How much colocation power is usable after a failure?

How much contracted power remains usable after a failure?

Buy continuous IT kW at a named delivery point, in a named hall, under named failure conditions. Separate capacity available for sale, capacity reserved for your organization, energized capacity ready for use and future capacity awaiting construction. A delivery date belongs beside each allocation, not beside a campus-wide MW announcement.

Four quantities determine the purchase: installed equipment rating, operating electrical limit, contractual entitlement and evidence-supported allocation. A signed entitlement does not enlarge a feeder, and a successful test does not reserve capacity for a new tenant.

Read what “contracted kW” purchases. It can describe a continuous ceiling, a minimum monthly billing commitment or an average demand measure. Establish whether A/B pricing buys one protected load or two separately rated circuits, then put surviving kW, permitted density, load shedding, cooling duties and evidence delivery in the technical schedule. Commercial remedies follow the agreement's actual terms.

Which load must keep running?

Specify uninterrupted electrical service, restored service, sustained cooling and application performance separately for each load class. Identify protected racks, interruptible racks, single-cord devices, transfer switches and enforced power caps. Record allowed interruption, restoration time, outage duration and initial battery/fuel state; each condition changes what the promised kW means.

Use the colocation fit-out guide for tenant/operator responsibilities and the grid-capacity evidence checklist for future utility dependencies. Reclaiming headroom belongs to stranded capacity and power utilization; this register establishes which buyer can use that headroom after the specified failure.

Where is colocation power measured?

Measure at the contractual delivery point and trace every upstream constraint to that boundary. A utility meter includes facility services; rack inlets measure the tenant load. Record both boundaries when billing and delivery differ, using the power distribution chain and rack PDU distribution guide to identify the intervening equipment.

Real power in kW describes consumption; apparent power in kVA and conductor current constrain equipment; kWh records accumulated energy. Capture voltage, power factor, phase configuration and each conductor's current. Sum synchronized A/B readings for rack consumption, then calculate the surviving capacity of each feed separately.

The Green Grid's undated PUE glossary defines PUE as total data center energy divided by ICT equipment energy. Annual PUE cannot establish instantaneous rack failure capacity: downstream measurements and scenario-specific auxiliaries and losses determine that allowance. See PUE and energy efficiency for the energy boundary. [S7]

Request instrument ID, CT/PT ratios, accuracy specification, calibration or verification record, time zone and aggregation period. Keep interval averages, captured maxima and event waveforms distinct: an average can conceal an overload or interruption.

Resolve conflicting meter readings by aligning boundaries, clocks, transformer ratios and averaging windows before attributing the difference to losses. Carry stated uncertainty and an agreed acceptance rule into a close pass/fail decision. The CDU capacity uncertainty worksheet applies that discipline to thermal measurements; this buyer register sets no universal meter accuracy class.

Do A/B feeds, N+1 or a Tier label prove failover capacity?

A/B labels identify feeds; the one-line establishes their dependencies. Trace common switchgear, UPS bypass, control power, cooling, fuel and distribution. Test A unavailable and B unavailable independently, then identify shared-component failures that neither directional feed test covers.

N is the capacity needed for the design load; N+1 adds one capacity component, while 2N provides two complete capacity systems. Neither label reserves tenant kW or proves that distribution, cooling and controls are independent. For interchangeable modules sharing a common load, sum the surviving derated ratings, then check their distribution limits.

Uptime Institute's April 2021 explanation defines Tier III by concurrent maintainability and Tier IV by fault tolerance. Obtain the exact award type and facility phase: design certification, constructed-facility verification and operational certification answer different questions. [S15]

Uptime's July 20, 2017 certification guidance explains why topology changes can affect previously certified outcomes. Match the award to the current built configuration and review subsequent changes, using Uptime Tier classification for the wider certification context. [S5]

How much normal loading leaves room for either feed to fail?

Under equal sharing, a fully transferable load at a 24 A surviving-feed ceiling draws 12 A per feed before failure. For two conventional 30 A circuits, Anvilfield estimates that as 40% of each breaker rating using Schneider's October 2024 continuous-application guidance. [S1]

Active/standby supplies, unequal sharing, changed PSU efficiency, inrush and phase imbalance alter the post-failure result. Obtain the actual transfer trace; a normal average cannot establish transient performance.

