ATS vs STS in Data Centers: What Actually Changes the Design

Ask a room of electrical engineers what the difference is between an ATS and an STS and you will hear "speed" — automatic transfer switches take seconds, static transfer switches take milliseconds. That is true and almost useless. The two devices are not faster and slower versions of the same job. They answer two different questions. An automatic transfer switch (ATS) answers "the source is gone, where do I go?" A static transfer switch (STS) answers "both sources are alive, but this path is no longer acceptable — how do I leave it without the load noticing?"

Put an STS where an ATS belongs and it adds nothing but cost, because one of its two required sources will not be energized. Put an ATS where an STS belongs and the load drops for seconds, because the mechanical contacts cannot close inside the ride-through window of a server power supply. Everything downstream — the topology, the standards, the maintenance calendar, and a surprising amount of continuous energy loss — follows from getting that placement right.

This guide is about the placement, not the brochures. It covers where each device sits in a compute-center power chain, what the millisecond numbers actually mean against the ITIC ride-through envelope, how the NFPA 110 ten-second window gets consumed item by item, why an STS burns kilowatts continuously while an ATS burns almost none, what changes when the load is a 120 kW AI rack, and the specification checklist that separates a compliant installation from one that fails on the drawings.

The one-line version, then the real version

Both devices are transfer equipment: they select which of two sources feeds a common load, and they are rated in amperes [16]. Beyond that they diverge in every way that matters.

Table 1 — ATS and STS side by side. The only row that decides the application is "sources it expects": if one of the two sources is not energized when you need it, an STS cannot help you.

Where each device sits in the power chain

A compute center power chain reads, in simplified form: utility → ATS or transfer switchgear → UPS → distribution (PDU, RPP, busway) → rack PDU → IT equipment [8]. The ATS occupies the upstream slot. The STS occupies a slot downstream of the UPS, in front of a load that has only one power input.

The asymmetry is structural. An ATS is a source-selection device for the facility: it decides whether the building runs on the grid or on its generators [7]. An STS is a source-selection device for a load: it decides whether a specific rack, panel, or device draws from bus A or bus B [6]. A large facility can contain one ATS lineup and dozens of STS units, and both can be doing useful work in the same outage.

That is also why the two devices appear in the same Tier III and Tier IV designs as a matter of routine: the ATS handles the utility-to-generator handoff at facility level, and the STS gives single-corded equipment access to the redundant UPS buses that feed dual-corded equipment natively [1][3]. They are not competitors in that topology; they occupy different layers of the same defense.

The ATS: built for a dead source

An ATS is an electromechanical device with a controller. Its seven jobs are: carry current continuously, detect loss of the normal source, signal the alternate source to start and initiate transfer, transfer the connected load, sense restoration of the normal source, retransfer the load back, and withstand and close on fault current [16]. Everything about its design follows from the fact that its alternate source — usually a diesel generator — is not running when the outage begins.

Three transition modes, and why the middle one is dangerous to misuse

• Open transition (break-before-make). The load is briefly disconnected while the switch moves. This is the cheapest and most common arrangement, and the interruption lasts from about 0.5 s up to 30 s depending on how long the generator takes to start and stabilize [2]. It is acceptable for one reason only: the UPS downstream rides through the gap.
• Delayed transition. The switch pauses in a center-off position so that motor and transformer fields decay before reconnection. On a pure IT load this is usually unnecessary; on a hall full of chilled-water pumps and CRAH units it prevents out-of-phase reclosing, which can damage motor windings [2][5].
• Closed transition (make-before-make). Both sources are paralleled for under 100 ms, producing no perceived interruption. The price is synchronization: sources must agree within roughly ±5° of phase angle, ±0.2 Hz of frequency, and about 5% voltage before the controller will parallel them — and the local utility normally has to approve the brief parallel operation with its grid [2].

What "seconds" means against a server power supply

The tolerance envelope that governs this industry is the ITIC (formerly CBEMA) curve, which describes the AC input voltage range that information technology equipment can tolerate, including a complete voltage dropout that must resolve within about 20 ms [2][5]. Twenty milliseconds is one and a fifth cycles at 60 Hz. Hold that number next to the ATS outage window from the table above:

An ATS open transition exceeds the ITIC dropout window by one to three orders of magnitude. That is not a defect; it is the definition of the application. The ATS is not meant to protect IT equipment directly — it is meant to protect the UPS, and the UPS is meant to protect the IT equipment. The battery carries the critical load across the entire generator start and transfer sequence [7].

