On-Site Generation: Electrical Integration

Chapter 3.5 made the strategy call: for a gigawatt-class campus in 2026, behind-the-meter generation is the primary power strategy, not a diesel afterthought, because the interconnection queue cannot deliver firm megawatts on an AI revenue clock. Chapter 4.9 covers the molecules, getting fuel to the prime movers. This chapter covers the electrons in between: wiring a privately owned power plant to a privately owned supercomputer, with no utility on the far side of the bus to hide your mistakes. The grid is a stiff source with seconds of spinning inertia and fault current measured in tens of kiloamps. An islanded gas plant has to make its own inertia and fault current, and the load it serves is synchronized AI work that steps up and down fast and often, through power-electronic front ends.

Three decisions dominate everything downstream: island vs grid-parallel, where the transient absorber lives, and grid-forming vs grid-following. Each one drives its own protection, stability and capital consequences. Under all three sits one fact: an AI cluster is mostly power-electronic load, with harmonic current that depends on topology and operating point and load steps synchronized to the workload. A self-built plant has to be engineered around that from the synchronizing relay outward, not retrofitted after the first synchronized GPU pulse trips a generator.

Classical generator-and-bus design assumes a load that is small relative to any single source, statistically smooth, and tolerant of a few hundred milliseconds of frequency excursion. A training cluster violates all three. Tens of thousands of GPUs run the same synchronous collective on the same clock, so their power transitions can align across the hall. To the plant, such a load can swing tens of megawatts in a single step (Meta's Llama 3 405B training, on up to 16,384 H100s; the step grows with GPU count), repeatedly, with a sharp di/dt and a power-factor and harmonic signature set by thousands of switch-mode front ends. Size against the waveform of the specific rack (its idle floor, amplitude and duration), not a millisecond step inferred from an OEM peak rating. SemiAnalysis modeled gigawatt-scale training-load fluctuation as a genuine grid-blackout risk, and xAI's Colossus reportedly burned tens of millions of dollars a year running dummy workloads to flatten its profile (DCD / SemiAnalysis, 2025–26). Colossus 2 has since been fitted with Tesla Megapacks: Canary Media counted 720 containers in satellite imagery from July 2026. xAI has not said what they do.

The prime movers cannot see this load raw. A large gas engine such as Wärtsilä's 50SG is rated to ramp over 100%/minute when hot and to reach full load in under two minutes on a fast start (regular start under five minutes). A GE LM6000 aeroderivative turbine (about 50–57 MW) ramps about 50 MW/minute, close to its full rating, and reaches full power about five minutes after a start (Wärtsilä; GE Vernova product data). Both are geologically slow next to a millisecond GPU step. Keep three responses apart: the online machine's immediate electromagnetic and inertial response, its slower governor and fuel ramp, and an offline start. Place storage where the residual exceeds what the bus can tolerate; Chapter 4.5 sizes repetitive absorption alongside reserved outage duty. The plant's job is to hand the engines a load they can follow and the GPUs a bus stiff enough that they never learn how soft the source behind it is.

Sizing an islanded plant is not 'IT load plus a margin.' It is a stack of multipliers that compound. Start with an IT-to-facility allowance for retained mechanical and electrical draw (1.4–1.5x is an illustrative planning range, not a universal island multiplier). Add a redundancy overbuild (N+1 or N+2 at the generation tier). Then take each engine or turbine's output at the site's ambient temperature and altitude, including its hot-climate derate, and leave headroom for the block- and step-loading transient the machines can actually accept. Stacked, the real-world number is steep. Cleanview's equipment tally for Vantage's Frontier campus in Shackelford County (accessed 2026-10-12) lists about 2.58 GW of planned gross Jenbacher reciprocating-engine generation for 1.4 GW of planned critical IT, roughly 1.84 times the IT load. (VoltaGrid's 2.3 GW is the size of its October 2025 contract with Oracle, a separate figure.) Every multiplier becomes a fuel-supply, emissions-permit and capital cost, so the engine-versus-turbine choice (Chapter 3.5) and its fuel consequences (Chapter 4.9) cannot be made apart from this electrical sizing.

Engine count follows unit rating. Cleanview's planned Frontier mix works out to roughly 120 engines for every 500 MW of gross capacity: 500 × 620 ÷ 2,581, using its listed 250 engines at 3.22 MW and 370 at 4.80 MW. That is the guide's arithmetic, not Cleanview's. By contrast, 27 Wärtsilä 50SG units at the vendor's 18.875 MW 60 Hz rating give about 510 MW gross. Derating, auxiliaries and redundancy raise the count needed to serve retained load.