Which deratings belong in the capacity register?

Record each limiting asset's ID, rated value, operating condition, document revision, permissible value and verification date. Keep equipment derating, redundancy reservation and buyer growth allowance separate.

Schneider Electric FAQ FA104355, revised October 14, 2024, applies an 80% current ceiling to standard-rated continuous applications of three hours or more and allows up to 100% for qualifying 100% rated applications. Confirm the breaker, conductors, terminations and assembly installation conditions together. Apply the locally adopted code and equipment listing; the US manufacturer example does not establish a universal IEC circuit rule. [S1]

Schneider FAQ FA408273, modified January 5, 2026, gives a 320 A continuous limit for a standard-rated 600 A breaker set to 400 A. The as-left setting, rather than frame size, determines that calculation. [S2]

Apply the installed model's derating rule: extending the Galaxy VS temperature slope to VXL gives the wrong result. UPS intake and server inlet are separate thermal boundaries. Obtain OEM treatment of simultaneous temperature and altitude derates.

Cummins' PowerHour explanation of ISO 8528-1:2018 separates duty, annual operating hours and average-load restrictions. An average-load restriction is distinct from an instantaneous ceiling, and ISO DCP is distinct from Cummins DCC. Match the purchased duty sheet, site derates, fuel plan and operating permission to the promised outage duration.

How do you calculate the failure-capacity worksheet?

Anvilfield's buyer worksheet defines usable colocation capacity as the lowest supported contractual, surviving electrical, cooling or enforced-control allowance across normal operation and the required failure scenarios at one tenant boundary, with the agreed interruption and duration conditions met.

At each bottleneck, reconcile surviving capacity, other customers' commitments, non-IT loads, downstream losses and separately held reserves. Convert the remaining allowance to tenant IT kW before comparing it with another boundary. Keep demand and capacity on the same side of every loss conversion.

Anvilfield allocation method, 2026: U_s = max(0, min(C_s, E_s, H_s, P_s)); U = min_s(U_s), including normal operation. C_s is the scenario's contractual allowance; E_s is the tightest surviving electrical allowance; H_s is cooling-supported IT load; P_s is the enforced cap. All values are tenant kW after commitment and reserve reconciliation. Mark a confirmed absent cap “not limiting.” An unknown essential limit or failed transition leaves the allocation unsupported, even if the other entries are numeric. The zero floor indicates no spare allowance and triggers correction when commitments exceed capacity.

Each row also needs duration, initial state, limiting asset, evidence ID/revision, test load, hold time, transient result, status, owner and due date. A normal limit below the surviving-path limits still controls deployment.

Reconcile signed but uninstalled reservations as well as running loads at every shared device. A customer's commitment belongs in both UPS and PDU checks when it traverses both; compare the remaining allowances rather than adding them. At each device, count each commitment once. Calculate either the buyer's total allowance excluding its own existing reservation from other commitments, or incremental headroom including it; label the result. Metered utilization alone does not reveal capacity already sold.

The sustained kW result needs passing transfer, voltage/frequency, thermal and workload conditions over the required duration. Include battery/fuel autonomy, replenishment and return transfer. Use power density planning for load placement and power capping and oversubscription for enforced allocations.

What evidence proves contracted rack failover?

Request the signed script and results, actual connected load, load placement, power factor, topology revision, as-left settings, firmware, ambient conditions, exceptions and synchronized traces. “Full load” needs a named boundary and a number. The electrical commissioning and power QA guide explains the underlying acceptance records; integrated systems testing covers scenario execution.

Trace each test to the buyer's delivery point. A generator load bank connected upstream of the UPS does not establish rack delivery downstream. A resistive load differs from a dynamic GPU workload; record load characteristic and the validation covering that difference. For single-cord equipment, retain the rack transfer device's rating, source-compatibility limits and measured transfer result at the device inlet.

Match battery runtime at the specified load and condition to the complete transfer sequence. Retain recharge state, unavailable strings, capacity evidence and time until another event can be supported. Use UPS and static transfer switch checks and generator acceptance and load-bank turnover to locate the relevant records.

