On-Site & Bring-Your-Own-Power Generation (Energy-Supply Strategy)

For two decades, on-site generation at a data center meant a yard of diesel gensets that ran a few hours a year on an NFPA test schedule. By 2026, on-site and bring-your-own-power (BYOP) generation has become an increasingly common primary supply strategy for new gigawatt-scale AI campuses. Operators do not want to run power plants. The grid simply cannot energize them on the schedule their capital plan assumes. When the scarce resource is megawatts-and-when rather than chips or capital (Chapter 3.2), waiting in the queue is one option and generating your own electrons is another.

This chapter is strategy, not engineering. Electrical integration (sizing, paralleling, step-loading into synchronized GPU pulse loads, protection coordination, grid-forming inverters) is in Chapter 4.8; gas-process and fuel-conditioning engineering is in Chapter 4.9. A strategist and a lead engineer should agree on five forks before any prime mover is ordered: bridge or permanent, speed or efficiency, islanded or grid-parallel, firm or interruptible fuel, and which emissions basis governs the whole plant.

Speed has a price. Getting 200 MW online six months early brings forward about $1.2–1.3B of gross revenue (Chapter 2.2 works the arithmetic), while the GPU capital inside sits idle and its economic life runs down. Whether on-site power buys you those months depends on how your campus's grid-service date compares with the first-power date an on-site plant can contract for. RMI's November 2025 analysis of PJM puts the average wait at more than eight years: generating projects that came online in 2025 had taken that long from interconnection application to commercial operation. Those were finished generators, not data centers waiting on load service, so judge your campus by its own executed service offer and energization milestones. For ERCOT's large-load queue, Batch Zero and how energization works there today, see Chapter 3.2. Your realistic options are on-site generation you can contract and permit, an existing powered facility, or grid service that is actually available. On-site generation wins when it beats the grid-service date and turns a powered shell into a revenue-earning asset on the schedule the silicon demands.

A common bridge scenario runs in three phases. Phase 1 (now): reciprocating engines or aeroderivative turbines, paired with BESS, supply first power if the delivery and commissioning dates you can contract for come before the grid-service date. That first-power date comes out of five schedules together: the equipment slot, the air permit, the fuel lateral, site construction and commissioning. Wärtsilä's 282 MW data-center plant, announced October 5, 2026, shows a real timeline. The order was booked in Q3 2026, the equipment arrives in 2028, and full operation is expected in 2029. Phase 2 (2028 onward): add efficiency where the plant will live for years, with steam tails, combined cycle or solid-oxide fuel cells, and firm up fuel supply. Each option runs on its own clock: fuel cells can be installed in months, but a frame turbine ordered in 2026 delivers in 2028–2032 and then needs a 24–36 month install. Phase 3 (2027–2035): existing-nuclear restarts and uprates arrive this decade, SMRs and 24/7 carbon-free renewable-firm hybrids arrive in the 2030s, and where grid service has arrived, on-site gas steps down to backup. The phases overlap in time, so treat them as parallel tracks rather than steps. The danger sits in the seam between phases. A bridge plant that cannot be cleanly demoted, because the grid connection never arrived or its air permit assumed temporary operation, becomes a permanent plant you under-engineered or a stranded asset you cannot recover. Whether you can demote it turns on an executed grid-service plan: how much capacity it can deliver, and when it energizes. A permanent plant, islanded or grid-parallel, ends differently, because it can remain the primary supply. Whether you are building a bridge or a permanent plant is the first fork, and it governs every choice below it.

Once you know whether you are buying speed or efficiency, three variables settle the prime-mover choice: lead time (how fast you can get the iron), efficiency or heat rate (what you burn per MWh) and permitting posture (what the plant emits and how hard that is to permit). The four combustion options plus fuel cells and nuclear span the tradeoff space. Read the table below as what you give up to get what you need, not as a spec catalog.