Block loading is the largest single load step a generator can accept without an unacceptable frequency dip, and it governs the transient spec. No universal ranking puts engines ahead of turbines. Judge each offered machine or package on five points: initial loading, fuel mode, site conditions, the voltage and frequency excursion it permits, and recovery time. Together they fix how much transient the BESS must still cover. For engine gensets, write the requirement as an ISO 8528-5 performance class plus your own load-step profile. Classes G1 to G3 set the allowed transient frequency and voltage deviation and the recovery time for each load-acceptance step, with G3 the tightest. G4 is agreed case by case with the manufacturer. ISO 8528-5 covers reciprocating-engine sets, so write the same parameters into a turbine package's guarantee. Stewart & Stevenson, for example, rates its 2.4 MW GTS-2400 and 3.2 MW GTS-3200 standby turbine packages for 100% block load within ±4.5% transient frequency deviation, a rating that does not carry over to other turbines.

Combustion torque cannot track an arbitrary millisecond GPU edge, but the rotating machine's electromagnetic and inertial response starts at once. One 2026 design path is engines for bulk power, qualified BESS for the residual transient: the battery sources or sinks the fast di/dt in the first milliseconds to seconds and holds bus frequency while the governors catch up over seconds to minutes. NVIDIA's production-BESS guidance gives the facility battery three missions, transient absorption, ride-through and demand response. It warns that they can compete, so it calls for an explicit state-of-charge priority strategy that keeps the bridge charged for the next pulse (NVIDIA, June 2026).

Capacity and bridge energy pass; island acceptance still needs the dynamic test. Three survivors offer 3 × 9 = 27 MW against 23 MW, leaving 4 MW of steady margin. Define the initial deficit D = 23/4 MW, about 5.8 MW. The assumed bridge energy is [D × 0.020 s + ½ × D × 8 s]/3.6/0.95 = about 6.8 kWh, so the 6 MW / 10 kWh reserve passes this power-and-energy screen. The screen does not calculate frequency nadir, bus voltage or reactive-current headroom.

The reserve runs out if governor recovery takes longer than [10 kWh × 3.6 × 0.95 − D × 0.020]/(½ × D), just under 12 s. At a 20 s recovery the bridge needs about 17 kWh and fails, even though three generators still have enough steady MW. Keep the candidate only if the machines demonstrate recovery inside that time; otherwise add reachable reserve or shed load through an approved local control. The plant controller has to detect the lost source, keep the voltage and frequency reference continuous, hold reactive and current limits, carry cooling through the event and move protection to its island settings. Commissioning (Chapter 13.4) witnesses the trip and tests resynchronization permissives before island operation is accepted, with the power-quality recorders of Chapter 4.12 capturing voltage, frequency and state of charge. Fuel delivery has to be proven through the same event (Chapter 4.9), because 27 MW on a generator schedule cannot overcome a failed fuel train.

The four placements form a layered defense. A gigawatt design picks the layers it needs to cover the measured residual, and gives each one a defined reachable load and a defined power and energy duty. Assume one layer covers another's job and you pay for it. You size the BESS for the genset's seconds-to-minutes energy and forget it must also hold frequency in the first cycle, or you oversize engines to chase a transient they cannot physically follow and strand capex in machines that idle inefficiently below their minimum-load sweet spot. Decompose by time constant instead. Chip-level decoupling handles the fastest edges, and rack power-shelf capacitance smooths the sub-second pulse before it reaches the facility bus. The BESS owns milliseconds to minutes and the frequency, and the prime movers own steady-state energy and spinning reserve.

Firming 500 MW–2 GW from reciprocating engines means paralleling dozens to hundreds of engines onto a common bus. Unit size varies: Frontier's Jenbacher mix runs 3.22 and 4.80 MW per engine, Wärtsilä's 50SG is 18.875 MW, and modular blocks are larger still. INNIO's October 2025 order from VoltaGrid covers 92 power packs of 25 MW each, 2.3 GW in all (INNIO, 2025). A power pack is a modular multi-engine block, not a single engine. Each machine must be synchronized to the live bus before its breaker closes (matching voltage, frequency, phase angle and phase sequence within the sync-check relay's window). It must then share load proportionally through droop or isochronous load-sharing controls, so no engine hunts or reverse-powers. At this unit count the synchronizing and load-sharing controller is the plant's nervous system, not a commodity panel. Its failures are catastrophic in a way they never are behind a stiff utility: a mis-synchronized close that slams a machine onto the bus out of phase, or a load-sharing loop that oscillates under the GPU pulse.