A test at 60% load needs a bounded engineering basis before supporting a 100% tenant claim. Identify the uncovered path, load placement or operating envelope and require targeted verification. After a change, choose affected-scenario verification or a full integrated retest according to the documented impact, rather than treating every change as identical.

Collect recovery evidence: return transfer, restored redundancy, temperatures, alarms and defect closure. An accepted deviation needs a stated kW/scenario consequence, interim restriction and corrective action.

What sampling interval is fast enough?

Specify the event to detect and the shortest contractual acceptance interval, then match waveform sampling, bandwidth, triggers and clock accuracy to it. Logger time resolution alone does not establish interruption-detection capability. Anvilfield recommends 1-second power/temperature trends through the scenario and 1-minute operating exports for the preceding 30 days, plus event waveforms; instrumentation must still resolve faster equipment responses.

Does usable power also support cooling and GPU performance?

Cooling evidence must follow heat to the rack through failure and recovery. DOE/FEMP's July 2024 design guide summarizes the ASHRAE fifth-edition recommended inlet band of 18°C to 27°C for classes A1 to A4. That band is neither a universal GPU limit nor a chilled-water setpoint. [S8]

ASHRAE's May 31, 2022 errata correct the Class H high-density Table 2.2 allowable dry-bulb interval to 15°C to 25°C and identify incorporation into the revised 2022 fifth edition. Select the applicable equipment class and OEM envelope; the two published temperature ranges address different scopes. See data center thermal envelopes. [S17]

Measure the hottest relevant server inlets rather than substituting room averages or return air. Record heat location and air/liquid split when a load bank supports a cooling claim. Obtain evidence for hot weather, restricted water or other design-limiting conditions when the current season does not exercise them.

For liquid cooling, request installed CDU and pump duty, minimum/peak flow, supply temperature, differential pressure, water-quality requirements, residual air cooling and failover trends. Put the actual model limits into the register. CDU commissioning supplies the test context, while the capacity uncertainty worksheet addresses close thermal decisions.

Does “still running” mean full GPU performance?

The DGX H100/H200 has a specified 10.2 kW maximum system draw and six PSUs in a 4+2 arrangement. NVIDIA's October 17, 2026 DGX H100/H200 User Guide states that operation with only three powered PSUs continues at reduced performance, so powered operation alone does not establish sustained GPU performance after a feed loss. With three PSUs on each feed, losing either feed leaves three powered PSUs and invokes that reduced-performance condition. [S11]

Record server model, PSU-to-feed mapping, cold-start capability and ownership of facility, tenant and cloud power caps. Aggregate PSU nameplate watts are not server input draw. Retain workload throughput, GPU clocks, caps and throttling events beside facility traces, and specify the accepted performance or restoration obligation.

How fresh must capacity and test evidence be?

Under Anvilfield's evidence-register method, a capacity confirmation date establishes inventory freshness, while test date, configuration revision and intervening changes establish whether a failover result supports the current claim.

Colocation Scout's verification policy, reviewed July 24, 2026, removes operator-reported regional inventory rows it cannot reconfirm within 45 days. That is an inventory policy, not a failover retest standard. The buyer still needs evidence matching the reserved load, configuration and service conditions. Colocation Scout verification policy

The intervals above are Anvilfield buyer recommendations. A twelve-month register review is not a universal annual IST expiry; retain the site's actual maintenance, certification and change-control obligations separately. A recent test before a protection or firmware change can be less relevant than an older test of the unchanged configuration.

Connect current meter exports and configuration changes to DCIM and asset records. Reconfirm reservations before deployment: sales change shared-capacity headroom without changing equipment.

Claim-to-evidence matrix: what is supported?

Classify each material assertion separately. “Supported” means matching evidence for the stated service; “conditional” means a named dependency with a quantified consequence; “unsupported” means evidence is missing, mismatched or fails the agreed acceptance conditions. Record the lower supported kW separately from the requested but unproven allocation.