Reciprocating engines (RICE) from Wärtsilä, INNIO Jenbacher, Caterpillar and Bergen/Everllence are the speed-and-flexibility play. They deliver high simple-cycle efficiency, up to 50% for the 60 Hz Wärtsilä 50SG. Wärtsilä rates the 50SG to reach full load in under two minutes on a fast start and under five on a regular one, to ramp hot at more than 100% of rated load per minute, and to turn down to a 10% unit minimum load. Those figures describe the 50SG, not every engine. A ramp rate is also not step-load acceptance. A synchronized AI cluster can swing its draw by tens of megawatts almost instantly, faster than any engine's governor and turbocharger can follow, so an islanded plant needs BESS, supercapacitors or grid-forming inverters to absorb the pulse whatever the prime mover (Chapter 4.8). Where engines win is start time, deep turndown and modular reserve. The cost is modularity as a burden: a 500 MW plant needs 25–250 units across the 2–20 MW range, or 27 Wärtsilä 50SG units at its 18.875 MW 60 Hz rating, before reserve and maintenance margin. That is a lot of paralleling, switchgear and O&M headcount. Aeroderivative turbines (GE LM2500 ~34 MW, LM6000 ~52–57 MW; refurbished ProEnergy PE6000 from repurposed jet cores) are denser per unit and start in 5–10 minutes. They run lower simple-cycle efficiency (37–42%), derate hard in heat and bottleneck on turbine-blade casting supply. Heavy-frame turbines and combined cycle pass 60% efficiency in CCGT configuration and carry 24–36 month installs. They are the efficiency endgame, not a bridge. Their slots run years out. Mitsubishi says orders booked in its first quarter of fiscal 2026 will deliver in 2028–2030, and on September 16, 2026 GE Vernova reported reservations for delivery in 2032. A slot reservation commits a manufacturing position. It is not an equipment order or a commissioned plant.

Three announcements from mid-2026 show how big the on-site and grid-enabled path is getting, and all three are still permits and plans rather than energized plant. PORTS-Pike (Pike County, Ohio) follows a grid-enabled path. OpenAI takes up to ~8 IT-GW on a 20-year lease at SB Energy's campus on the former Portsmouth site, backed by ≥10 GW of new generation including 9.2 GW of gas and ~$4.2B of AEP Ohio transmission structured to shield other ratepayers. OpenAI's August 17 announcement expects the first 800 MW in 2028, largely using existing AEP infrastructure, with expansion needing new grid-connected generation and transmission. NVIDIA is the exclusive compute provider and credit support behind the initial tranche (→ financing structure in Chapter 2.5). PORTS is not DOE's only former enrichment site going to AI. At the Paducah site in Kentucky, Brookfield and DOE announced on July 29, 2026 a campus of more than 1.2 GW of compute, backed by up to 2 GW of gas that NextEra would build and own and up to 2.6 GW of batteries. Big Rivers Electric would provide wholesale service, and the deal still needs definitive documentation. GW Ranch (Pecos County, Texas): Amazon confirmed ownership of the 8,000+-acre site. In TCEQ's preliminary determination the plant has 35 turbines, a nominal 5,000 MW output, and no way to import power from a utility or export power to one. Pacifico's January 2026 announcement lists its own figures: 7.65 GW gross of gas, 1.8 GW of batteries and 750 MWac of solar. The site is initially islanded from ERCOT entirely. In August 2026 Amazon told Data Center Knowledge it was looking at moving to front-of-the-meter grid service, but gave no date. That is the bridge-or-permanent fork playing out inside one project. So far none of the three adds to the ~2 GW of BTM generation running in mid-2026.

Combined cycle, CHP and thermodynamic optimization are an energy-strategy choice, not a default. Bolting a steam tail onto simple-cycle exhaust lifts a 42% plant toward 60%+, but it adds a steam turbine, a heat-recovery steam generator, a condenser and a water demand that collides with the water-availability siting gate (Chapter 3.7). For a bridge plant that steps down to backup in three years, the steam tail is capital you mothball. For a permanent plant, it is the difference between competitive and uncompetitive delivered energy. The bridge-or-permanent fork decides it. Combined heat and power, where reject heat warms an adjacent load, is a tempting middle path. In most AI siting there is no co-located thermal host, so CHP economics rarely work unless heat reuse or district heating is itself a siting feature (Chapter 15.5).

