Data center on-site power generation, fuel cells, and microgrids

On-site generation, and the constraint that put it back on the table

On-site generation is power produced at the data center itself, behind the utility meter, to serve the campus as a primary source instead of only backing it up when the grid drops. The difference from the standby generators most people picture is the duty. A standby set sits idle and runs a few hours a year on a utility failure. An on-site primary plant runs as the working source, carrying real load for long stretches or continuously, with the grid behind it or sometimes not connected at all.

The reason this went from a niche idea to a leading way to power a campus is the grid. Power availability is now the gating constraint on the AI buildout, and the interconnection queue to connect a large new load runs years in the busiest markets. The grid-utility guide covers that wait in depth. The short version is that operators can buy the chips and raise the money faster than the utility can deliver the megawatts, so they make their own power on site and skip the queue.

This guide is about the power plant the data center builds for itself: the generation options, the microgrid that ties them together, how it islands or runs alongside the grid, the fuel and emissions that come with running a power plant, and the economics that actually drive the decision. The grid connection upstream of the fence is the subject of the grid-utility guide. The standby generators that exist only to ride a grid outage are the subject of the generator sizing guide. Read those for the grid and the backup. Read this for the primary power the site makes itself.

Why are data centers building their own power?

Data centers build their own power to get energized on a schedule the grid cannot meet. A large new load files into the utility interconnection queue, waits for the study and the agreement, and often waits again on grid upgrades miles away, a process that has been reported at four to seven years in constrained markets. The building goes up in 12 to 24 months. The gap between a finished shell and a powered one is the schedule risk that on-site generation answers by making the power the project controls.

The push behind it is AI compute. Rack densities jumped from the old 5 to 15 kW range into 50 kW, 100 kW, and beyond, and a campus full of those racks lands in load territory the local grid was never planned to serve all at once. Around 50 GW of behind-the-meter gas generation was reported announced in a single recent year, and one forecast put on-site generation as the primary source for more than a third of data center facilities by 2030, up from a small fraction a year earlier. Forecasts like these move, so read them as direction, not gospel.

Speed is the metric that matters, not cost. On-site generation is usually not cheaper than buying grid power once you account for the fuel and running a plant. What it buys is a connection date the operator owns instead of one the utility controls. Where the grid can deliver power on a schedule the project accepts, the grid wins on cost. Where it cannot, on-site power is how the project gets built at all, which is the trade most of these decisions come down to.

What is the difference between prime and standby power?

Standby power runs only when the normal source fails. Prime power runs all the time as the normal source. That single distinction reshapes the machine, the rating, the maintenance, the fuel logistics, and the emissions permit, which is why treating a standby genset as a prime plant is the most expensive mistake in this whole subject.

A standby-rated machine is built and rated for limited hours at a varying load with no sustained overload, commonly held around 200 hours a year. It is sized to ride an outage and then shut down. A prime-rated or continuous-rated machine is built for unlimited hours under load, with the cooling, the bearings, the service intervals, and the warranty that match a working duty. The generator sizing guide covers the ISO 8528 standby, prime, and continuous ratings in detail, including the data center continuous rating that sits between them. The point here is that on-site primary generation lives at the prime and continuous end, not the standby end.

Run a standby-rated set as prime power and you wear it out fast, void the rating, and run afoul of the emissions exemption that assumes a true emergency machine. The nameplate kW also differs by rating: the same physical engine carries a lower continuous number than its standby number, because it is allowed to run harder for longer. Spec the rating for the duty the plant will actually serve, confirm it against the manufacturer's published ratings, and do not let a standby number stand in for a prime load.

The on-site generation options

The technologies competing for on-site primary power are gas turbines, reciprocating gas engines, and fuel cells today, with solar and battery storage as a supplement and nuclear small modular reactors talked about for the longer term. Each trades speed, scale, efficiency, emissions, and fuel against the others, and most large campuses end up with a mix rather than one technology, tied together by the microgrid.

Pick by what the site is short on. Where the constraint is raw megawatts fast, large gas turbines and engine plants carry it. Where emissions or efficiency drive the choice, fuel cells move up the list. Where a carbon target matters, solar and storage and clean-power contracts come in alongside. The table is the quick read; the sections that follow take each one apart.