Retain document ID, issuer, issue/test date, locator, revision, delivery boundary, scenario, supported kW, duration, limiting condition and acceptance signature. Distinguish issuer records, engineering calculations and witnessed measurements. A weighted score must not let an attractive certificate offset a failed essential capacity claim.

Operator engineering owns topology and allocation reconciliation; the commissioning provider supplies test records. The buyer's technical reviewer reconciles the claim, and procurement carries the agreed deliverables into the commercial schedule. These acceptance roles do not grant switching authority.

When confidentiality prevents unrestricted disclosure, request redacted records preserving load, boundary, scenario and result, a controlled data-room review or a scoped independent witness statement. If essential evidence remains unavailable, leave that claim unsupported and reduce or defer the affected purchase.

Worked example: two 30 A feeds are not 9.48 kW of protected load

The worked rack has two independent 208 V, 30 A single-phase feeds, a fully transferable 4.20 kW IT load and 0.95 power factor. Anvilfield estimates the following limits using Schneider Electric FAQ FA104355, revised October 14, 2024. [S1]

Multiply voltage by allowable current and power factor to obtain real watts, then divide by 1,000: kW = V × A × PF / 1,000. Compare that feed result with every upstream and downstream limit on its surviving path. For balanced three-phase input, multiply line-to-line voltage by line current, power factor and the square root of three, then divide by 1,000. Check every phase and neutral loading separately.

The smaller surviving path controls either-feed service. The 20 A sensitivity therefore rejects the 4.20 kW deployment even though the other feed retains its original rating. Lower voltage or power factor also reduces real-power allowance at a fixed current ceiling; use installation measurements when completing the site worksheet.

A-loss cooling sets the 4.00 kW allocation, below the 4.20 kW design load. Reduce deployment by 0.20 kW or resolve the cooling constraint and close the scenario evidence.

Normal aggregate capacity assumes both complete upstream paths support the sum. The contract, cooling and cap entries are Anvilfield planning inputs. Attach matching site records, duration and passing transition results before accepting the allocation.

Method note: Anvilfield estimates use the November 5, 2026 source check, Schneider's continuous-application rule and the declared voltage, current, power factor and load inputs. Calculate with unrounded values and display two decimals. The evidence pack supplies measured values, applicable ratings and contractual allowances.

What does a UPS temperature rule do to unit-loss capacity?

Select three compatible complete 100 kW Galaxy VS UPS units in parallel at 45°C, with installed parallel support and controls confirmed. Under revision 990-91141L-001's temperature rule, each retains 87.5 kW; Anvilfield estimates 175 kW after one unit is lost, compared with 200 kW from unadjusted ratings. [S13]

For a shared-bus continuation, select 5 kW of non-IT load and downstream loss, leaving 170 kW at the tenant boundary. Deduct 110 kW of other commitments and 10 kW of separate reserve: Anvilfield estimates 50 kW remaining. A downstream PDU with 40 kW remaining limits the electrical allowance to 40 kW; with a 60 kW contract, 48 kW cooling allowance and 45 kW cap, the buyer allocation is 40 kW.

Now commit another 20 kW on the UPS through a different PDU. UPS headroom falls to 30 kW, so the buyer allowance falls to 30 kW even though its local PDU still has 40 kW free. Reconcile each shared device independently and take the tightest allowance; do not add overlapping headroom.

How much time and money does buyer evidence review take?

Anvilfield estimates 44 professional hours and $8,300 for the desktop review below, using selected billing allowances. With one reviewer and eight productive hours per day, that is 5.5 working days after the complete pack arrives. Operator response time, site access and retesting are separate schedule items.

The deliverables are a meter map, completed registers, scenario gap list and technical schedule for procurement. Quote site witness work separately by scenario, test window, travel, instrumentation, vendor attendance, load banks and repeat mobilizations. An incomplete pack adds question cycles; define who must answer and when before fixing the delivery schedule.

Estimating method: Multiply the selected hours by each rate and sum once. The $8,300 allowance excludes contingency, tax, travel, live switching, load-bank or instrument rental and test crews.

Compare provider charges over the same term and delivery boundary using the same included recurring charges divided by evidence-supported contracted IT kW. Show energy, reservation, setup, cross-connect and overage terms separately. A physical capacity shortfall does not create an assumed contractual refund.