Nuclear is the option whose strategic role is most often misstated, and it splits into tracks on very different clocks. Existing-plant restarts, uprates and life extensions are real megawatts on a near-term clock, but each follows its own schedule and slips like any megaproject. Constellation aims to restart Crane, the 835 MW plant behind Microsoft's PPA, in 2027. Crane's first PJM study found upgrades running as late as December 2030. FERC's June 1, 2026 waiver allows Eddystone's 760 MW of capacity interconnection rights to be transferred, which may shrink those constraints or remove them. PJM's studies determine deliverability. Separately, DOE's Federal Power Act §202(c) emergency orders have kept Eddystone Units 3 and 4 available to operate through November 20, 2026, at a cost PJM allocates to load (Chapter 3.11). At Palisades, Holtec's September 21, 2026 update says the switchyard reconnected on September 11 and testing and restoration continue, but fuel loading stopped after an assembly tilted. It gives no commercial-operation date. NextEra's September 8, 2026 update targets the 615 MW Duane Arnold restart by the first quarter of 2029. Meta's Clinton deal works another way, keeping an operating plant running and adding a 30 MW uprate.

SMRs (Kairos, X-energy, Oklo, NuScale) belong to the 2030s, and their licensing risk is still live. OPG plans to connect its first Darlington unit to the grid by the end of 2030. Two 2026 steps narrowed that risk without removing it. The NRC finalized its Part 53 rule in March 2026, an optional risk-informed licensing path under which NRC staff expect design approvals in 18 months or less. Four DOE-authorized advanced test reactors reached zero-power criticality by July 4, 2026. A zero-power test is not a power plant, so keep SMRs off the energization schedule. No single cost figure covers them. DOE's September 2024 reference model puts the median overnight cost of a 300 MW small boiling-water reactor at ~$13,000/kW, with a wide upper tail, against ~$8,500/kW for a large reactor (DOE Liftoff, Fig. 23). EPRI's August 2023 report 3002026582 assumes lower overnight costs for light-water SMRs in its Table 5-1: $8,460/kWe for the first of a kind and $5,134/kWe for the nth, in 2022 dollars. In May 2025 OPG budgeted C$6.1B for the first 300 MW Darlington unit, plus C$1.6B for systems shared by all four units. Spread the full C$7.7B over that first unit and you get ~C$25,700/kW. That is project budget divided by capacity, not an overnight cost. In March 2026 the NRC voted to award TerraPower a construction permit for its Natrium sodium-cooled fast reactor at Kemmerer, Wyoming, which TerraPower expects to complete in 2030 (TerraPower, 4 March 2026). On September 28, 2026, the NRC issued TVA's Clinch River Unit 1 construction permit. That permit lets TVA build. It gives no commercial-operation date and no delivered-power price, and financing, completion, operating authorization and deliverable electricity each remain milestones to track on their own.

The strategic mistake is to treat either track as a fast answer to power you need in 2026. Restarts are a scarce, largely one-time inventory of clean-firm capacity that a handful of hyperscalers have already locked up. SMRs are a hedge you can underwrite as optionality but should not put on the critical path. Deloitte's April 2025 analysis estimates that new nuclear capacity could meet about 10% of the projected rise in US data-center power demand over the next decade: meaningful, late and clean, but not the answer to 'I need 500 MW in 2027.' For firm-clean-supply procurement structures (17–20 year PPAs, take-or-pay, FERC co-location treatment), see Chapter 3.4.