The NFPA 110 ten-second window, spent item by item

Where life safety or a legally required load is involved, NFPA 110 sets the clock. A Type 10 system — the classification used for hospital essential loads and commonly adopted as the data center benchmark — must restore power to the load within 10 seconds of the loss of the normal source [19]. Critically, that window is measured from loss of normal power to energization of the load terminals on the emergency side of the transfer switch, so every intermediate delay counts: voltage and frequency sensing, the time-delay-on-transfer (TDOT) setting, generator cranking, generator voltage and frequency stabilization, and the mechanical operation of the switch itself [13].

Table 2 — A workable Type 10 allocation. There is exactly one second of slack in the sequence, and it is not the place to be generous with timers.

Now run the failure case that shows up in real commissioning reports: a 5-second TDOT (set generously to avoid nuisance transfers on utility dips) plus a 7-second generator start-to-stable. That is 12.2 seconds before the mechanical contacts have even finished moving — 2.2 seconds past the limit, with no load energization at all [13]. The switch is fine. The generator is fine. The settings are not.

Bypass isolation, and why Tier III asks for it

An ATS is a single point of failure sitting in the critical path, and mechanical contacts wear. A bypass-isolation ATS adds a second switching path so a technician can isolate and service the main switch while the load stays energized — the physical expression of the concurrent maintainability that Tier III requires, and a practical requirement under NFPA 110 for mission-critical sites where the transfer switch must be exercised and maintained without an outage [2][5]. Without it, maintaining the switch means dropping the load, which is the same as saying the switch never gets maintained.

Neutral poles: the 4-pole question

Where the standby source is a separately derived system — a generator with its own neutral-to-ground bond — the transfer switch has to switch the neutral as well as the phases. A 4-pole ATS with a switched neutral prevents the two neutral systems from being connected together, which would otherwise create parallel ground paths and nuisance ground-fault trips [2]. It is one of the least glamorous lines on a specification sheet and one of the most reliable ways to fail an inspection.

Codes and the ATS

Transfer equipment is listed to UL 1008, whose scope covers automatic and non-automatic transfer switches rated up to 6,000 A and 600 V, including their control devices such as voltage and frequency sensing relays and time-delay relays [9][16]. The application article follows the load: Article 700 for emergency systems, 701 for legally required standby, 702 for optional standby, and 708 for critical operations power systems [3]. Sizing follows the load class — emergency systems are commonly specified at 125% of the continuous load, and the switch also has to withstand and close on the available fault current, not merely carry it [1].

Related reading: the battery that carries the load across this whole window is sized by a different calculation — see how to size UPS battery backup for an AI data center, and the redundancy tiers that determine whether you need one transfer switch or two independent paths are compared in the US data center Tier standards comparison.

The STS: built for a source that is alive but unusable

A static transfer switch is, functionally, an AC/AC solid-state converter: a single-pole double-throw switch implemented with a pair of silicon-controlled rectifiers (SCRs), so that the output is fed from one of two input lines with no moving parts [6]. That construction is the whole story. Because there are no contacts to travel, the transfer is limited by detection and gate firing rather than by mechanics, and the device can switch between two live sources inside a fraction of a cycle [6].

Where it goes, and the one condition it cannot escape

The canonical placement is between the outputs of two independent UPS systems, feeding a load that has only one power cord [6][7]. Dual-corded servers do not need this: they already draw from an A and a B feed, and each cord survives the loss of the other path. But an enormous amount of equipment in a modern hall is single-corded — network switches, older storage arrays, out-of-band management, PDU monitoring, and the pumps and controllers inside coolant distribution units. The STS is the standard way to give those loads the benefit of two UPS systems that they did not physically ask for [5].

Because it switches between two live sources, the STS inherits a hard precondition: both sources must be present, and they must be within the synchronization window the controller accepts. If the two sources drift too far apart in phase, a well-designed STS delays or blocks the transfer rather than closing onto a source that would drive a large inrush into a downstream PDU transformer and trip an upstream breaker [5]. This blocking behavior is a feature, and it is also the reason an STS is not the right answer for a utility-to-generator transfer: the generator is not alive at the start of the event.

