The vocabulary of the Data Centre layer.

The point where the grid supply enters the facility's ownership: the intake switchroom or substation that receives, meters and passes the feeders to the medium-voltage switchgear. The Power layer ends here and the Data Centre layer begins. A resilient facility takes two feeders, ideally from different upstream substations, and sizes the intake for the end-state load from the first phase, since enlarging it later means a second connection.

See also Medium-voltage switchgear, Facility load, Phased delivery

Switching and protection equipment at the distribution voltages inside a large facility, typically 11 kV to 33 kV, routing power from the intake to the transformers and isolating faults. It is built as lineups of circuit breakers and protection relays, with bus ties that let one intake feed both halves of a 2N design. Lead times now rival transformers', and the choice between air-insulated and gas-insulated designs is constrained by floor area and sulphur hexafluoride restrictions.

See also Utility intake, Transformer, Arc flash

A static device that steps voltage down, in a data centre from medium-voltage distribution to the low voltage at which UPS systems and switchboards operate. Transformers are rated in kilovolt-amperes, sited outdoors as oil-filled or indoors as dry-type units, and their losses count against efficiency every hour. They are now among the longest-lead items in a build, and high-density designs push medium voltage closer to the hall to shorten heavy low-voltage runs.

See also Medium-voltage switchgear, Low-voltage switchboard, Power usage effectiveness

The main distribution board on the low-voltage side of each transformer, holding the air circuit breakers that feed the UPS systems, mechanical plant and house services, and the generator connection point. It is built in separated sections so one can be worked on while others stay live. It concentrates most of a facility's arc flash energy and hosts the generator transfer logic, so commissioning tests it hardest.

See also Transformer, Automatic transfer switch, Arc flash

A system that holds the IT load on stored energy during the seconds between a mains failure and the generators taking over, and conditions the power it passes through. Static systems use power electronics with batteries; rotary systems use a spinning mass, sometimes coupled to a diesel engine. A double-conversion static UPS loses a few percent of its throughput as heat every hour, so partial-load efficiency and a high-efficiency mode bear directly on operating cost.

See also Battery autonomy, Battery chemistry, Standby generator

The time a UPS can carry its rated load from batteries alone, typically five to fifteen minutes for a static system and tens of seconds for a flywheel. It only needs to exceed the generators' start time with margin for a failed first start; longer autonomy buys little beyond a tidy shutdown if the generators fail. Autonomy is quoted at full load and new battery capacity, so a lightly loaded hall with aged batteries behaves differently.

See also Uninterruptible power supply, Standby generator, Battery chemistry

The cell technology behind a UPS, chiefly valve-regulated lead-acid or lithium-ion, with sodium and nickel-zinc cells appearing. Lead-acid is cheap but heavy, short-lived and temperature-sensitive; lithium-ion is lighter, longer-lived and tolerates more cycles, but costs more, needs a battery management system and falls under fire codes that may require separate rooms, suppression and thermal-runaway testing. The chemistry decides battery room size, replacement cycle and insurance terms, which outweigh purchase price.

See also Uninterruptible power supply, Battery autonomy, Fire detection and suppression

An engine-driven alternator, usually diesel, that carries the facility when the grid fails. It must start and accept load within the UPS autonomy, and a Tier III or IV facility must rate its generators for continuous full-load operation and hold at least twelve hours of fuel on site at design load. Generators are the largest items in planning terms and emissions limits may restrict testing hours, so gas engines and fuel cells are being trialled.

See also Battery autonomy, Automatic transfer switch, Load bank

A switch that moves a load between two sources, typically from the failed utility to the generators and back, without an operator. Transfer takes seconds and is covered by the UPS. Open transition has a brief break; closed transition parallels the sources briefly so returning to utility causes no second break. In large facilities main switchboard breakers and control logic do the job, and that logic is what integrated systems testing exercises most.

See also Low-voltage switchboard, Standby generator, Integrated systems test

A solid-state switch that moves a load between two independent power sources within a quarter of a cycle, fast enough that equipment downstream does not notice. It gives single-corded equipment two upstream paths, and in block redundant designs it steers a block of load to the reserve system. An STS is itself a single point for the load behind it, so dual-corded equipment is preferred and an STS used only where that is impossible.

See also A and B feeds, Distributed redundant and block redundant topologies, Automatic transfer switch

A floor-standing cabinet that splits a UPS output feed into breakered, metered circuits for rows of racks, in some markets through a transformer stepping down to rack voltage. Remote power panels, smaller boards among the racks, bring breakers closer to the equipment. The same term is applied to power strips inside racks, which confuses capacity discussions. Its rating sets the load a row can draw, and in high-density halls the circuit count runs out first.

