A watt does not jump from a power plant into an accelerator. It crosses planning studies, protection zones, voltage transformations, energy storage, conductors, and power electronics—each with a different job and failure mode.
A hyperscale campus is not merely a large building. Its size, power-electronic controls, rapid ramps, and ride-through behavior can affect transmission planning and bulk-system operations.
Many data centers run near-continuously, but AI training can switch between compute and checkpoint phases quickly.
UPS rectifiers, server power supplies, variable-speed cooling drives, and GPU regulators do not behave like passive resistance.
A brief voltage sag may leave utility breakers closed while customer protection transfers or sheds compute.
Utility service MW, total facility demand, critical IT MW, installed nameplate, and usable redundant capacity are different numbers.
A duplicate UPS does not create a duplicate path if both feeds share one transformer, switchboard, fuel pump, or rack transfer device.
Transformer, UPS, conductor, and power-supply losses add electrical demand and cooling demand.
Voltages are representative U.S. examples, not a prescription. The most honest “loss” entry is often a mechanism and a meter location: topology, loading, conductor length, and vendor curves decide the number.
| Stage | Representative voltage | Purpose | Loss / how to verify | Primary failure mode | Operator check |
|---|---|---|---|---|---|
| 1. Bulk grid / utility supply | 115–500 kV class | Moves bulk energy to the campus area and establishes the source fault level and reliability envelope. | Upstream of site PUE Utility system losses are not a facility-stage constant. | Generation shortage, transmission fault, congestion, or delayed network upgrades. | Confirm firm service terms, allowable ramps, curtailment rights, and energization milestones. |
| 2. Receiving substation | Example: 230 → 34.5 kV | Terminates utility service, transforms voltage, meters demand, and defines protection zones. | Core + copper loss Use transformer test report/nameplate and interval meters; do not assume one percentage. | Transformer outage, bus fault, relay miscoordination, or common incoming-line exposure. | Check N-1 load transfer, spare strategy, relay study, fault duty, and actual loading. |
| 3. Campus medium voltage | 13.8 or 34.5 kV examples | Moves power between substations, buildings, data halls, cooling plants, and generation with lower current than LV distribution. | I²R + dielectric Calculate from route, conductor, current, and power factor; trend feeder meters. | Cable damage, splice failure, ground fault, or a loop that cannot isolate a section. | Validate selective coordination, sectionalizing points, route diversity, and ampacity at worst ambient. |
| 4. MV/LV switchgear | MV in; often 480 V after step-down | Interrupts fault current, isolates equipment, and transfers/parallels sources under a defined sequence. | Contact + bus I²R Inspect thermal scans and millivolt drop; no voltage conversion in a plain breaker. | Breaker fails to open/close, controls lose DC power, arc fault, or transfer logic acts incorrectly. | Test trip units, interlocks, protection selectivity, control power, and maintenance bypass. |
| 5. UPS / ride-through | Common example: 480 V AC in/out | A VFI/double-conversion UPS rectifies AC to a DC link and inverts it back, bridging sags and the generator-start interval. | ≥95% efficient DOE reported 95%+ for 2023 double-conversion systems; verify the vendor curve at actual load. | Battery string, inverter, bypass, controls, or a transfer outside the IT ride-through envelope. | Trend module load, battery impedance/temperature, bypass availability, and end-to-end transfer tests. |
| 6. LV distribution / PDU | 480 V or 415/240 V examples | Feeds rows or busways; transformer-based PDUs may create 208/120 V for legacy loads. | Transformer + I²R Higher voltage and fewer conversion steps reduce current and avoid avoidable heat. | Overloaded neutral/phase, transformer hot spot, breaker miscoordination, or shared PDU choke point. | Balance phases, measure harmonics, verify transformer load factor, and reserve real—not paper—capacity. |
| 7. Overhead busway | Often 415/240 or 480/277 V | Distributes high-current power along a row with movable, metered tap-off boxes. | Bus + joint I²R Calculate voltage drop and inspect joints thermographically under load. | Loose joint, incorrect tap rating, phase imbalance, or arc flash exposure during work. | Record tap location/rating, torque/thermal history, fault rating, and A/B physical separation. |
| 8. Rack PDU + cords | 208 or 240 V AC examples | Meters and branches row power to server power supplies; dual-corded equipment should land on independent A/B paths. | Cord + connector I²R Meter per outlet/branch and watch connector temperature; higher voltage means less current for the same power. | Branch overload, accidental unplug, bad connector, or both cords on one source. | Enforce A/B labeling, phase balance, breaker headroom, locking cords, and per-outlet telemetry. |
| 9. Server PSU → DC bus → VRMs | AC → DC bus → chip rails | Power supplies rectify rack AC; point-of-load regulators create the low-voltage, high-current rails used by GPUs, CPUs, memory, and fans. | Conversion curve Use measured server input versus component telemetry; every conversion loss appears as heat in the rack. | PSU/VRM trip, firmware power cap, transient overshoot, or cooling-driven throttling. | Compare rack input, PSU telemetry, accelerator power, throttling, and workload phase timestamps. |
Start with the compute profile, decide what must ride through, add distribution losses and cooling, then test each normal and contingency state. Starting with campus nameplate reverses the logic.
This is a transparent arithmetic model, not a “typical campus.” PUE is an energy ratio; using it with constant power assumes the same load for the full interval.