Two switching algorithms, and the 4 ms versus 8 ms question

Table 3 — The 4 ms and 8 ms figures on STS datasheets are not a quality ranking. They describe two different algorithms chosen by what is downstream of the switch.

Vendors publish the sub-cycle figure as "under 4 ms under normal operating conditions," with graceful degradation when the two sources are grossly out of phase — one major STS line specifies detection and transfer performance even at a 180° phase difference that is up to ten times faster than many competing switches, precisely because the worst case is where the design either holds or fails [20]. For sizing purposes, what matters is that even the degraded figure stays inside the 20 ms ITIC window, while the ATS figure does not come close [2][5].

The hidden variable: an STS burns power continuously

This is the part of the ATS-versus-STS comparison that most design discussions skip, and it is the one that keeps paying out year after year. A mechanical contactor has milliohms of contact resistance; a conducting SCR has a forward voltage drop that is essentially fixed regardless of the current through it — 1.3 V typical in one commercially documented STS power-switch module, with the industry spread running from about 1.0 V to 2.0 V [4][10]. A fixed voltage drop multiplied by a large current is a resistive loss that never turns off.

For a three-phase STS, one SCR pair conducts per phase, so the standing loss is three times the drop times the line current:

Put those losses next to the useful power the switch is carrying at 480 V three-phase, and the efficiency looks excellent — because it is. Then put them next to an annual electricity bill, and the picture changes:

Table 4 — Standing SCR loss on a three-phase STS at a 1.3 V drop. Efficiency is computed at unity power factor; at 208 V the same current carries far less useful power and the percentage efficiency falls toward 98.9% because the loss is set by current, not voltage.

Three consequences follow, and they all belong in a specification:

• The loss is current-driven, not voltage-driven. A 400 A STS at 208 V carries 144.1 kW and still loses 1,560 W — an efficiency of 98.93% rather than 99.53%. Low-voltage, high-current STS applications pay proportionally more for the same transfer capability.
• The heat has to go somewhere. 6.24 kW of continuous dissipation at 1,600 A is a space heater inside a cabinet. That is why an STS enclosure runs two to three times the size of an equivalent ATS enclosure and needs forced ventilation and thermal monitoring; in the documented module, fans start at a 45 °C heatsink and the SCRs go into thermal shutdown near 70 °C [1][10].
• Two paths means two losses. A mechanically redundant hall with an STS on each of the A and B paths is dissipating the SCR loss twice, continuously, whether or not a transfer ever happens.

The comparison is not that the STS is inefficient — 99.5% is an excellent number and the solid-state path is what buys the 4 ms transfer. The comparison is that the ATS's zero-loss characteristic is a real operating-cost advantage, and it should be entered on the right side of the ledger when someone argues that the cheaper switch is the more expensive one.

Current path, withstand, and the fuse question

Because an STS has no mechanical contacts to separate a fault, its current path design is a safety-critical choice rather than a detail. Some designs are deliberately fuseless, using disc-type SCRs oversized by 200–600% relative to the rating, so that the same semiconductor that conducts the load current also survives the fault current until an upstream device clears it — one such unit is certified by UL to safely sustain fault currents as high as 100 kAIC without fuses or thermal interruption devices in the path [20].

That matters because most installations sit on a bus with available fault current well above 45 kAIC, and a switch whose withstand rating is below the available fault current is a switch that fails destructively when it is needed most. The rule is the same one that applies to the ATS: a 400 A circuit on a bus with 35 kA available fault current needs a switch rated for at least that 35 kA of withstand and close-on capability, not merely a 400 A continuous rating [1].

Overload tolerance and derating

STS products publish overload envelopes that read like UPS specifications rather than switchgear: 110% of rated current continuously, 125% for 30 minutes, 150% for 2 minutes, and an inrush capability on the order of 500% for a fraction of a cycle to accommodate load startup surges [17]. Continuous ratings span a wide range, from rack-mount single-phase units of a few kVA up to three-phase cabinet units in the thousands of amperes [14][15][17]. What follows from that is a sizing habit borrowed from UPS practice: the STS should be selected for the load's transient behavior and future growth, not just its steady-state amperes.