The bus-tie architecture is the reliability and maintainability fork. A single bus is cheapest and simplest but a single point of failure, and servicing the switchgear means a full shutdown. A double-ended (main-tie-main) bus splits generation across two sections with a normally open or normally closed tie, so a correctly coordinated fault or maintenance outage takes out one section, not both. It is the workhorse for block-redundant AI campuses and the natural electrical companion to the distributed, block-redundant UPS topology Chapter 4.5 favors. A ring bus gives every circuit two breaker paths: a bus-section fault opens the two breakers bounding it and removes only the circuit tapped there, so the other sources stay connected, at the price of more breakers and harder protection coordination. The choice cascades into relaying. More sections and ties mean more zones, more directional and differential relays and a harder selectivity study, which is why the protection design (Chapter 4.2) cannot be finalized until the bus-tie topology is frozen.

An islanded inverter-and-engine microgrid breaks two assumptions that decades of protection practice rest on: abundant inertia and abundant fault current. A utility-fed bus has both: seconds of synchronous spinning mass damping frequency, and tens of kiloamps of fault current that make overcurrent relays trip crisply and selectively. Strip away the utility and replace half the source with inverters, and both collapse. Inverter sources are nearly inertia-less, so frequency can move fast under a step. They are also current-limited, typically to ~1.2–1.5x of rating; design to the offered unit's phase and sequence current-versus-time limits. A fault on an inverter-dominated bus may not produce enough current to operate a conventional overcurrent relay at all. Protection has to change. Test minimum-fault sensitivity and selectivity in every source state, and add differential, directional or fast communication-assisted schemes where the study requires them (fast bus differential, line-current differential over fiber, source-transfer logic). You can no longer assume a fault announces itself with a 20x current surge.

For an islanded AI plant, a synchronous condenser, optionally fitted with a flywheel, is one design path: the synchronous machine adds physical inertia and fault current, easing the load on grid-forming controls and low-fault-current protection. ABB's orders for VoltaGrid total 27 synchronous condensers with flywheel booked through 2025 plus 35 more in a March 2026 extension. Those 62 machines are meant to make a low-inertia, inverter-heavy island behave, electrically, more like the grid it replaced (ABB / DCD, 2025–26). The choice is study-driven: quantify minimum system strength, RoCoF and nadir, inrush, inverter overload and fault-detection selectivity, then pick a condenser-assisted or condenser-free architecture. A condenser buys mechanical inertia and higher fault current at the cost of capital and losses. A condenser-free design has fewer rotating assets but needs tighter validation of its grid-forming controls and low-fault-current protection.

This fork decides whether a behind-the-meter plant can island at all. A grid-following (grid-feeding) inverter is a current source that needs an existing voltage and frequency reference to synchronize to. Strip the grid away and a fleet of grid-following inverters has nothing to lock onto: they cannot establish a bus or black-start. A grid-forming inverter is a voltage source. It imposes voltage and frequency on the bus, acts as the reference the other sources follow, and can emulate inertia and damping within its power and energy limits. It may support black start when its energy source, auxiliaries, controls and energization duty are qualified. The consequence is binary: an island needs a voltage and frequency reference from a suitable synchronous generator or a qualified grid-forming inverter, and a BESS label alone does not establish that capability. A grid-parallel-only plant can get away with grid-following because the utility is the reference. That is why cheaper grid-following hardware proliferated first, and why the shift to grid-forming is the defining 2026 inverter-control story.

Grid-forming BESS now runs at transmission scale: 200 MW and 275 kV at Zenobē's Blackhillock Phase 1 in Scotland (Phase 2 takes the site to 300 MW / 700 MWh in 2027), the first transmission-connected battery selling inertia and short-circuit-level stability services to NESO (March 2025). Two reported examples show black start at transmission voltage. In 2023 at Redhouse, NESO's Distributed ReStart project used an 8 MW grid-forming battery to energise a 10.6 km, 132 kV line and its 33/132 kV transformer. In its July 2026 report, ZTT says EDC's GS12 battery in Cambodia (100 MW / 200 MWh) energized four 230 kV substations and 182.8 km of 230 kV line during black-start and restoration tests, and supported over 55 MW of regional demand. That is the vendor's account, not an independently verified outage restoration.

