customer asks: "Why should I move to 415 V from 480 V?"

Answer: 415 V three-phase gives you 240 V single-phase line-to-neutral, which is exactly what modern server PSUs run on most efficiently. You eliminate the in-row step-down transformer (480→208), which saves roughly 1–2 points on PUE, reduces capex, and frees floor space. It's also the global standard (most of the world is already 400/415 V), so equipment is broadly available.

5What's the difference between online double-conversion UPS and line-interactive?

Answer: Double-conversion runs the load continuously through rectifier-inverter — the AC waveform is rebuilt from DC, so there's no transfer event when utility fails (zero break time) and the output is conditioned regardless of input quality. Line-interactive feeds the load directly from utility through a buck/boost transformer and only switches to inverter on actual outage (a few ms break). All data-center UPS are double-conversion.

6Customer says, "We want 30 minutes of battery runtime." How do you respond?

Answer: Ask why. Standard practice is 5 minutes because gensets start in 10–20 seconds and you want enough margin for a couple of failed start attempts. 30 minutes triples the battery cost and floor space. Legitimate reasons: site has no gensets (rare); customer wants to monetize the battery in grid services (legitimate — sell them on UPSaaS); they want to ride out short utility events without running gensets (debatable economics). Push back politely and align on the why.

7You're quoting a 2 MW UPS in N+1 with 1 MW modules. How many modules?

Answer: Three. Two carry the 2 MW load, one is the +1. Each module runs at ~67% loaded (2 MW ÷ 3 modules). With any one out of service the other two run at 100% — at the upper end but still within rating. If you wanted to keep modules under ~80% even during a failure, you'd use four 1 MW modules (load = 50%, failure = 67%).

8Why do generator sizes cluster at 2–3 MW for data centers?

Answer: Sweet spot of engine efficiency, transportability (still fits a single truck/skid), reasonable N+1 economics, and matches the typical 2 MW "power block" architecture that hyperscalers use to slice up a data hall. Below 2 MW you carry too many gensets; above 3 MW you lose granularity for N+1 and the single-engine failure mode becomes punishing.

9What's the difference between "standby" and "prime" rating on a generator nameplate?

Answer: Standby = variable load, no overload, limited run hours per year (typically <200), only for emergency backup. Prime = continuous variable load, no run-hour limit but typically 70–80% average load factor. A standby-rated 3 MW genset might only be a 2.5 MW prime-rated unit. The "behind-the-meter prime power" trend is forcing customers to spec prime-rated engines, which derates capacity vs. classic standby duty.

10A customer is choosing between VRLA and Li-ion batteries for a 1 MW UPS. They care about capex. What's your pitch?

Answer: Show the 10-year TCO, not the day-one capex. VRLA needs full replacement every 5–7 years (often two replacements over the UPS lifetime); Li-ion lasts 12–15 years on a single set. VRLA needs a dedicated battery room with ventilation and spill containment; Li-ion can go in a cabinet next to the UPS, saving ~70% of the floor space. Telemetry is built into Li-ion cells — you see degradation in real time instead of discovering it during the next discharge test. By year 4 Li-ion is cheaper. If they truly can't move on capex, sell them VRLA and a Li-ion upgrade path.

11What's "selective coordination" and why do customers ask about it?

Answer: The protective device closest to a fault should be the only one that trips, isolating the smallest possible area. Achieved through trip-curve coordination across the upstream chain (instantaneous, short-time, ground-fault settings cascaded). Required by NEC for emergency systems and for Tier III/IV. Without it a fault on one rack can take down the whole switchgear lineup.

12Customer is in the PJM interconnection queue with a 4-year wait. What do you sell them?

Answer: Behind-the-meter prime power — natural-gas reciprocating engines (Cat G3520, Cummins HSK78G, INNIO Jenbacher) or gas turbines (Solar Turbines Mercury 50, Mars 100). Pair with a substation-grade BESS to handle dynamic load steps. Get the site running on gas at year 1, transition to utility primary with gas as backup once the interconnect closes. There's also an air-permitting story to navigate — that's a service revenue opportunity for the right partner.

13What is "ECOnversion" or "eco-mode" on a UPS and why is the customer cautious about it?

Answer: The UPS bypasses the rectifier-inverter path during clean utility and feeds the load directly through the static switch, with the inverter on standby. Efficiency jumps from ~97% to ~99%. The catch is the transfer back to inverter on utility loss takes a couple of milliseconds — which a properly designed IT PSU rides through, but cautious customers worry about it. Best-practice answer: use ECOnversion on UPS units feeding non-mission-critical loads, or on hyperscale fleets that are tolerant of microbreaks.

14Why is SF₆-free switchgear suddenly a default spec?

Answer: SF₆ is the most potent greenhouse gas regulated (GWP ≈ 23,500). EU F-Gas Regulation, EPA proposals, and state-level PFAS regulations are phasing it out. Vacuum (medium voltage) and clean-air or fluoronitrile alternatives (Schneider's AirSeT, Hitachi's EconiQ, GE's g³) have caught up technically. Customers with corporate ESG targets specify it; some utility tariffs now penalize SF₆.

15What's the difference between a closed-transition ATS and an open-transition ATS?

Answer: Open transition breaks the connection to the source before making the new connection — the load experiences a brief outage (typically 100–300 ms) that the UPS rides through. Closed transition makes the new connection before breaking the old one, allowing live transfer with no outage — but requires the generator to be synchronized with utility, so it's only used for planned retransfer back to utility or for monthly genset test load transfers. Most customers want closed-transition for the operational benefit of testing under live load.

16You're at a meeting and the customer says, "We're seeing nuisance trips on UPS during AI training runs." What's likely going on?

Answer: GPU training workloads have brutal dynamic load steps — racks can swing from 30% to 100% in milliseconds when a synchronous training job kicks off, and back when it checkpoints. Older UPS rectifier and inverter control loops weren't tuned for these transients and either trip on overload or feed back grid disturbance that fights the generator. Fix: newer UPS platforms (ABB MegaFlex, Schneider Galaxy VXL, Vertiv Trinergy) have AI-aware control loops; right-size the modules so loaded margin is higher; consider DC-bus tied designs that smooth the input.

17Why is busway gaining share over hardwired conduit feeders?