Facility power = IT power × PUE
100 MW × 1.20 = 120 MW
Annual energy = 120 MW × 8,760 h
= 1,051,200 MWh/year
If that entire 100 MW IT segment is behind a 95%-efficient UPS, the UPS needs 105.26 MW input and rejects 5.26 MW as loss. At 97%, input falls to 103.09 MW; the 2.17 MW difference equals about 19,010 MWh/year at continuous full output. Do not add this again if the measured PUE already includes it.
Use when continuous conditioning and a defined AC–DC–AC ride-through path justify conversion loss.
Both bridge an interruption; duration, cycling, footprint, maintenance, and fault behavior drive the choice.
Keep voltage higher longer when equipment supports it; the same power at higher voltage requires less current.
Engine generators sustain the site through a long utility outage; the UPS bridges the start/transfer gap and conditions sensitive load.
N, N+1, and 2N describe spare capacity and paths; they do not, by themselves, certify availability. Uptime’s Tiers are performance outcomes: basic capacity, redundant capacity, concurrent maintainability, and fault tolerance.
| Pattern | Concrete meaning for N = 4 modules | Relative scope | What it can provide | Use when | Gotcha |
|---|---|---|---|---|---|
| N | Four required modules; zero spare. | Lowest | Capacity only; maintenance or one failure can shed load. | The workload can stop, migrate, or checkpoint without unacceptable loss. | N at UPS does not mean N at transformer, cooling, fuel, or controls. |
| N+1 | Five equal modules; any four carry design load. | Moderate | One redundant capacity component; can permit component maintenance if paths and isolation cooperate. | One component failure must be tolerated and software/site diversity covers larger events. | A common bus remains a shared failure domain; low module loading can reduce efficiency. |
| 2N | Two independent four-module systems, A and B; either path carries 100%. | Highest of these patterns | Full duplicate capacity and distribution paths when the duplication truly reaches dual-corded IT. | The facility must maintain service through a complete path outage. | Cross-ties, shared control power, rack ATSs, or both cords on A can collapse “2N” into one path. |
| Distributed redundant / catcher | Example: three loaded blocks plus a shared spare block that can accept one failed block. | Between N+1 and 2N; topology-specific | Capital-efficient block failover with fewer fully duplicated assets. | Large modular halls can transfer blocks selectively and operators can manage switching complexity. | Protection, transfer capacity, and simultaneous maintenance states must be modeled—not inferred from the block count. |
Onsite generation can accelerate or reinforce supply, but it creates a second interconnection problem: fuel, emissions, protection, islanding, synchronization, and utility operating agreements. It does not erase the grid.
| Option | EIA reference scale | Lead-time assumption | Capital anchor | Regulatory / operating posture | Use / do not use |
|---|---|---|---|---|---|
| Natural-gas combustion turbine | 211 MW aeroderivative; 419 MW industrial frame | 2 years | $2.275M/MW aero; $1.158M/MW frame | Air permit, firm gas and often pipeline work, noise/water review, utility protection and parallel-operation agreement. | Use: dispatchable bulk onsite power. Not: as a substitute for UPS ride-through or without firm fuel analysis. |
| Gas reciprocating engines | 7 MW average new unit installed in 2024 | Vendor/site-specific; EIA actual-cost series gives no comparable lead time | $1.490M/MW average installed in 2024 | Modular air and noise permits, gas service, many-unit controls, maintenance and utility interconnection. | Use: modular phasing and black-start-capable designs. Not: when emissions/noise footprint or unit count is unacceptable. |
| Solid-oxide fuel cell | 10 MW | 3 years | $8.287M/MW | Gas supply and air permitting still apply; inverter controls and island/parallel behavior need study. | Use: modular steady generation where the emissions/footprint case works. Not: assume “fuel cell” means fuel-free or zero-emission. |
| Small modular reactor | 480 MW plant (6 × 80 MW) | 6 years; first available 2031 in AEO2026 model | $9.831M/MW | NRC licensing, environmental/site review, security, emergency planning, fuel and waste obligations; project-specific approvals remain decisive. | Use: treat as a long-horizon firm-supply scenario. Not: present it as a near-term drop-in campus generator. |
| Solar PV + 4-hour battery | 150 MW PV reference; storage bundled in cost case | 2 years | $1.903M/MW of PV reference capacity | Land, local permits, inverter interconnection, charging source and dispatch rules; capacity accreditation varies by system. | Use: energy/capacity shaping and resilience as modeled. Not: claim 24/7 firm coverage from nameplate MW or four-hour storage. |
The serving utility/transmission owner studies whether the new demand violates thermal, voltage, stability, short-circuit, or reliability limits.
Onsite supply that operates in parallel—or a remote plant built for the campus—may need a separate generation study and agreement.
Training transitions can move load faster than conventional scheduling intervals, so the grid needs ramp rate, protection, and ride-through models.
A tariff converts engineering risk into contract terms: minimum billing demand, ramp/curtailment provisions, collateral, dedicated-facility cost, and service duration.
Most bad explanations mix a time scale, a boundary, or a reliability term. Correct the category and the architecture becomes easier to reason about.
Numbers and dated process claims link to the issuing agency, lab, standards body, or operator. Vendor-specific loss claims were deliberately excluded from the chain table.