Past awards and winning amounts in PorchRocket Bid Control let estimators compare agency power-upgrade scopes with the metering, testing and turnover deliverables in the current bid.

Buyer evidence request checklist

Send this pack to the operator with the proposed rack schedule and service envelope. Assign each unresolved item an owner, due date and effect on the named kW claim. The operator's approved process governs testing; the request pack identifies required evidence and does not prescribe a switching sequence.

• Name site, hall, suite, racks, deployment date, continuous protected IT kW, required duration and accepted interruption.
• Attach the contract schedule and identify its meter boundary and billing units.
• Provide the current one-line with both rack feeds traced through shared dependencies.
• Provide installed ratings, breaker settings and applicable OEM derating revisions.
• Reconcile current load, reserved load and this buyer's allocation independently at every shared bottleneck, then take the tightest remaining allowance.
• Fill normal, A-loss and B-loss rows; add module, utility and cooling scenarios required by the service.
• Attach electrical and thermal test traces, load distribution, hold time, instrumentation, uncertainty/acceptance rule and clock information.
• Confirm battery condition, transfer timing, generator duty, fuel plan and recovery sequence.
• Identify single-cord equipment, transfer devices and any unprotected controls or cooling loads.
• For GPU service, attach power-cap settings, workload performance and throttling records.
• Attach defect closures and repeat-test records for affected scenarios.
• List configuration changes since testing and the evidence that closes their effects.
• Date the availability confirmation and identify who reserves the capacity.
• Record exact certification type and facility phase for each certification claim.
• Assign unresolved claims, permitted interim load and evidence delivery dates; record evidence IDs, configuration revision and acceptance owner.
• Obtain buyer technical acceptance and carry the agreed evidence duties into the commercial schedule.

Accept the allocation when the required scenarios, limits and evidence match. For a conditional claim, record the permitted interim load and the dependency that must close before expansion. Carry ongoing evidence delivery, configuration-change notice and reservation duties into the agreed commercial schedule so the register remains useful after handover.

Sources, methodology and how to cite this page

Anvilfield checked the following eleven primary documents on November 5, 2026. Public manufacturer and agency explanations support the stated applications; the registers, allocation method, review intervals and estimating allowances are Anvilfield's buyer evidence pack.

• S1: Schneider Electric, FAQ FA104355, last modified October 14, 2024.
• S2: Schneider Electric, FAQ FA408273, displayed last modified January 5, 2026.
• S5: Uptime Institute, Myths and Misconceptions Regarding the Uptime Institute's Tier Certification System, web article dated July 20, 2017.
• S7: The Green Grid, Power Usage Effectiveness (PUE), undated glossary.
• S8: DOE/FEMP, Best Practices Guide for Energy-Efficient Data Center Design, revised July 2024.
• S9: Cummins, Generator Set Ratings for Data Centers and Other Applications, PowerHour slide deck, hosted in January 2020 directory; printed title has no completed presentation date.
• S11: NVIDIA, DGX H100/H200 User Guide, HTML last updated October 17, 2026.
• S13: Schneider Electric, Galaxy VS external-battery IEC Technical Specifications, displayed revision 990-91141L-001.
• S14: Schneider Electric, Galaxy VXL IEC Technical Specifications, displayed revision 990-55333E-001.
• S15: Uptime Institute, Explaining the Tier Classification System, April 2021 update.
• S17: ASHRAE, May 31, 2022 errata to Thermal Guidelines for Data Processing Environments, fifth edition (2021).

Use stable links to the meter demarcation register, derating register, failure-capacity worksheet, test freshness register and claim evidence matrix when citing individual evidence units.

Update log: 2026-11-05: Final article review completed; source applications, allocation rules and buyer evidence pack checked.

How to cite this page: Anvilfield, How much colocation power is usable after a failure?, 2026, https://anvilfield.com/field-guides/datacenter/colocation-power-capacity-evidence-register/

Related tools

FAQ

People also ask

• What is stranded capacity in a data center?
• How long does data center electrical commissioning take?
• What is an integrated systems test?
• What is colocation?
• Does a utility letter prove grid capacity?