Answer: Speed and flexibility. Conduit is custom-cut, terminated, and tested on site — weeks of labor. Busway sections are factory-made, snap together, and let you drop a tap-off plug anywhere along the run for a new rack — minutes. As AI deployments retrofit existing data halls to higher rack densities, the ability to add 100 A circuits without re-pulling conduit is decisive. Starline (Legrand), Universal Electric, Schneider Canalis are the brands to know.

18A customer asks about UPS efficiency at "low load." Why?

Answer: Day-1 utilization in a new data hall is often 20–40% of design — racks haven't all moved in yet. A UPS that's 97% efficient at 50% load might be only 92% at 20% load, which is real money in opex. Modular UPS architectures let you turn off entire modules at low load, keeping the active modules in the sweet spot. Lead with the part-load efficiency curve, not the peak nameplate number.

19Customer wants to know if Li-ion batteries are a fire risk in a UPS room.

Answer: LFP (lithium iron phosphate) — the chemistry used in essentially every modern UPS — is thermally stable and very hard to push into thermal runaway. NMC (used in EVs and some older UPS) is more energetic and more prone. Modern UPS Li-ion cabinets include cell-level fusing, BMS thermal monitoring, and often a built-in suppression agent (Stat-X, Novec) plus venting. NFPA 855 and IFC 1207 are the relevant codes. The honest answer is yes there's a risk, no it's not unmanageable, and the data center insurance market has accepted it.

20Apply it all: A customer wants to build a 50 MW AI training campus in central Texas, online in 24 months. Walk through your equipment shopping list and the biggest risks.

Answer: Power load model: 50 MW IT × ~1.3 PUE = ~65 MW facility. Generator capacity ~75–80 MW (16 × 2.5 MW units, N+1 in two blocks). Utility request ~75 MVA. UPS: 50–60 MW in 1.25 MW modules, 2N (two systems of ~30 MW each, lithium-ion). MV switchgear: 13.8 kV or 34.5 kV class, SF₆-free vacuum. Transformers: 1× 75 MVA substation, ~25 × 2.5 MVA pad-mounts.

Risks (in order):

Substation transformer: 2–4 year lead. Solve with a refurbished/used 75 MVA unit or get on an allocation list immediately.
Utility interconnect (ERCOT): typically 18–36 months. May force prime-power gas gensets to bridge.
Generators: 16 × 2.5 MW units in 24 months requires Caterpillar or Cummins allocation slots booked at contract signing.
MV switchgear: 12–18 months — order with the transformer.
Cooling (see next section): AI workloads at 50 MW likely require liquid cooling; CDUs and cold plates are themselves supply-constrained.

Selling angle: sell schedule certainty (allocated factory slots, prefabricated power skids that combine transformer + MV gear + LV gear) rather than the cheapest unit price. The customer's opportunity cost on a 6-month delay is enormous.

Cooling — moving heat from chip to atmosphere

Cooling is half the operational cost of a data center and the discipline that's changing fastest. The fundamental task is to move heat from a 1–2 cm² silicon die out to the sky — a chain of thermal resistances and fluid loops, each of which has its own equipment, vendors, and selling story.

Heat load fundamentals

Every watt of electricity into a server comes out as heat. You don't "cool" a data center so much as move that heat outdoors. The chain looks like this:

Chip → die TIM → heat sink/cold plate → rack air or coolant → CRAH/CDU → chilled water → chiller → condenser water → cooling tower → atmosphere

Each link has a temperature drop. The hotter you can run the chip-side (ASHRAE A1/A2/H1 envelopes), the warmer your supply water can be, the more hours per year you can run "free cooling" without a chiller. This is the single biggest lever on PUE.

The point of PUE is that everything to the right of "IT load" is overhead — money spent moving heat, conditioning power, and keeping the lights on. Modern cooling design (warm supply temps, free cooling, liquid cooling) shrinks the cooling bar dramatically; modern UPS technology (high-efficiency double conversion, LFP batteries, modular load matching) shrinks the UPS bar.

PUE, WUE, and the metrics customers track

Metric	Formula	Typical	Best-in-class
PUE (Power Usage Effectiveness)	Facility kWh ÷ IT kWh	1.4 air-cooled / 1.2 modern	1.05–1.10 (hyperscale liquid)
WUE (Water Usage Effectiveness)	L water ÷ IT kWh	1.0–1.8 L/kWh (cooling tower)	0 (air-cooled chiller or dry cooler)
pPUE (partial PUE)	Same as PUE but for a hall/zone	—	—
ERE (Energy Reuse Effectiveness)	(Facility – reused) ÷ IT	1.4	<1.0 (heat reuse to district heating)

PUE 1.5 means 50% overhead — every 1 kW of compute costs 1.5 kW at the meter. The cooling system is most of that overhead, which is why every cooling sales call eventually becomes a PUE conversation.

The bullseye is "recommended" (18–27 °C dry-bulb, 8–80% RH). Around it are allowable envelopes A1–A4 — each looser, each unlocking more free-cooling hours. The industry has been pushing setpoints from the cold corner of recommended (~20 °C) to the warm edge (~27 °C) over the last 15 years, and many hyperscale halls run into the A2 zone on purpose.

ASHRAE thermal envelopes — the customer's spec book

Class	Recommended inlet	Allowable	Typical use
A1	18–27 °C (64–81 °F)	15–32 °C / 8–80% RH	Enterprise, mission-critical
A2	18–27 °C	10–35 °C / 8–80% RH	Most commercial DCs
A3	18–27 °C	5–40 °C / 8–85% RH	Some volume hyperscale
A4	18–27 °C	5–45 °C / 8–90% RH	Hyperscale extreme
H1 (high-density)	18–22 °C tighter	n/a	AI / HPC, high power density
W class (liquid)	2–45 °C supply water (W1–W5)	W4: up to 45 °C, W5: >45 °C	Direct liquid cooling

Modern hyperscale runs cold-aisle at the top of A1/A2 recommended (~27 °C / 80 °F) on purpose — every degree warmer = more free cooling hours = lower PUE.

Discovery questions — cooling strategy overall

What PUE and WUE targets are you committed to publicly or internally?

Why this matters: Anchors every cooling decision. PUE drives architecture choice (air vs. liquid, chilled-water setpoint, free cooling). WUE drives tower vs. dry cooler. Customers with ESG commitments need these numbers from us.