Poles and neutrals, again

The STS inherits the same neutral problem as the ATS, with a sharper edge. A three-pole STS assumes the two sources share a neutral or that no neutral switching is required. Where two completely independent four-wire sources feed the switch, a four-pole device is required, and specifically a four-pole switched neutral — because a four-pole non-switched design leaves both input neutrals connected together at all times, tying two separately derived systems together and recreating exactly the parallel neutral path the fourth pole exists to prevent [6]. The specification line reads something like: "630 A, 400 V, 4-switched poles" [6].

Codes and the STS

Static transfer switches are listed to UL 1008S, the Standard for Solid-State Transfer Switches, whose scope covers solid-state automatic transfer switches rated up to 6,000 A and 600 V for use in ordinary locations, providing lighting and power in optional standby systems per Article 702 — and which notes that a solid-state transfer switch is not to be used as service entrance equipment unless it is marked as such [9]. That scope note is worth reading twice: it tells you where the standard expects the device to live, and it is downstream of a UPS, not at the utility entrance.

Head to head: the comparison that actually decides a design

Table 5 — Read the second and third rows together. An STS is only the right answer when the load cannot ride through an outage and both candidate sources are already hot.

What changes when the load is an AI rack

Two trends push this comparison in opposite directions. Power density is going up, and it is going up most sharply in exactly the single-corded equipment classification that benefits from an STS.

An NVIDIA GB200 NVL72 rack is designed around 120 kW per rack [5]. Run that against a feeder ampacity and the practical consequences of switch placement become concrete:

Table 6 — Rack count per STS feeder at AI densities. A 2 MW AI hall needs 6.3 parallel 400 A STS units, 3.2 at 800 A, or 1.6 at 1,600 A — which is why the higher ratings exist and why the SCR loss column in Table 4 grows with them.

Three design consequences specific to AI halls:

• The STS current rating becomes a capacity-planning number, not a detail. The mismatch between 120 kW per rack and a 400 A feeder means more parallel switching devices and more parallel loss. Choosing 1,600 A over 6× 400 A saves device count but concentrates 6.24 kW of heat into one cabinet.
• Single-corded equipment gets more critical, not less. Network fabric, management planes, and liquid-cooling controls are frequently single-corded while the compute they serve is dual-corded. In a hall where a single lost rack means 120 kW of GPU capacity offline, the STS protecting the switch that feeds that rack is not a secondary concern.
• Liquid cooling moves the risk into new places. CDUs have pumps, controllers, and sensors that are typically single-fed, and a CDU trip takes a thermal-critical hall down faster than a power interruption would. Whether those controls sit behind an STS is a design decision that has to be made at the single-line-diagram stage, not after the racks arrive — the same discipline described in CDU maintenance and service planning.

If the hall is being built as a power-first program rather than a retrofit, the upstream architecture decisions that constrain transfer-switch count and rating are covered in reinventing power for AI data centers. And because every one of these devices has to be proven under load before the hall goes live, the sequence described in step-by-step load bank testing applies to transfer equipment just as much as to generators — a transfer switch that has never been exercised under real current is an untested assumption.

The decision checklist

Choose an ATS when all of the following are true:

• The alternate source is a generator, or any source that will be dead at the moment of failure [7].
• Something downstream — a UPS — will ride through the seconds-long outage without help [2].
• The load class is emergency, legally required standby, optional standby, or critical operations, and the applicable NEC article has been identified [3].
• Concurrent maintainability is required, which means specifying bypass isolation rather than a plain ATS [2][5].
• Neutral switching is required because the standby source is separately derived — specify 4-pole switched neutral [2].

Choose an STS when all of the following are true:

• Two live, synchronized sources exist — typically two UPS outputs [6][7].
• The load is single-corded and cannot tolerate a transfer longer than the ITIC window [5].
• The available fault current has been compared against the switch's withstand rating, and the switch wins [1].
• The neutral configuration has been resolved: shared neutral for 3-pole, independent sources for 4-pole switched neutral [6].
• The continuous SCR loss and its cooling load have been added to the hall's power and thermal budget (Table 4).