Serving a flat, inflexible load from wind and solar takes a large overbuild. In LUT University's Finnish study, updated June 16, 2026, the modeled renewable baseload system needs roughly 7x as much wind and solar nameplate as demand, and costs €103–121/MWh. To supply the load without interruption, the study also adds renewable-fuel backup big enough to carry the whole data center, since wind, solar and batteries cannot ride through long stretches of low output on their own. The multiplier is a site-and-storage result, not a universal figure. If about 10% of demand can flex at the right times, the study finds annual generation costs fall by roughly 15–24%. It does not say what that does to the capacity multiplier. Overbuild of that order is why pure renewables off-grid rarely anchors a gigawatt AI campus today: land, capex and storage all balloon. Renewables-anchored hybrids do work as the clean fraction of a firmed portfolio: solar and wind for the energy, gas or fuel cells for firmness, and BESS for the seconds-to-minutes ride-through and the inertia an islanded low-rotating-mass plant otherwise lacks. Be careful what that portfolio is called. Wind and solar backed by ordinary natural-gas engines or gas-fueled SOFCs give you reliable power that is partly clean. 24/7 carbon-free energy (CFE) means carbon-free supply meets the load in every hour, so the firming energy has to count as carbon-free under the same accounting. Report the hourly CFE share and plant availability as two separate numbers. The decision is how much clean fraction you will pay the overbuild premium for, against a corporate CFE commitment and an increasingly carbon-aware permitting and community environment (Chapter 3.11 on social license; Chapter 15.4 on stewardship).

The most under-appreciated failure mode in BYOP is solving the electrical critical path and forgetting the molecular one. A gas plant is only as firm as its fuel, and firm pipeline transport is scarce. The lateral and midstream buildout to bring enough gas to a gigawatt-class site can take as long as the turbines, or longer, and INGAA projects North America needs roughly +39% pipeline capacity by 2052 to serve the load. A gas-supply interruption is functionally a generation outage, so fuel firmness is a reliability-engineering problem, not a procurement footnote. The lead-time register in Chapter 2.3 therefore carries the pipeline lateral as a long-lead item alongside the prime movers and transformers.

The decision is firm versus interruptible transport, and how to build firmness when firm transport is unavailable. Firm transport buys reserved pipeline capacity and priority of service for a reservation charge; you buy the gas itself separately. It is expensive and hard to get, and a physical outage or force majeure can still stop delivery. Judge the whole fuel path and how long the site can run on its fallback (Chapter 4.9). Interruptible transport is cheaper but can be curtailed when you need it most, during a correlated cold snap when residential heating has priority. The practitioner answer is 'synthetic firm': layer multiple pipelines, interruptible transport, on-site LNG/CNG storage and dual-fuel switching so no single curtailment takes the plant down. LNG/CNG also lets a site run before the lateral is complete. Choose between them on the offered trailer’s usable HHV energy, sendout, round-trip time and delivered cost. No standard truck count or haul radius tells you how long the site can run. Sizing the days of storage that hedge a cold-snap curtailment is a fuel-process engineering problem, detailed in Chapter 4.9. The strategic point is that supply interruption is a firmness risk co-equal with the electrical interconnection, and a BYOP plan that treats fuel as a given has an unhedged single point of failure.

The islanded-or-grid-parallel fork is as much a regulatory decision as a technical one, and the answer depends on the market. Fully islanded generation never touches the grid. That is the cleanest regulatory path, since you avoid interconnection queues and most FERC/ISO jurisdiction, at the cost of the full overbuild obligation above and no grid backstop. Grid-parallel or co-located generation can lean on the grid as reserve, but only up to the firm withdrawal capacity you have bought, and only as much of it as can physically be delivered at the boundary. Picture a site with 900 MW of its own generation and firm grid service of 100 MW: if the plant trips, the site cannot pull 1,000 MW from the grid to cover it. Non-firm service can be curtailed when the grid runs short.