Gas turbines for large prime power

Gas turbines are the workhorse where a campus needs large prime power in a hurry, because a single machine puts out tens to hundreds of megawatts and a few of them can carry a whole site. A simple-cycle turbine burns gas to spin a generator and exhausts the heat, which is fast to install and quick to start but less efficient. A combined-cycle plant captures that exhaust heat to raise steam and drive a second turbine, which lifts efficiency well into the higher range at the cost of more equipment, more water, and a longer build.

The catch is the order book. Turbine demand from the AI buildout has run ahead of supply, with heavy-duty units facing multiyear backlogs, even as makers expand. GE Vernova has said it expects to reach roughly 20 GW of annualized gas turbine output in 2026, with more targeted after, which tells you both how hot the demand is and that the lead time is real. A turbine you cannot take delivery of for years does not solve a schedule problem, so the order date matters as much as the rating.

Gas turbines also need a firm gas supply and an air permit, and both can govern the project. A simple-cycle machine can be permitted and started faster but emits more per megawatt-hour than a combined-cycle plant or a fuel cell, which pulls it into stricter emissions review in many jurisdictions. Confirm the turbine availability, the gas pressure and volume the machine needs, and the permit path early, because any one of them can move the energization date the whole plant exists to protect.

Reciprocating engine generators

Reciprocating engine plants are the modular alternative to a big turbine, and they win where a site wants prime power in smaller increments with strong part-load behavior and fast block-load acceptance. These are large gas or dual-fuel engines, each a few megawatts, ganged in numbers to reach the site load. Because the plant is built from many units, it scales unit by unit, holds efficiency across a wider load range than a single large turbine, and lets you take one engine off for service while the rest carry the load.

Block-load and start speed are where engines shine for a load that cannot wait. Manufacturers cite fast ramps: some engine platforms reach full load in seconds, others in a couple of minutes, with rapid starts measured in minutes rather than the longer spin-up a large turbine needs. That responsiveness matters when the plant has to pick up load fast, ride a transient, or follow a swinging load without dipping out of band. The exact start and load-step numbers come from the engine data sheet, not from memory.

The trade against turbines is footprint and maintenance. A field of engines takes more space, more units to maintain, and more exhaust aftertreatment points than one large machine of the same total output, and engine emissions and noise have their own permit and siting demands. Where the site values modularity, redundancy, and part-load efficiency over the simplicity of a single big machine, the engine plant is the usual answer, which is why engine power plants have surged alongside turbines in this build cycle.

Can data centers use fuel cells?

Data centers can and do use fuel cells for on-site primary power, and the technology reached commercial maturity for baseload data center duty in this cycle. A solid-oxide fuel cell, the Bloom-type SOFC most often deployed, converts natural gas or hydrogen to electricity through an electrochemical reaction rather than combustion, which is why it runs clean and quiet. Around 7.65 billion dollars in fuel cell deals for data centers were reported closed between late 2025 and early 2026, with major operators among the buyers, which is the signal that this moved from pilot to mainstream.

The pull is efficiency, emissions, and modularity at once. SOFC systems deliver electrical efficiency of roughly 60 percent and above, higher than a simple-cycle turbine, and because there is no combustion they emit very little of the criteria pollutants, the NOx and particulate, that drive air permitting for turbines and engines. They are built in modular blocks you add as the load grows, and they run on natural gas now with a path to hydrogen or biogas as that supply develops. For a site fighting an emissions limit or a tight permit, that low-criteria-emissions profile can be the deciding factor.

Fuel cells also carry a speed-to-power advantage that lines up exactly with the queue problem. Modular fuel cell systems have been cited as deployable in under a year, with a speed-to-power edge of roughly nine months to a year over heavy turbine orders stuck in backlog. They still depend on a gas supply, they still emit carbon dioxide when run on natural gas, and they accept load differently than an engine, so they are not a free lunch. But where clean, fast, and modular all matter, the fuel cell has become a serious primary-power option, not a curiosity.

Nuclear and small modular reactors

Nuclear is the long-horizon clean-power option, and over the last few years every major hyperscaler has signed at least one nuclear deal for data center capacity. The arrangements split into two kinds: power purchase agreements for the output of existing or restarted large reactors, and deals with small modular reactor developers for new units to be built. As of recent reporting, the announced projects across these deals committed many gigawatts of nuclear capacity, which is real money and real intent, but most of it is years out.

The headline deals show the shape of it. One operator signed a 20-year power purchase agreement for the entire output of a restarted reactor, with first electrons expected in 2027. Another secured nearly 2 GW from an existing nuclear plant under a multi-decade agreement and is exploring new SMRs at that site. A third signed an SMR fleet deal, with an early unit set to supply a modest first block around the end of the decade. These are contracts and PPAs, often with the reactor co-located near the load, not on-site machines a project commissions like a turbine.