What rack density profile are you designing for — average and peak per row?

Why this matters: <10 kW = simple air cooling. 10–30 kW = containment + in-row. 30–80 kW = RDHx or DLC. 80+ kW = DLC mandatory. Density dictates the cooling stack.

What inlet temperature setpoint are you running — 22, 24, 27 °C?

Why this matters: Higher setpoints unlock more free-cooling hours and lower PUE. If they're at 22 °C, there's an immediate optimization opportunity. If they're at 27 °C, they're modern.

Any heat-reuse opportunity at the site — adjacent district heating, industrial process, on-site greenhouse?

Why this matters: ERE<1.0 is achievable with reuse. Strongest in EU markets (mandatory in some jurisdictions). Worth raising in every conversation, especially in cold climates.

CRAC and CRAH units Generally available

The white-space cooling unit. CRAC (computer room air conditioner) has its own DX refrigeration loop with a remote condenser. CRAH (computer room air handler) is just a coil + fan; chilled water comes from a central plant. CRAH dominates anything above ~1 MW because central chillers are more efficient than dozens of distributed DX units.

Same cabinet shape, fundamentally different machine. A CRAC carries the full refrigeration cycle (compressor, expansion valve, evaporator) and rejects heat through refrigerant lines to a remote condenser outside. A CRAH is just a coil and a fan — chilled water comes in cold from a central plant, leaves warmer. Above ~1 MW the central-plant CRAH approach wins on efficiency and serviceability.

Rule of thumb

A 30-ton CRAH (~105 kW) cools ~80–90 kW of IT load at typical airflow. Plan one CRAH per ~80 kW of IT, plus N+1. EC fans (electronically commutated) are universal now — vs. older belt-drive AC units, they cut fan energy by 30–50%.

Common solutions

Vertiv Liebert PCW/PDX, Schneider/APC Uniflair, Stulz CyberAir, Airedale OnRak/UltimateAir, Munters. Down-flow (under raised floor) is traditional; in-row CRAH (Vertiv CRV, Schneider InRow, Stulz CyberRow) sits between rack rows for high-density (10–30 kW/rack); fan walls replace dozens of CRAHs in greenfield hyperscale.

Supply chain reality

Standard CRAH units are 12–20 weeks. Custom large air handlers and fan-wall units are 30+ weeks. Vertiv reports $9.5B backlog across the cooling portfolio. Tariff pressure on steel and aluminum (heat-exchanger coils, frames) raised prices ~8–15% in 2025.

Key selling angles

EC fans + variable-speed control + AI-based airflow optimization. Integrated dewpoint/temperature sensing across the data hall (Vertiv iCOM, Schneider EcoStruxure IT). For retrofits, in-row units that sit beside the rack and target high-density pockets without re-engineering the whole hall.

Discovery questions for the customer

Do you have a central chilled-water plant, or do you need self-contained DX units?

Why this matters: CRAH (chilled water) wins above ~1 MW. CRAC (DX) makes sense for small sites or edge. Determines which product line entirely.

What's the rack density per row and per hall?

Why this matters: Density drives per-CRAH/CRAC capacity and unit count. Also flags whether you'll need in-row CRAHs or RDHx for high-density pockets.

Down-flow under raised floor, in-row, or fan-wall architecture?

Why this matters: Down-flow is traditional but fading. In-row targets high-density pockets. Fan-wall is the hyperscale greenfield direction. Reveals architectural maturity.

Which DCIM are you running — Vertiv iCOM, Schneider EcoStruxure, or third party?

Why this matters: Integration scope matters. If they have an existing platform, your CRAH must speak that language. If they don't, propose one.

Are EC fans + AI airflow optimization in your scope, or budgeted as future upgrade?

Why this matters: EC fans cut fan energy 30–50% vs. belt-drive AC. AI control loops further optimize. Easy upgrade story if they have older units.

Hot & cold aisle containment Available

The single highest-ROI thing a data center operator can do for cooling. Physically separate the supply (cold) and return (hot) air streams so they don't mix. Without containment, ~30% of supply cold air bypasses the racks; with containment, almost 100% goes through equipment.

Cold-aisle (open) / hot-aisle (contained) configuration. Cold air rises through floor tiles, passes through the racks front-to-back, dumps into the contained hot aisle, and returns to the CRAH above the ceiling.

Rule of thumb

Containment alone typically cuts cooling energy 20–40%. Hot-aisle containment is generally preferred over cold-aisle because the rest of the room stays comfortable for techs.

Common solutions

Hard ceiling/roof panels, drop-down vinyl strip curtains, gap panels and blanking plates in racks, brush grommets at floor penetrations. Many hyperscalers use chimney cabinets (each rack vents directly into a ceiling plenum).

Selling angles

ROI in <18 months on retrofit. Bundle blanking panels, floor grommets, and door kits. Fire-code: ensure containment design is compatible with the room's suppression strategy (FM-200, Novec, sprinkler).

Discovery questions for the customer

Hot-aisle or cold-aisle containment preferred?

Why this matters: Hot-aisle is generally preferred (room stays comfortable for techs). Cold-aisle is older and still common in retrofits. Confirms their design philosophy.

What's your fire suppression — sprinkler, FM-200, Novec, or aspirating?

Why this matters: Containment design must integrate with suppression. Drop-down panels and sliding doors have different code paths depending on the agent.

Retrofit into an existing hall, or new build?

Why this matters: Retrofits typically pay back in <18 months on energy savings — easy ROI sale. New builds get containment as a default; the conversation is about which type and brand.

Are you considering chimney cabinets vs. room-level containment?

Why this matters: Chimney cabinets give each rack its own return path; more granular but more expensive. Common for hyperscale; rare for enterprise.

Have you done a leakage / bypass-airflow study at this site?

Why this matters: Most legacy halls leak 30%+ of cold air. A tracer-gas study or CFM-balance audit identifies the gaps and justifies containment + blanking-panel packages.

Chillers Tight

The big refrigeration machines that produce chilled water for CRAHs and CDUs. Two architectural choices dominate: air-cooled chillers (reject heat directly to ambient air via a packaged unit on the roof or pad) and water-cooled chillers (reject heat to a condenser-water loop, which then goes to a cooling tower).