See also Busway, Uninterruptible power supply, Stranded capacity

An overhead track of enclosed conductors running along a row of racks, from which tap-off boxes draw power at any point. It replaces fixed PDU-to-rack cabling with circuits that can be added, moved or re-rated by fitting a different tap-off, without shutting the row down. It is now standard in slab halls because it keeps the floor clear for liquid pipework and a single run can match a row of accelerator racks.

See also Power distribution unit, Raised floor and slab, Design density

The notation for how many spare units a system has beyond the number, N, needed to carry the load. N+1 adds one spare on a shared path, 2N duplicates the whole system and its path, and 2N+1 adds a spare to each duplicate. The notation describes components, so a facility can be 2N in UPS yet N+1 in chillers. A claim should say which system the letters refer to and whether the paths are physically separated.

See also Distributed redundant and block redundant topologies, A and B feeds, Tier classification

Two ways of achieving redundancy with less equipment than full duplication. In a distributed redundant design, three or more systems share the load so any one can fail and the others absorb it. In a block redundant design, several primary systems share one reserve, reached through static transfer switches when a primary fails. Both give near 2N resilience for less capital, at the price of maintaining the load mapping as tenants come and go.

See also Redundancy topology, Static transfer switch, Capacity planning

Two independent power paths to each rack, each from its own UPS and distribution, so a dual-corded server keeps running when either is lost. Each path carries half the load and must be sized for all of it, so resilience depends on keeping neither path above fifty percent. Accelerator servers with several power supplies complicate this, since they may run on the remaining supplies with less margin than specified, which should be tested before loading a hall.

See also Redundancy topology, Static transfer switch, Fault tolerance

A bank of resistive and reactive elements that draws a controlled electrical load so generators, UPS systems and distribution can be tested at full rating before any IT equipment is installed. They are used at every commissioning level and in annual generator tests. A facility never load-tested at full capacity has never been proven at it, and generators that only run lightly loaded accumulate deposits that reduce reliability.

See also Commissioning levels, Integrated systems test, Standby generator

The explosive release of energy when an electrical fault arcs through air, producing heat, pressure and molten metal that can kill anyone nearby. IEEE 1584 incident energy calculations for each switchgear item set protective equipment, approach distances and live work practices. It governs how a power train can be maintained live, so arc-energy-reducing switchgear, remote racking and arc-resistant construction protect staff and decide which maintenance needs a shutdown.

See also Low-voltage switchboard, Medium-voltage switchgear, Maintenance regime

How closely the voltage delivered to equipment keeps the sinusoidal shape, magnitude and frequency it was designed for. Harmonics from power electronics, voltage sags from grid faults and switching transients all count as poor quality, and a UPS exists partly to correct them. Accelerator clusters are themselves a power quality problem: thousands of synchronised power supplies stepping together produce current swings and harmonics that the transformers, generators and grid connection must absorb.

See also Uninterruptible power supply, Accelerator load step, Transformer

Power a facility has paid for and cannot sell, because another limit, cooling, floor space, structural load, circuit positions or a tenant's contracted but unused allocation, is reached first. A low-density hall with a large power allocation strands megawatts behind its cooling limit; a tenant that contracts 5 MW and draws 2 MW strands 3 MW behind its contract. A facility's real value is how much connected power it can deliver and bill.

See also Oversubscription, Capacity planning, Design density

Contracting more IT capacity to tenants than the power train can deliver at once, relying on tenants rarely drawing their full allocation together. The assumption holds for mixed enterprise workloads. It fails for accelerator clusters, which run near full draw for days during training, so a facility hosting AI tenants must size its power train to the sum of their allocations, and an operator carrying enterprise assumptions has less saleable capacity than its plans show.

See also Stranded capacity, Accelerator load step, Take-or-pay

The rapid change in facility power when a synchronised training cluster starts, pauses or checkpoints and thousands of accelerators move between idle and full draw within a fraction of a second. The UPS, the generators on standby power and the grid connection all see the step, and none were designed for a load that behaves as one machine. Operators specify UPS and generator step response; schedulers smooth the ramp or keep accelerators busy in pauses.

See also Power quality, Oversubscription, Uninterruptible power supply

Staged testing that proves a facility works as designed before it carries IT load, numbered one to five. Level 1 is the factory witness test, level 2 inspection on delivery, level 3 start-up and verification of each item, level 4 functional testing of each system, and level 5 the integrated systems test of the whole building. Each has scripts, witnesses and sign-offs, and compressing them to meet a date is the commonest cause of early-life outages.