Co-location also drags you into the rules that took shape in 2026. FERC's December 18, 2025 PJM order directs three co-location transmission services: Firm Contract Demand, Non-Firm Contract Demand, and Interim NITS, available only to customers seeking NITS while required network upgrades are built. Existing retail arrangements that net behind-the-meter generation get a three-year transition to December 2028; other co-location service does not. On June 18, 2026 FERC set the services' terms, requiring non-firm service to be offered every day capacity allows, and ordered further PJM tariff filings, so the docket is still live. Both orders are specific to PJM, and neither is an ERCOT tariff, so price each site under its own service and tariff. The live controversy is cost causation: who pays for the grid backstop a co-located load relies on, and how to keep that cost off other ratepayers. The market changes the answer. ERCOT's lighter-touch, energy-only structure differs sharply from PJM's capacity-market co-location order and from a vertically integrated utility where the PUC governs the tariff. Regulatory treatment and PPA structures are in Chapter 3.4; the point here is that where you island determines whether islanding is the cheaper path.

The recurring irony of bridge power is that operators turn to on-site gas to escape a multi-year grid wait and then run into a multi-year air-permit wait. A combustion plant of any size is a stationary source. Which rules apply turns on the site's potential emissions, its operating limits and the area's attainment status for each pollutant. In attainment or unclassifiable areas, a major new source or major modification triggers PSD review and BACT. For a pollutant where the area is in nonattainment, major nonattainment NSR brings LAER, emission offsets and public participation. Securing the offsets can decide whether the site can be built at all. Title V and GHG-reporting applicability are tested separately. In a nonattainment area, expect dispersion modeling and community health-risk assessment even for nominally 'emergency' fleets. Even Texas, long the easy state for permits, often takes more than a year to issue an air permit. Since September 21, 2026 there is also a hold: the governor told TCEQ it should issue no permits sought by data-center projects before ERCOT's audit is done, and TCEQ reads the hold as covering air and water permits for infrastructure that directly supports those projects. The schedule advantage you bought with fast turbines can be eaten whole by the permit. That is why the air permit belongs on the critical-path register (Chapter 2.3) and why the full federal, state and local matrix is in Chapter 3.9.

Two 2026 developments sharpen the fork. First, in January 2026 EPA finalized an additional 'temporary turbine' subcategory in the combustion-turbine NSPS (subpart KKKKa). The subcategory covers units of up to 850 MMBtu/h that stay at one location for 24 months or less. Bridge turbines get a lighter permit path because the rule assumes they will be removed. The cliff is hard. A unit that stays longer loses temporary status for its whole time in service, and swapping in another temporary unit does not reset the clock. A bridge plant permitted as temporary that the grid connection never relieves faces permit-cliff and citizen-suit risk, because the assumption that made it fast to permit becomes a liability when 'temporary' turns permanent. Second, SOFCs avoid combustion-turbine-specific standards and sharply cut criteria pollutants, which is strategically valuable. The site still goes through whatever air-permit and emissions-inventory review applies to it. The emissions decision feeds the prime-mover choice and the bridge-or-permanent fork at the top of the tree. Water and discharge permitting for any steam-cycle or evaporative-cooling component compounds it (Chapter 3.7).

The energy-supply strategy is the product of four nested decisions, taken in order. (1) Bridge or permanent? It sets your efficiency tier and permit posture. (2) Speed or cheap energy? It sets your prime mover: engines or aeroderivatives for speed, CCGT or nuclear for cost. (3) Islanded or grid-parallel? It sets your overbuild obligation and your FERC/ISO exposure. (4) Firm or synthetic-firm fuel? It sets your second critical path and your cold-snap resilience. Each answer constrains the next, and each downstream cost lands years later: the mothballed steam tail, the stranded high-heat-rate bridge, the cost-causation surprise, the curtailed pipeline. The operators getting this right in 2026 settled bridge-or-permanent first and let every other choice inherit from it. The cleverness of the prime mover matters far less.

Only a bridge plan ends up on grid service, and the 2027–2035 pipeline of queue reform, grid upgrades, existing-nuclear restarts and clean-firm hybrids can reach your site in time or miss it. Model the bridge and permanent cases side by side over one horizon, and count maintenance, fuel, permit conditions and any future grid charges in both. Use BYOP to win the speed race without over-committing to a plant that later grid service would make redundant. The grid-supply alternatives BYOP is racing against are in Chapter 3.3 and Chapter 3.4.