Read nuclear as a hedge and a horizon, not a near-term on-site build. SMRs promise firm, carbon-free, around-the-clock power that fits a flat data center load better than intermittent renewables, which is the appeal. The timelines, the licensing, and the first-of-a-kind cost are the catch, and the first commercial units serving these deals are years away. For a campus that needs power now, the gas, engine, and fuel cell options carry the load while nuclear, if it lands, arrives later. Confirm any nuclear plan against the actual agreement and the developer's schedule, because the announcements run ahead of the steel.

Solar and battery storage on site

On-site or co-located solar paired with battery storage contributes clean power to a campus, but it cannot carry a data center on its own because the load is flat and around the clock while the sun is not. Solar offsets energy during the day and helps a carbon target, and the battery, a large lithium-iron-phosphate BESS in most builds, shifts some of that energy into the evening and smooths the swings. What neither does by itself is provide firm capacity at three in the morning under full IT load.

Storage on a data center site does more than time-shift solar. A battery can ride the seconds-to-minutes gap on a source transition, support a microgrid through a disturbance, supply reactive support, and provide grid services where the utility allows it. The energy-storage design, the sizing of the BESS for ride-through versus energy shifting and the safety systems that go with lithium batteries, is its own subject worth treating separately. The point for on-site generation is that storage is the glue that lets intermittent and firm sources work together, not a primary source on its own.

The honest framing is that solar and storage are a supplement to firm generation on a data center, not a replacement for it. A campus that wants to lean hard on renewables needs either a large amount of storage, a diverse portfolio, or a firm clean source behind the solar to cover the hours it does not produce. Size the solar and the battery for the role they actually play on the site, and keep a firm source, gas, fuel cell, or grid, carrying the baseload the renewables cannot.

What is a microgrid?

A microgrid is a local power system that ties the on-site generation, the storage, and the load together under one control scheme, with the ability to run connected to the utility grid or to disconnect and run on its own. It is what turns a collection of turbines, engines, fuel cells, and batteries into a coordinated plant instead of a yard full of separate machines. The defining feature is that it can island, separating cleanly from the grid and continuing to power the load, and then resynchronize and reconnect when the grid returns.

The pieces of a data center microgrid are the generation sources, the energy storage, the loads, the switchgear that ties them, and the controller that runs the whole thing. The controller is what makes it a microgrid rather than a parallel set of generators. It decides which sources run, how they share load, when to charge or draw the battery, when to island and when to reconnect, and it holds voltage and frequency stable across all of it through every transition. The controls and the energy management system are covered below; here the point is that the controller is the part that has to be engineered and commissioned, not just installed.

For a data center the microgrid is what makes on-site generation deliver the availability the mission demands. A single generator is a single point of failure. A microgrid with multiple sources, storage, and a controller can lose a source, ride a disturbance, and island from a failing grid without dropping the critical load, which is the behavior a Tier-rated site needs from its power. The microgrid is the system that integrates the on-site power; the individual machines are just the parts inside it.

Islanding versus grid-parallel operation

A microgrid runs in one of two modes: islanded, disconnected from the utility and powering the site entirely from its own sources, or grid-parallel, connected to the utility and running alongside it. The mode determines what the controller has to do and what protection has to be in place, and a well-built microgrid moves between the two without dropping the load. Most data center microgrids are designed to do both: parallel with the grid most of the time, island on a grid disturbance, and resynchronize back when it clears.

Islanded operation is the resilience case. The on-site plant holds voltage and frequency on its own, which is harder than running parallel because there is no big grid behind it to set the reference and absorb swings. The generation and storage have to balance the load moment to moment, which is why storage and a fast controller matter, and why the plant has to be sized and tuned to ride load steps and source trips while islanded. A site designed to island can shrug off a grid outage that would force a grid-fed site onto its standby generators.

Grid-parallel operation opens the other uses. Connected to the grid, the site can import to cover a shortfall, export surplus where the agreement allows, and participate in demand-response or grid-services programs that pay a flexible load. Running parallel also means the microgrid has to protect the grid, which is where the interconnection requirements and anti-islanding protection come in, covered below. Pick the modes the site needs, design the transitions between them, and prove them in commissioning, because the value of a microgrid is in the transitions working, not in either mode alone.