See also Integrated systems test, Load bank, Practical completion and ready for service

The final commissioning stage, in which the whole facility runs at design load on load banks while failures are induced: utility cut, generators started, UPS discharged, chillers tripped and transfers exercised, with hall temperature watched throughout. It is the only point at which electrical and mechanical systems are proven together, and where control logic faults appear. A tenant or its adviser should witness it, and hall acceptance is usually conditional on the results.

See also Commissioning levels, Load bank, Automatic transfer switch

Two milestones often confused. Practical completion is the construction contract's term for the builder finishing, bar minor defects, and handing the building to the owner. Ready for service is the lease's term for a commissioned hall where the tenant can install and rent starts. The gap, holding commissioning, certification and fit-out, can run to months, and who bears a slip's cost is among the most negotiated points in a lease.

See also Commissioning levels, Ramp schedule, Take-or-pay

The control system that runs a facility's mechanical plant: chillers, pumps, air handlers, valves and dampers, with the electrical side usually monitored by a separate power monitoring system. It executes the sequences of operation that commissioning tested, including the response to a cooling failure, and holds the trend data. It is most often implicated when an outage is traced to a control fault, so its software changes deserve the same change control as the plant.

See also Data centre infrastructure management, Integrated systems test, Chilled water system

Software that records what is installed where, how much power, cooling, space and weight each position uses and has left, and how those figures change over time. Where the BMS controls plant, DCIM manages capacity and sources the utilisation reports an operator gives tenants and financiers. Its value depends on keeping it current, and a facility whose DCIM disagrees with the floor has no reliable view of its stranded capacity.

See also Building management system, Capacity planning, Stranded capacity

The schedule in a colocation contract stating what the operator guarantees, typically power availability at the rack, inlet temperature and humidity within a stated envelope, remote hands response times, and the credits paid when one is missed. Credits are capped, usually at a fraction of a month's fees, so an SLA defines the service without compensating for its loss. The measurement points, exclusions for planned maintenance and definition of an outage decide what the headline figure means.

See also Availability nines, ASHRAE thermal envelope, Remote hands

Stating availability as a percentage of time in service, where 99.9 percent allows about 8.8 hours of outage a year, 99.99 percent about 53 minutes and 99.999 percent about 5 minutes. The figure says nothing about distribution, and one five-hour event in a training run costs far more than sixty five-minute events over a decade. Nines follow the SLA's own definitions, so a facility can report full availability through an event a tenant saw as an outage.

See also Service level agreement, MTBF and MTTR, Tier classification

The planned programme of inspection, servicing and replacement for a facility's plant: generator runs and load tests, UPS battery checks, switchgear inspection and thermography, chiller servicing, filter changes and water treatment. Preventive regimes work to fixed intervals; condition-based ones use sensor data to service equipment when needed. The regime is what concurrent maintainability exists to allow, and the fraction completed on time is a better leading indicator of reliability than outage history.

See also Concurrent maintainability, MTBF and MTTR, Load bank

Mean time between failures and mean time to repair, which together give availability, since availability equals MTBF divided by the sum of the two. Redundancy makes a single failure harmless, lengthening the system's effective MTBF without changing that of its parts. Operations decide MTTR through spares on site, staff on shift and vendor response times, and a long MTTR on a critical component leaves a redundant design exposed for the whole repair.

See also Availability nines, Maintenance regime, Redundancy topology

Continuously matching the power, cooling, space, structural load and circuit positions a facility has to the demand its tenants have contracted and will draw. It runs on DCIM data and tenants' ramp forecasts, and its product is how much saleable capacity remains and when it runs out. Accelerator tenants make it harder in one way and easier in another: their draw stays close to contracted load but changes in large steps with each hardware generation.

See also Data centre infrastructure management, Stranded capacity, Ramp schedule

Colocation sold by the rack or kilowatt in a shared hall, on terms of one to a few years, with the operator providing power, cooling, security and remote hands. Retail is priced per kilowatt per month, with higher unit prices and operating cost than wholesale because of tenant numbers, churn and services. It is the segment least suited to AI hardware, since a shared air-cooled hall cannot supply the density, though operators increasingly offer liquid-ready zones for inference.