The same four-step cycle is inside every air-conditioner on Earth, just scaled up. Compressor adds energy → condenser dumps it (to outdoor air or a cooling-tower loop) → expansion valve drops pressure → evaporator absorbs heat from the chilled-water loop. Efficiency (kW/ton) is mostly determined by the compressor stage; magnetic-bearing centrifugals win because they eliminate friction losses.

Type	Efficiency (kW/ton)	Use	Water?
Air-cooled scroll/screw	0.9–1.1	<500 ton, low-water sites	None
Water-cooled screw	0.55–0.65	200–800 ton	Cooling tower
Water-cooled centrifugal	0.45–0.55	500–4,000 ton	Cooling tower
Magnetic-bearing centrifugal (oil-free)	0.40–0.50 (part-load 0.30!)	500–2,500 ton	Cooling tower
Air-cooled with adiabatic precooling	0.65–0.80	Modern hyperscale	Minimal

Rule of thumb

1 ton refrigeration = 3.517 kW heat removal = 12,000 BTU/hr. For 1 MW of IT load you need roughly 285 tons of cooling (1,000 kW / 3.517). Add chiller capacity for redundancy (N+1 typical, 2N for Tier IV) and oversize ~10% for ambient extremes. Customers spec the chilled-water supply temp — historically 7 °C (45 °F), now often 18 °C (65 °F) or warmer to maximize free cooling.

Common solutions

Water-cooled centrifugal with magnetic bearings (Trane Series E CenTraVac, York YK/YZ, Carrier 19DV/AquaEdge, Daikin Magnitude). For modern campuses, air-cooled chillers with adiabatic pre-cooling (Trane Sintesis Advantage, Carrier AquaForce Vision) — uses far less water than a cooling tower while approaching its efficiency.

Supply chain reality

Large water-cooled chillers: 30–50 weeks. Air-cooled chillers: 20–40 weeks. Magnetic-bearing centrifugals (oil-free) are particularly tight because Danfoss Turbocor compressors are the single supplier for most OEMs — that's been a bottleneck. Steel/aluminum tariffs added ~10–15% to coil and shell costs in 2025. Order chillers at the same time as transformers.

Key selling angles

Part-load efficiency: data center chillers spend 90% of life at 30–70% load. Magnetic-bearing centrifugals dominate part-load. Lead the conversation with IPLV/NPLV, not full-load kW/ton.
Free-cooling chiller (integrated economizer): when ambient is cold enough, bypass the compressor and use the cooling tower or coil directly. 4,000–6,000 free-cooling hours/year in northern climates.
Refrigerant transition: R-134a being phased out in favor of low-GWP HFOs (R-1233zd, R-1234ze, R-513A). Position your platform on the right side of the transition.
Heat reuse: capture chiller condenser heat for district heating or industrial process — sells well in EU/Nordic markets.
Modular chiller plants: skid-mounted, factory-tested, single-trip delivery. Cuts site labor by 60%.

Discovery questions for the customer

Water-cooled or air-cooled chillers?

Why this matters: Water-cooled is most efficient at scale but needs a tower and water. Air-cooled with adiabatic precool wins in drought regions. Their answer drives the entire mechanical plant.

Are magnetic-bearing centrifugal (oil-free) chillers acceptable, given the single-supplier (Danfoss Turbocor) risk?

Why this matters: Oil-free dominates part-load efficiency, but lead time is constrained by Turbocor compressor supply. If they want oil-free, schedule conversation comes next.

Integrated free-cooling / waterside economizer in scope?

Why this matters: Doubles or triples free-cooling hours per year. Big PUE lever — every customer in a temperate climate should be considering it.

What refrigerant are you specifying — R-134a, R-1233zd, R-1234ze, R-513A?

Why this matters: R-134a is being phased out for high-GWP. Low-GWP HFOs are the modern direction. If they're still spec'ing 134a, raise the regulatory risk.

Modular / skid-mounted plant, or traditional stick-built?

Why this matters: Modular cuts site labor 60% and accelerates schedule. Worth the premium when they're behind on commissioning timelines.

Is heat reuse an option? Any potential district-heating off-taker or industrial process on site?

Why this matters: ERE<1.0 is achievable with heat reuse, big ESG win and possible tax incentive. Mostly relevant in EU/Nordic markets but worth asking everywhere.

Cooling towers & dry/adiabatic coolers Tight on large units

The atmospheric rejection device for the condenser-water loop. Open-circuit (counterflow or crossflow), closed-circuit (fluid coolers), or hybrid adiabatic. Water-use questions have made the cooling tower a politically sensitive piece of equipment in drought-prone regions.

An open cooling tower is brutally simple: spray hot water down through fill, draw outside air up through it, ~1% of the water evaporates and cools the other 99% — but consumes a lot of water (the WUE story). An adiabatic cooler keeps the process water inside a closed coil and only sprays the outside coil surface on the hottest days. Dramatically less water at slightly worse efficiency.

Rule of thumb

Tower sizing follows condenser load (chiller IT load + chiller compressor work). For a 1 MW IT facility figure roughly 1.2–1.4 MW thermal at the tower. Water consumption: ~1.5–2.0 L per kWh of IT load for an open cooling tower system. Dry/adiabatic systems trade some efficiency for near-zero water.

Common solutions

Open counterflow induced-draft (BAC Series 3000, SPX/Marley NC, EVAPCO ESW) for large WCC plants. Closed-circuit fluid coolers (BAC FXV, EVAPCO eco-Air) where water quality is bad or hygiene is a concern. Dry coolers for water-scarce regions. Hybrid adiabatic increasingly the spec for new hyperscale in arid climates.

Supply chain reality

Large factory-assembled towers: 20–35 weeks. Field-erected (concrete basin or wood/FRP structures): 12+ months including site work. Adiabatic coolers tight because demand has surged with WUE-conscious specs.

Selling angles

Water story is the lead in 2026 — Arizona, Texas, Nevada are increasingly hostile to high-WUE designs. Pair an adiabatic cooler with a small chiller for peak ambient days; spend most of the year on free cooling. Variable-speed EC fans, plume abatement, drift eliminators, legionella mitigation.

Discovery questions for the customer

What WUE target are you committed to (publicly or internally)?

Why this matters: Drives open-tower vs. closed-circuit vs. dry-cooler choice. Customers with corporate water pledges (hyperscalers) push hard for low WUE.