Eight mistakes that show up in real projects

1. Sizing on the mechanical transfer time. The load's outage is measured from loss of the normal source to load energization, which includes sensing delay, TDOT, generator cranking, and stabilization — not just contact travel [13].
2. Setting TDOT generously "to avoid nuisance transfers." A 5-second TDOT plus a 7-second generator start-to-stable is 12.2 seconds against a 10-second requirement [13].
3. Forgetting that the STS needs two hot sources. An STS fed by utility and generator brings no stored energy to a configuration where one source is not yet alive [1][6].
4. Ignoring the synchronization window. An STS that closes onto a badly out-of-phase source can drive transformer inrush and trip the upstream breaker; the blocking behavior is the protection working [5].
5. Specifying a withstand rating below the available fault current. Continuous amperes and interrupting amperes are different numbers, and the second one is what protects the install [1].
6. Choosing 3-pole on independent four-wire sources. A non-switched 4-pole STS keeps both input neutrals connected, tying separately derived systems together [6].
7. Leaving the SCR loss out of the thermal budget. 6.24 kW at 1,600 A is continuous, and it lands inside the same cooling envelope as the IT load [10].
8. Deploying a transfer switch that has never been exercised under real load. Every standard cited in this article assumes the equipment has been tested, not merely installed [11][12].

Ten questions for the supplier

1. Which standard is the device listed to — UL 1008 or UL 1008S — and can I see the listing for the exact model and rating? [9][16]
2. What is the measured transfer time at my load, and what is the worst-case figure when the two sources are out of phase? [2]
3. For the ATS: what is the minimum TDOT you will support, and what is the sum of TDOT plus my generator's start-to-stable time? [13]
4. For the STS: what is the accepted synchronization window in degrees, hertz, and volts, and what does the controller do outside it? [5]
5. What is the withstand and close-on rating, and how does it compare to the available fault current at my bus? [1]
6. What is the SCR forward voltage drop at rated current, and what is the resulting heat dissipation in watts? [4][10]
7. What cooling does the enclosure require — airflow, ambient limit, and thermal shutdown setpoint? [1][10]
8. Can the switch be serviced without dropping the load, and does it include mechanical and key-interlocked bypass isolation? [2][5]
9. How many poles, and is the neutral switched or non-switched? [6]
10. What is the documented overload envelope at 110%, 125%, and 150% of rated current, and the fault-withstand duration? [17]

The bottom line

ATS and STS are not two speeds of the same device. An ATS is an electromechanical source selector for a facility whose alternate source has to be woken up, and its job is to buy the UPS time. An STS is a solid-state source selector for a load whose two sources are both awake, and its job is to make the transfer invisible. The ATS costs less up front and almost nothing to run, and needs bypass isolation if you ever intend to maintain it. The STS costs two to three times as much and a few kilowatts of continuous SCR loss, and buys you a quarter-cycle transfer and a service life measured in decades rather than operations.

Get the placement right first — generator behind the UPS on the ATS, single-corded critical loads between the UPS outputs on the STS — and the rest of the specification writes itself: standard for the device, withstand for the fault current, poles for the neutral, thermal budget for the loss, and a load test before anyone signs off. Get it wrong, and you will spend the difference twice: once on hardware that cannot do the job it was bought for, and once on the outage it did not prevent.

Specifying power distribution for an AI hall, a colocation build, or a containerized pod? Request a quote and we will work through the transfer-switch schedule against your topology, fault current, and redundancy target.