A gigawatt AI campus carries two distinct stability risks. The workload itself can force oscillations, and the controls of the sources can interact with one another. Periodic AI training load can force and amplify wide-area oscillation modes (Ko and Zhu, arXiv 2508.16457, simulated on WECC and NPCC test systems). Separately, many fast inverter, governor and condenser controls on one island can interact unless they are tuned together, so the inverters, the synchronous condensers and the engine governors must share the frequency and voltage job without fighting each other. On standards, IEEE 2800 covers interconnection of inverter-based resources, and UL 1741 / IEEE 1547 cover the inverter functions. IEEE 2030.7-2017 specifies microgrid-controller functions above component-control level, and IEEE 2030.8-2018 gives controller testing and performance metrics. Those standards do not qualify your particular source. Before relying on it, test the inverter firmware you chose on how it holds island voltage and frequency, how it limits current, whether it can black-start, and how it behaves alongside the other sources.

Whether or not a campus runs permanently islanded, almost all of them want the option to interconnect: to import during planned maintenance, to export or provide grid services for revenue (Chapter 15.8), or simply to keep the utility as a backstop. That makes the point of interconnection (POI) a two-way electrical and regulatory boundary, and the transition across it the riskiest routine event in the plant's life. The hardware is the synchronizing breaker and its check-sync relay at the POI, the anti-islanding / intertie protection the utility mandates (so the plant never energizes a de-energized utility line and endangers line workers), and the transfer scheme that decides open, closed or seamless transition. Each must be coordinated with the utility's own relays and its dynamic model of the load, the same model submittals Chapter 4.3 governs. To the utility, a self-generating gigawatt campus is at once a large load, a potential generator and a reliability variable it must study.

The regulatory ground moved in late 2025. FERC's December 18, 2025 order directed PJM to offer co-located load three transmission services: network service with an interim non-firm option, Firm Contract Demand and Non-Firm Contract Demand. FERC's April 16, 2026 order partly accepted PJM's interconnection compliance filing and imposed its own Co-Located Load definition. On June 18, 2026, FERC ruled on the transmission services themselves, requiring non-firm service to be available every day when capacity allows, and ordered further tariff filings. A separate docket moved on October 9, 2026: FERC accepted PJM's definition of a Large Load (50 MW or more at a single site) and the Large Load Registry that goes with it. The same order rejected PJM's Interim Resource Adequacy Service, which would have put new large loads lacking capacity of their own first in line for curtailment, so the service is not available. The co-location docket is still live; Chapter 3.2 holds the tariff mechanics and current status. For electrical integration, the tariff does not fix your transition scheme. Non-firm service sets interruptible import rights. It does not decide whether the POI is open, and a grid-connected campus can meet a curtailment with on-site supply or by cutting load. Do not count non-firm imports as guaranteed contingency supply. For islanding, size generation, storage and controllable load to your own retained-load target, and engineer electrical separation as its own design item.

The electrical design cannot be frozen against today's rack power. The roadmap points to a 15–25x increase from H100-class rack power: H100 ~40 kW, GB200 ~132 kW, GB300 up to 142 kW (NVIDIA RA; OEM peaks run to 155 kW), Vera Rubin NVL72 ~188–228 kW operating (OEM Max Q / Max P) and planned at 330 kW TDP in NVIDIA's DSX reference design, and a signposted ~600 kW–1 MW Kyber-class rack (NVIDIA targets H2 2027; SemiAnalysis reported 2028 in July 2026). A denser rack raises potential local power demand, but the transient that reaches the island bus depends on the workload, how well it is synchronized, power caps, and rack and facility smoothing. NVIDIA describes added rack storage on Vera Rubin to cut AC power variation. Size storage power and energy, and the response of the source controls, from waveforms you have measured or a vendor has qualified, one platform and operating mode at a time (Chapter 4.5). Conductor and facility capacity are sized separately.

Hold back BESS pad space, DC-bus headroom, synchronous-condenser foundations, spare switchgear positions and bus-section isolation points, so later additions can be staged while the unaffected sections stay energized. Any tie-in that forces a shared retained-load path to be de-energized means an outage; mark those boundaries in the expansion plan. Treat controller and firmware changes separately, through validation and commissioning. Size the firm-power tier to the contracted IT ramp, and the stability and transient tier to the waveforms you measure.