Open cooling tower, closed-circuit fluid cooler, or hybrid adiabatic?

Why this matters: Open is cheapest and most efficient but uses the most water. Closed-circuit eliminates legionella and water quality concerns. Hybrid adiabatic is the modern compromise in arid regions.

Is the water permit secured at the site? What's the cost of water?

Why this matters: Arizona, Texas, Nevada, parts of Spain are increasingly hostile to high-water designs. Lack of water permit can kill a site — flag it early.

Is plume abatement required by local code?

Why this matters: Required in cold climates and where towers are near roads/airports. Adds cost — better to know upfront.

Drift eliminator standard, and legionella mitigation program?

Why this matters: Legionella outbreaks at data centers have happened. Modern drift eliminators and chemical-treatment programs are non-negotiable but often overlooked in spec.

What's your design approach temperature?

Why this matters: Tighter approach (5–7 °F) = bigger tower but more free-cooling hours. Looser approach (10–15 °F) = smaller tower but more chiller hours. Driving the tradeoff at design time is your value-add.

Economizers & free cooling Available

"Free cooling" means moving heat to atmosphere without running a refrigeration cycle. Three architectures: airside economizer (direct outside air, with filtration), indirect airside economizer (outside air through a heat exchanger to a separate indoor loop), and waterside economizer (cool the chilled-water loop via the cooling tower or dry cooler when ambient is cold enough).

Three ways to get heat out of the data center without a compressor. Higher chilled-water supply temperatures dramatically widen the free-cooling window — the modern hyperscale move is to design the IT side for warmer supply (18–27 °C) and use the chiller only as backup for peak ambient days.

Rule of thumb

Free-cooling hours per year (waterside, depending on supply water temp):

7 °C / 45 °F chilled water: 2,000–3,000 hrs (Northern US/EU)
18 °C / 65 °F chilled water: 5,000–7,000 hrs
27 °C / 81 °F supply (warm-water cooling): 7,500–8,500 hrs — essentially year-round in most climates

Common solutions

Integrated free-cooling chillers (Trane Sintesis Advantage, Carrier AquaEdge with WSE). Plate-and-frame heat exchangers between condenser loop and chilled-water loop. Indirect adiabatic air handlers (Munters Oasis, Nortek StatePoint, Stulz CyberAir 3 PRO E2) are the modern hyperscale staple — outside air cools an internal air loop via a polymer/aluminum heat exchanger, evaporative spray on the outside surface when needed, no contaminant exposure inside the data hall.

Supply chain

Custom indirect-adiabatic AHUs: 25–40 weeks. Plate heat exchangers (Alfa Laval, SWEP): generally available.

Selling angles

PUE lever. Pair a high supply-water setpoint with indirect adiabatic and you get a sub-1.2 PUE in most U.S. climates with zero chiller hours half the year. Lead with the climate analysis and the kWh savings dollar figure for the customer's specific site.

Direct-to-chip liquid cooling (DLC) Booming, tight

The fastest-growing category in 2026. Coolant (water-glycol or dielectric fluid) is piped through a cold plate mounted directly on each CPU/GPU. The cold plate replaces the air heat sink. A CDU (coolant distribution unit) at the rack or row separates the IT loop (clean coolant) from the facility loop (chilled water) via a brazed-plate heat exchanger.

DLC architecture. Two isolated loops — clean IT loop and facility loop — couple at the CDU's plate heat exchanger. Warm supply temperatures mean the facility loop can usually reject to a dry cooler with no chiller, killing PUE.

Rule of thumb

DLC handles 70–80% of rack heat on the liquid side (the chips); the rest (memory, NICs, power supplies) still needs air cooling. So a "liquid-cooled" rack usually still has CRAH or rear-door units for the air side. Typical IT-loop supply temp: 30–45 °C (W3/W4 ASHRAE) — warm enough to free-cool with a dry cooler in any climate.

Common solutions

In-rack CDU (40–100 kW) for retrofit; in-row CDU (300 kW–2 MW) for new builds; sidecar CDU adjacent to OEM racks (NVIDIA GB200 NVL72 reference). Cold plates are typically OEM-integrated (Dell, HPE, Supermicro ship them on the server) — the CDU, manifolds, hoses, and dripless quick-disconnects are the open-market opportunity.

Supply chain reality

Hottest category in the market. CDUs and high-capacity cold plates are essentially supply-constrained — Schneider bought Motivair specifically to secure cold-plate and CDU manufacturing capacity for the AI build-out. CoolIT announced a 2 MW CDU offering in April 2025. Lead times: 20–40 weeks for high-capacity CDUs, longer for custom integrations.

Key selling angles

AI is the wedge: NVIDIA GB200, AMD MI300X, Intel Gaudi all spec DLC at the rack scale customers actually want.
Density unlock: air-cooled tops out at ~30–40 kW/rack; DLC reaches 130–200 kW/rack with room to grow.
PUE 1.05–1.10: warm-water supply means dry-cooler free cooling year-round in most climates.
Retrofittable: in-rack and sidecar CDUs let customers add DLC to existing data halls without redoing the chilled-water plant.
Service: coolant chemistry management, filter cartridges, leak detection — recurring revenue.
Risk story: dripless quick-disconnects, leak sensors at every joint, automatic pump shutdown — customers fear water near the electronics, so over-engineer the failure modes.

Discovery questions for the customer

What's your target rack power density — 40, 80, 130, or 200+ kW per rack?

Why this matters: 40 kW is air-cooling's ceiling. 80+ requires liquid. 130+ likely needs DLC + RDHx hybrid. 200+ approaches immersion territory. Density is the single most important number.

Which GPU / accelerator platform — NVIDIA GB200, AMD MI300X, custom?

Why this matters: NVIDIA's reference designs (especially GB200 NVL72) specify DLC. AMD MI300X is similar. Custom silicon may have unique cold-plate requirements. The chip platform usually dictates the cooling architecture.

What IT-side coolant supply temperature are you targeting — 30, 35, or 45 °C?

Why this matters: Warmer supply (W4/W5 class) unlocks year-round free cooling with a dry cooler. Cold supply (W1/W2) still needs a chiller. Drives the entire facility-loop architecture.

CDU placement — in-rack, in-row, or sidecar?