References

1. ATS vs STS Switch — 7 Differences That Actually Matter, SENTOP / OnesTo EP (transfer time ranges, cost and lifespan comparison, UL 1008 vs UL 1008S, NEC Article 700 sizing at 125%, withstand vs available fault current, enclosure footprint).
2. Transfer Switch Requirements for Data Centers & Mission-Critical Facilities, ValuAdd (ITIC/CBEMA 20 ms dropout envelope, open/delayed/closed transition modes and their synchronization limits, Tier III bypass isolation and Tier IV 2N implications, POG and VSS algorithms, 4-pole switched neutral).
3. Electrical Solutions for Data Centers: Power Switching & Transfer, ValuAdd (medium-voltage vs low-voltage placement, NEC Articles 700/701/702/708, ATS 50–200 ms per UL 1008, STS sub-4 ms, 2–3× cost comparison, Tier III–IV application).
4. Understanding Transfer Switches: Types, Applications, and Design Considerations, Mission Critical Engineers (manual, automatic, breaker-pair, and static switch types; SCR and IGBT characteristics and ratings; 1.5–2 V forward drop; surge capability).
5. Static Transfer Switch vs Automatic Transfer Switch: 4 ms vs 10 s, eINDUSTRIFY (STS operation and phase synchronization described as inrush protection for downstream PDU transformers, ITIC ride-through, NFPA 110 Type 10 ten-second benchmark, open/delayed/closed transition, bypass-isolation ATS, GB200 NVL72 at 120 kW and LBNL data center energy share).
6. STS in Data Centres, Borri (STS as an AC/AC solid-state converter with SCR pairs, 2–3 ms versus 100 ms+ for ATS, 3-pole and 4-pole arrangements, switched versus non-switched neutral and the parallel-neutral hazard, distributed redundancy and failover schemes, single-phase STS at PDU level).
7. ATS vs STS in Data Centers: How Transfer Switches Protect Critical Loads, Data Center Strategy (ATS between utility and standby source, STS between independent UPS outputs and a single-corded load, the generator-start sequence, and why the UPS is what protects IT equipment).
8. Data Center Power Distribution: From Switchgear to Rack PDU, Data Center Strategy (the utility-to-rack power chain, protective device coordination, and where transfer equipment sits in it).
9. UL 1008S — Standard for Solid-State Transfer Switches (scope), UL Standards & Engagement (6000 A / 600 V envelope, Article 702 optional standby systems, service-entrance restriction).
10. Static Transfer Switch Operation and Installation Manual, Exeltech (1.3 V typical SCR voltage drop, published current-versus-power-loss-versus-efficiency table at 98.90%, fan turn-on at 45 °C and thermal shutdown at 70 °C heatsink temperature).
11. NFPA 110: Level 1 vs Level 2 Emergency Power Systems, US Made Supply (Level, Type, and Class definitions; Type 10 at 10 seconds; Class runtime table; testing frequency; transfer equipment as part of the EPSS).
12. Understanding NFPA 110 Generator Testing Requirements (2025 edition), MGIEPS (monthly loaded exercise at 30% nameplate for 30 continuous minutes, cold-start requirement, and the annual supplemental load-bank protocol).
13. The Transfer Failure — NFPA 110 transfer equipment requirements and testing, Uptime Compliance (Section 6.2.1 ten-second measurement from loss of normal power to load energization, the TDOT failure case, and ATS transfer failure as the leading emergency-power citation).
14. Static Transfer Switch product specification, EverExceed (≤ 4 ms and ≤ 8 ms switching modes, break-before-make operation, overload and shutdown behavior, 16–125 A range).
15. STS Static Transfer Switch — seamless dual-power switching for server rooms, BVT Power (dual-source combinations including dual mains, dual UPS, mains plus UPS, and UPS plus generator; sub-4 ms break-before-make operation; contactless construction; 16–125 A ratings).
16. Introduction to Transfer Switches, Electrical Generating Systems Association (UL 1008 scope for automatic and non-automatic transfer switches, the seven functions of an ATS, 6,000 loaded operations, and the UL 1008 versus UL 98 comparison table).
17. Static Transfer Switch (STS): Core Safeguard for Uninterrupted Power Supply, SolarMoo (efficiency above 99% at full linear load, overload envelope of 110% continuous, 125% for 30 minutes, 150% for 2 minutes and 500% inrush, 1–3200 A range, and supported source combinations).
18. Uptime Institute Tier Standard — Tier classifications, Uptime Institute (Tier III concurrently maintainable and Tier IV fault tolerant definitions, and the 99.982% and 99.995% availability targets associated with them).
19. Hospital Backup Power Plan — Understanding NFPA 110, AKCP (Class 10 ten-second restoration, two-hour fire-resistance-rated EPSS room, and Class X on-site fuel storage for hospital systems).
20. MGE Epsilon STS — 200/400/600 A static transfer switch catalog, Schneider Electric (sub-4 ms transfer and detection performance at up to 180° out-of-phase sources, UL-certified withstand to 100 kAIC, fuseless current path with disc-type SCRs oversized 200–600%, key-interlocked maintenance bypass, redundant isolated power supplies and fans).