See also Wholesale and hyperscale colocation, Pricing units, Remote hands

Capacity sold in megawatt blocks, usually whole halls or buildings, on terms of five to fifteen years to a few large tenants. Wholesale tenants take a turnkey hall and operate their own IT; hyperscale tenants may take a powered shell or build-to-suit with their own fit-out standards. Unit prices are lower than retail and the operating burden lighter; financiers underwrite the lease's credit quality and length, so the facility behaves more like real estate than a service business.

See also Retail colocation, Powered shell, Take-or-pay

The basis on which capacity is charged: per rack in retail, per kilowatt of contracted IT load in most colocation and per megawatt in wholesale, usually per month. Electricity is bundled into the rate or passed through at metered cost plus overhead, and the two give different exposures to power prices and PUE. Offers compare only on the same basis, since a low bundled rate at a poor PUE can cost more than a higher metered rate.

See also IT load, Power usage effectiveness, Take-or-pay

A contract term under which the tenant pays for its full contracted capacity whether or not it uses it. It transfers utilisation risk from operator to tenant and makes a lease financeable, since revenue does not depend on the tenant's ramp. For the tenant it is the cost of reserving capacity in a tight market, and stranded capacity sits on its ledger. Most wholesale and hyperscale leases are take-or-pay after an agreed ramp.

See also Ramp schedule, Stranded capacity, Wholesale and hyperscale colocation

The timetable over which a tenant's contracted capacity and payments step up from lease start to full commitment. A ramp gives the tenant time to install hardware and the operator time to phase fit-out, and typically runs months to a year or two. AI tenants compress ramps because hardware arrives in clusters and draws power within weeks, good for revenue and hard on commissioning, since full load arrives before the hall has run at anything less.

See also Take-or-pay, Practical completion and ready for service, Capacity planning

The points where telecommunications cables enter the building, from boundary ducts through an entrance room to the meet-me room. A resilient facility has at least two entries on separate sides, fed by separate duct routes, so one excavation cannot cut both. Entries are part of the shell and hard to add later, and their number and separation decide how much route diversity a tenant can buy, whatever the Connectivity layer offers outside the fence.

See also Meet-me room, Cross-connect, TIA-942

A secured room where carriers terminate their cables and tenants connect to them, with cross-connects on patch panels under the operator's control. A facility with two meet-me rooms on separate entries can offer genuinely diverse connections; one with a single room cannot, whatever its carrier count. The MMR is the boundary between the Data Centre and Connectivity layers, and an operator that charges for cross-connects earns a recurring fee from every connection made in it.

See also Fibre entry, Cross-connect, Physical security layers

A physical cable, usually fibre, run by the operator between two parties' equipment, most often a tenant and a carrier in the meet-me room. It is ordered as a service, installed in days and billed monthly while it stays in place. Fees are small individually and large in aggregate, and their pricing, and whether a tenant may run its own, affects every later connection. What the cross-connect reaches belongs to the Connectivity layer.

See also Meet-me room, Fibre entry, Remote hands

A service in which operator technicians perform physical tasks on a tenant's equipment: power cycling, swapping drives or cables, racking replacements. It is sold by the hour or included up to a limit, with response times set in the SLA. It lets a tenant run hardware in a facility it never visits; for an accelerator fleet, where a failed component idles a training job, response time and competence with liquid-cooled hardware matter more than the hourly rate.

See also Service level agreement, Retail colocation, Liquid cooling

The concentric controls between the public road and a tenant's equipment: perimeter fencing and vehicle barriers, a gatehouse, building access with identity checks, mantraps into the halls, caged or locked cabinets, and cameras and logs throughout. Vendors label these as security tiers or layers, with no common standard behind the numbers. Audit evidence matters, since the certifications a tenant relies on test whether controls are operated as described, and tenants' own compliance obligations often require inspection.

See also Certifications, Meet-me room, Data hall

Third-party attestations that a facility's controls meet a published standard. ISO 27001 certifies an information security management system; SOC 2 is an auditor's report, with no certificate, on controls against trust services criteria over a period for Type II; PCI DSS covers payment card data and, for a colocation operator, only its physical and environmental controls. None covers resilience or uptime, the concern of the Tier and TIA-942 schemes, and each is only as broad as its scope statement.

See also Physical security layers, Tier classification, TIA-942

Removing a tenant's hardware at the end of its life or lease and installing its replacement. IT hardware is refreshed every three to five years, accelerators faster, against a facility life of fifteen to twenty-five years, so a hall sees several generations, each denser and heavier. It involves certified data destruction, asset disposal and making good; the refresh is when a facility learns whether its density, floor loading and cooling carry the next generation.

See also Design density, Structural loading, Build-to-suit