Why this matters: In-rack is best for retrofits and small deployments. In-row scales to whole rows. Sidecar pairs with vendor-shipped racks (NVIDIA NVL72 reference). Tells you the integration scope.

What's your CDU redundancy target — internal pump redundancy, or full CDU N+1 / 2N?

Why this matters: Internal pump redundancy doesn't cover CDU failures. Tier III/IV needs CDU-level redundancy with isolation valves and manifold rings. Surfacing this early avoids design rework.

What coolant chemistry and service model do you require?

Why this matters: Water-glycol mix, propylene glycol, or specialty dielectric? Coolant management is a recurring service. Filter cartridges, biocide, conductivity monitoring — all ongoing revenue.

What leak detection and dripless quick-disconnect requirements do you have?

Why this matters: Water near electronics is the #1 fear. Customers want sensors at every joint, automatic pump shutdown, and dripless disconnects. Always over-engineer the failure-mode story.

Immersion cooling Niche, growing

Submerge entire servers in a tank of dielectric fluid. Two variants: single-phase (synthetic hydrocarbon or ester oil, heat removed by a coolant loop on the tank — coolant stays liquid) and two-phase (engineered fluorocarbon boils at ~50 °C, vapor condenses on a coil in the lid — far higher heat transfer but the fluid is expensive and PFAS-regulated).

Single-phase: servers sit in oil; a pump circulates the oil through an external heat exchanger. Simple, scalable. Two-phase: fluorocarbon boils at the chip surface, vapor rises and condenses on a coil in the sealed lid, condensate drips back. Higher density possible but the fluid is $50–100/L and PFAS regulations have constrained supply.

Rule of thumb

Single-phase immersion handles 50–250 kW per tank, two-phase up to ~500 kW. Eliminates server fans (saves ~10% of IT power). Coolant cost is the killer: $5–10/L for single-phase, $50–100+/L for two-phase. A typical 50 kW tank holds 600–1,200 L.

Common solutions

Single-phase: GRC ICEraQ, Submer SmartPodX, Asperitas AIC, LiquidCool Solutions. Two-phase: LiquidStack DataTank (until 3M exited Novec production), ZutaCore (uses HFE refrigerant). The 3M 3M Novec exit in 2022 disrupted two-phase; alternatives are still maturing.

Supply chain reality

Tanks themselves are available. Coolant for two-phase is the bottleneck and regulatory uncertainty (PFAS) has slowed adoption. Many "immersion" wins in 2025–2026 are actually single-phase with synthetic ester oils.

Selling angles

Crypto and inference workloads have been the early customers. Bitcoin miners adopted it first; HPC and AI inference are the next wave. PUE 1.02–1.05 is real. Selling against DLC: simpler thermals (no leaks into the equipment because everything is already in the fluid), but harder operations (techs need PPE, fluid handling, servers must be removed and drained for service).

Discovery questions for the customer

Single-phase (synthetic oil) or two-phase (engineered fluorocarbon)?

Why this matters: Single-phase is most "immersion" wins in 2026 (simpler, cheaper, less regulated). Two-phase has higher density but PFAS regulations and 3M Novec exit have constrained it.

What workload type — crypto mining, AI inference, HPC, AI training?

Why this matters: Crypto miners were first adopters — stable, replaceable hardware. Inference is the next wave. Training is harder because hardware is more dynamic. Tells you whether the customer's use case is mature for immersion.

Have your hardware OEMs validated their servers for immersion?

Why this matters: Dell, HPE, Supermicro have immersion-validated SKUs but most enterprise gear isn't validated. Voids warranty if used. Critical pre-sale check.

Is your ops team prepared for fluid handling — PPE, drain procedures, fluid management?

Why this matters: Service is messy (servers come out dripping). Operations cost is real and often underestimated. Surface it before they're committed.

What's your AHJ and insurance posture on immersion at this site?

Why this matters: Local code officials and insurers are still catching up on immersion. Successful deployments need pre-approval — flag the timeline impact.

Rear-door heat exchangers (RDHx) Available

A chilled-water coil mounted on the back door of an IT rack. Server fans push hot exhaust air directly through the coil; air leaves the rack at near-room temperature. Effectively turns each rack into its own contained cooling zone.

The "rear door" of the rack is the cooling coil. Hot exhaust from the servers passes through it before leaving the cabinet, so the back of the rack runs at almost the same temperature as the cold aisle. No raised floor, no CRAH, no aisle containment needed — each rack handles its own heat. Active variants add booster fans for racks up to 80 kW; passive variants top out at ~25 kW.

Rule of thumb

Active RDHx (with fan-assist) cools 30–80 kW/rack; passive (relies on server fans alone) tops out at ~25 kW. Common option for retrofitting high-density pockets — handful of AI racks in an air-cooled hall — without redesigning the whole cooling plant.

Common solutions

Motivair ChilledDoor (now Schneider), Vertiv Liebert XDR/RDHx, USystems ColdLogik, Coolcentric, Stulz RDHx, ZutaCore (with HFE refrigerant).

Supply chain

Generally available — 12–20 weeks. Demand has spiked with AI retrofits, particularly the 60+ kW active variants.

Selling angles

The retrofit weapon. Customer has 200 racks of legacy 8 kW gear and 20 new 40 kW AI racks — drop in RDHx on the AI racks, don't touch the rest. Warm-water RDHx (using 27 °C+ supply) can free-cool most of the year. No additional floor space needed.

Cooling section test (archived)

1How much cooling capacity in tons is needed for a 5 MW IT load?

Answer: ~1,425 tons. 5,000 kW / 3.517 kW/ton = 1,422 tons. Add 10% for redundancy/ambient = ~1,560 tons. In practice you'd spec two 800-ton chillers (1,600 tons) in N+1, or three 800-ton in 2N depending on tier.

2Why is PUE 1.5 considered mediocre in 2026 when it was best-in-class in 2010?

Answer: ASHRAE expanded the recommended/allowable envelopes (operators run warmer), containment became universal, EC fans replaced AC, magnetic-bearing chillers replaced reciprocating, free-cooling architectures matured, and warm-water DLC enabled chiller-free operation. The whole industry baseline shifted to 1.2–1.3, with hyperscale liquid-cooled facilities reaching 1.05–1.10. PUE 1.5 today says "the operator hasn't modernized" which is often a buying signal.

3A customer wants to know whether to use air-cooled chillers or water-cooled chillers with cooling towers. What's your decision tree?

Answer:

Water availability/cost? Drought region or expensive water → air-cooled with adiabatic precool.
Local water regulations or community optics (WUE goals) → air-cooled.
Cold climate with lots of free-cooling hours? Air-cooled with integrated economizer wins on lifecycle.
Hot, water-available, biggest scale (>10 MW)? Water-cooled centrifugals plus open towers will be most efficient and cheapest per ton.
Modern compromise: hybrid adiabatic coolers — air-cooled most of the year, spray water only on the few peak-ambient days.

4What's the difference between IPLV and full-load kW/ton, and which one matters for data centers?

Answer: Full-load kW/ton is efficiency at 100% capacity at standard rating conditions. IPLV (Integrated Part Load Value, AHRI 550/590) is a weighted average across 100%/75%/50%/25% capacity. Data centers spend almost all their life at part-load (40–70%), so IPLV is the number that matches reality. Magnetic-bearing centrifugals shine on IPLV (oil-free, infinite turndown), even if their full-load kW/ton is only marginally better.

5Why is "warm water cooling" (supply at 27 °C+) such a game-changer?

Answer: At 27 °C supply temperature, the chilled-water loop is warmer than ambient air for most hours of the year in almost any climate. That means the loop can be cooled by a dry cooler (just air-to-water fin coil) with no compressor running. You essentially delete the chiller from the equation. Combined with hot-aisle return at 40 °C+, you maintain workable ΔT while running zero chiller-hours from October to May in temperate climates. PUE drops to 1.1 and water use to near-zero.

6What's an "approach temperature" and why does the customer care?

Answer: The temperature difference between the leaving fluid and the entering air (or water). Tighter approach = bigger / more expensive heat exchanger but more usable free-cooling hours. A 5 °F approach is aggressive (expensive coil); 10 °F is common; 15 °F is loose. Tighter approach extends the operating window where the chiller can be off — that's the customer's lever for free-cooling hours.

7An AI customer says they need 130 kW/rack and asks what cooling architecture you'd recommend.

Answer: Direct-to-chip liquid cooling for ~75% of the heat (the GPUs), with rear-door heat exchangers or in-row CRAH for the remaining ~25% air-side load (memory, NICs, PSUs, switches). Facility loop at 30–35 °C supply allows year-round dry-cooler free cooling in most climates. Two-loop architecture: clean IT-side coolant isolated from facility water by a CDU. Specify dripless quick-disconnects, leak detection, redundant pumps. If the customer is greenfield and densities will keep rising, consider whether two-phase or single-phase immersion makes sense for a subset of the deployment.

8The CDU has redundant pumps — why is that not enough redundancy?

Answer: Internal pump redundancy protects against a pump failure but not a CDU failure (HX leak, control failure, valve failure). For Tier III/IV liquid cooling you want CDU-level N+1 or 2N — multiple CDUs feeding a manifold, with isolation valves that let one CDU be taken offline. Hyperscalers run "shared CDU" or "ring manifold" designs so a single CDU outage doesn't drop a row.

9Customer asks why you'd ever use immersion cooling instead of direct-to-chip.

Answer:

Cools 100% of the components, not just the hot chips — no air-side cooling needed.
No server fans (saves ~10% of IT energy and acoustically silent).
Higher density possible — 100+ kW/U in two-phase configurations.
Dust, humidity, and corrosion become non-issues — useful for harsh environments.

Trade-offs: hardware OEMs must validate immersion compatibility; service is messy (servers come out dripping); two-phase fluids are PFAS-regulated and expensive; insurance and AHJ understanding is still developing. So immersion wins where density is extreme, the workload is stable (low maintenance), and operations are simple — crypto, fixed-function inference, certain HPC.

10What's "WUE" and why does it matter more than it used to?

Answer: Water Usage Effectiveness — liters of water consumed per kWh of IT load. Open cooling towers typically run 1.5–2.0 L/kWh; dry coolers run 0. Customers and regulators in water-stressed regions (Arizona, Nevada, parts of Texas, Spain, Chile) now write WUE targets into permits and corporate ESG reports. AWS, Microsoft, and Google have public water-positive pledges. A cooling pitch that ignores WUE in those markets dies on the table.

11The customer's existing chilled-water plant runs at 7 °C supply. They want to add AI racks. What do you suggest?

Answer: Decouple the AI loop from the legacy 7 °C plant via a CDU. Run the AI-side facility loop at 30–35 °C supply, sourced from a new dry cooler or air-cooled chiller. Trying to feed 7 °C water to DLC racks wastes the chilled-water capex (you don't need it that cold) and prevents free cooling on the AI side. Two-tier plants like this are common in retrofits.

12What does "fan wall" mean and where does it fit?

Answer: A wall of EC fans pulling air through a large cooling coil (or air-to-air heat exchanger) at the perimeter of a data hall. Replaces dozens of individual CRAH units with a single large cooling-array volume. Hyperscale staple — better efficiency (fewer fans running at higher efficiency point), simpler maintenance, eliminates raised floors entirely. Vertiv, Schneider, Stulz, Munters all offer fan-wall solutions.

13Customer wants to recover the heat from their cooling system. What's realistic?

Answer: Recovering 30–40 °C return water from CRAHs is hard — too low-grade for most uses. DLC return water at 50–60 °C is genuinely useful — it can feed a district-heating loop directly, or a heat pump can boost it to 80 °C+ for industrial process or domestic hot water. Stockholm, Helsinki, and Dublin have working examples. ERE (Energy Reuse Effectiveness) can drop below 1.0 when reuse is substantial. The honest answer for U.S. greenfield is that without a co-located heat off-taker the economics rarely pencil — but for European and Nordic customers, mandatory.

14How do you size a CRAH for a 100 kW row?

Answer: Required airflow at 12 °F ΔT ≈ 285 CFM/kW × 100 kW = 28,500 CFM. Most 30-ton CRAHs deliver ~12,000–15,000 CFM — so you'd want 2 units to cover 100 kW with margin, plus an N+1 spare = 3 units total. Tonnage-wise: 100 kW / 3.517 = 28.4 tons, so 2 × 30-ton = 60 tons gives ~2× capacity, which sounds wasteful but matches CFM-needed and provides cooling redundancy. CFM rules airflow design; tons rules thermal capacity.

15Customer asks: "Can I run my data center at 30 °C cold aisle to save energy?"

Answer: Probably yes — ASHRAE A1/A2 allowable goes to 32 °C and almost all modern servers tolerate it. But: (1) Tech workforce comfort declines, (2) thermal margin during a cooling failure shrinks, (3) some legacy enterprise gear isn't rated above 27 °C, (4) higher inlet means hotter return which can push CRAHs/chillers off their design curve. Modern practice: run at the warm end of recommended (~27 °C) with allowable headroom for emergencies. Hyperscalers do run hotter; enterprises rarely should.

16Why did Schneider buy Motivair, and what does it tell you about the market?

Answer: Cold plates, CDUs, and high-capacity rear-door heat exchangers are the supply chain choke point for AI build-outs. Schneider acquired Motivair (2024) specifically to bring cold-plate and CDU manufacturing in-house and lock in capacity. Implication: the big four (Schneider, Vertiv, Eaton, ABB) are all racing to own the liquid-cooling stack, not just sell adjacent equipment. Expect more acquisitions and exclusive supply deals with NVIDIA/AMD/hyperscalers.

17What's "containment leakage" and how do you measure it?

Answer: The fraction of cold supply air that bypasses the IT equipment — leaks around blanking panels, under floor tiles, through unsealed cable cutouts, or through gaps in containment doors. Measured by tracer-gas studies or by comparing supply CFM to IT-equipment CFM demand. 30%+ in a poorly sealed legacy hall, <5% in a well-tuned modern one. Every percentage point of leakage reduction translates directly to fan energy savings.

18Apply it: 25 MW AI training facility, mixed deployment — 60% liquid-cooled NVIDIA racks, 40% air-cooled support infrastructure (storage, networking, mgmt). PUE target 1.15. Design the cooling system at a sketch level.

Answer:

Liquid loop (~15 MW IT, ~12 MW to liquid): Facility water at 32 °C supply / 48 °C return. CDUs at end-of-row, 1 MW capacity each, sized N+1 per row (~16 CDUs total). Heat rejection via dry coolers (Güntner or BAC closed-circuit fluid cooler) — sized for design ambient. Adiabatic spray section for summer days >35 °C. No mechanical chillers in the loop.
Air loop (~10 MW IT, ~13 MW with the 3 MW from the liquid side that still ends up as air): Indirect-adiabatic AHUs (Munters Oasis or Nortek StatePoint), polymer heat exchanger, with optional integral DX trim. Hot/cold aisle contained, supply temp 24 °C.
Backup/peaking chillers: 2× 1,500-ton air-cooled magnetic-bearing chillers, only run when adiabatic systems can't keep up.
Expected PUE: liquid side 1.05, air side 1.20, blended ≈ 1.11.
Risks: CDU lead time (Schneider/Motivair, CoolIT, Vertiv allocation); custom AHU lead 30–40 weeks; dry-cooler capacity in summer ambient (worth a TMY simulation, not a rule of thumb).

Standards, tiers & the customer's compliance world

Uptime Institute Tier classification

Tier	Definition	Power topology	Cooling	Availability target
Tier I	Basic, no redundancy	Single path, N	N	99.671% (~28h downtime/yr)
Tier II	Redundant components	Single path, N+1	N+1	99.741% (~22h/yr)
Tier III	Concurrently maintainable	Multiple paths, only one active, N+1 each	N+1, concurrently maintainable	99.982% (~1.6h/yr)
Tier IV	Fault tolerant	2 active independent paths, 2(N+1)	2(N+1), 2 independent paths	99.995% (~26 min/yr)

Most enterprise and colo is Tier III. Tier IV is reserved for financial, health, government. Hyperscalers often build to internal standards better than Tier IV (multi-site replication makes per-site uptime less critical).

Other standards worth knowing by name

Standard	What it covers	Where it matters
ANSI/TIA-942	Telecom infrastructure for DCs (cabling, spaces)	Network & cabling sales
ASHRAE TC 9.9	Thermal envelopes (A1–A4, H1, W1–W5)	Every cooling conversation
NFPA 70 (NEC)	U.S. electrical code	All power equipment design
NFPA 75 / 76	Fire protection for IT equipment / telecom facilities	Containment, suppression
NFPA 855	Stationary energy storage systems (Li-ion battery rooms)	UPS battery design
IEC 60364	Electrical installation (international)	EU/global projects
EU Code of Conduct for DCs	Voluntary efficiency framework	Hyperscale EU operations
EN 50600	European DC facility standard	EU customers, especially colo
EPA ENERGY STAR for DCs	PUE benchmark	Enterprise reporting
SOC 2 / ISO 27001	Operational controls (not equipment, but customers' compliance)	Audit-related design choices

Supply chain heatmap (mid-2026)

Equipment	Lead time	Current supply chain	SPM stock
Large power transformers (>30 MVA)	12–24 months	Severe	Available
Medium-voltage switchgear	6–12 months	Severe	Available
Large diesel generators (≥2.5 MW)	6–12 months	Severe	Available
Natural gas generators / turbines	8–15 months	Severe	Available
UPS (large, >500 kW)	3–6 months	Tight	Available
Li-ion batteries	8–18 weeks	Tight	Available
LV switchgear / switchboards	3–6 months	Tight	Available
ATS	8–16 weeks	Tight	Available
Floor PDUs / RPPs	8–10 weeks	Tight	Available
Busway	10–16 weeks	Tight	Available
Large air-cooled chillers	10–20 weeks	Tight	Available
Magnetic-bearing centrifugal chillers	4–6 months	Tight	Available
CDUs (large, >500 kW)	10–20 weeks	Tight	Available
Cold plates (high-capacity)	10–20 weeks	Tight	Available
CRAH / CRAC standard	8–10 weeks	Available	Available
Rear-door heat exchangers	8–10 weeks	Available	Available
Cooling towers (factory-assembled)	10–18 weeks	Tight	Available
Containment hardware	8 weeks	Available	Available
Rack PDUs (intelligent)	8 weeks	Available	Available

SPM stock shows current availability through our channel — all categories above are in stock or on confirmed allocation. The Current supply chain column reflects the wider industry condition for context.