The fuel supply and its risk

Most on-site primary generation runs on natural gas, which means the data center has traded a dependency on the electric utility for a dependency on the gas utility. Turbines, engines, and fuel cells all need gas, usually delivered by pipeline at a pressure and volume the machines require, and a gigawatt-scale gas plant needs a serious gas supply that may itself require new pipeline capacity. The gas pipe is a utility too, with its own availability, its own interconnection, and its own constraints, so securing it is part of the same problem as securing the power.

The risk is that a single gas supply is a single point of failure for a site that built its own power to be reliable. An event that takes down the electric grid, a severe cold snap, can also pull gas pressure down right when demand peaks, which has happened in real grid emergencies. A plant fed by one pipe with no alternative is exposed the same way a site on one electric feed is. The fixes are the usual ones: redundant supply paths, firm versus interruptible gas contracts, and on-site fuel storage or a dual-fuel capability that lets the plant fall back to stored diesel.

Treat the fuel supply with the same rigor as the power supply, because for an on-site plant it is the power supply. Confirm the gas pressure and volume the machines need, the firmness of the contract, the redundancy of the pipeline path, and what the plant does if the gas drops. A site that solved the grid queue and then runs on a single interruptible gas line has moved the single point of failure, not removed it.

Emissions and air permitting

Air permitting can govern an on-site plant as hard as the load does, because it decides which machines you are allowed to run and for how many hours. Combustion sources, the turbines and engines, emit NOx, carbon monoxide, particulate, and other regulated pollutants, and a large prime plant running continuously is a major emissions source, not a standby machine. That pulls it into serious permitting: federal New Source Review, Prevention of Significant Deterioration in attainment areas, and in nonattainment areas the stricter lowest-achievable-emission-rate review and the purchase of emission reduction credits that can be expensive or simply unavailable.

The controls and the limits are specific. EPA new-source standards set NOx limits for gas turbines that differ between simple-cycle and combined-cycle machines, and meeting them generally requires selective catalytic reduction with an ammonia or urea reagent for NOx and oxidation catalysts for carbon monoxide, with installed SCR costs commonly cited around 50 to 70 dollars per kilowatt for simple-cycle machines. Many states write tighter best-available-control-technology limits on top of the federal floor. Fuel cells sidestep most of this because they do not combust the fuel, which is a real permitting advantage where criteria emissions drive the review.

The standby exemption that makes backup generators easy does not extend to a prime plant, and conflating the two is a common and costly error. A true emergency engine running very few hours a year qualifies for a lighter emissions tier, but a machine running prime or continuous duty does not, and regulators have been moving to tighten how on-site data center generation is treated. Confirm the engine or turbine tier, the required aftertreatment, the permitted hours, and the major-source thresholds with the local air authority early, because a permit problem found late re-selects the whole plant.

Efficiency and combined heat and power

Combined heat and power, or cogeneration, raises a generator's total efficiency by capturing the waste heat the combustion or the fuel cell rejects and putting it to use, instead of throwing it away up the stack. A simple generator turns a third to half of the fuel's energy into electricity and loses the rest as heat. A CHP setup that uses that heat can push total fuel utilization much higher, which is the efficiency argument for combined-cycle turbines and for fuel cells with heat recovery, where combined efficiency can reach roughly 90 percent when the heat is actually used.

The awkward part for a data center is that the facility is already trying to get rid of heat, not use it. A data center's whole cooling system exists to reject the heat the IT load makes, so there is rarely a large on-site heat demand to soak up a generator's waste heat the way a factory or a district heating loop would. That mismatch is why classic CHP, sized around a building's heating need, does not map cleanly onto a data center, and why the heat-reuse story is often more aspiration than installed reality on these sites.

Where waste heat does find a use, it is in specific arrangements: heating adjacent buildings, feeding a district energy system, driving absorption cooling, or warming greenhouses or other co-located loads. Some sites pursue these for the carbon and efficiency credit. The practical read is to claim the CHP efficiency only when there is a real, year-round heat sink to use the heat, because waste heat with nowhere to go is just waste heat, and a combined-cycle plant earns its efficiency through the second power turbine regardless of whether anyone uses the final reject heat.

What does on-site generation cost compared to the grid?

On-site generation generally costs more per megawatt-hour than buying grid power, once you add the capital for the plant, the fuel, the maintenance, the emissions controls, and the staff to run it. The grid spreads the cost of generation, transmission, and reliability across many customers and decades of infrastructure. A single site building and running its own power plant carries all of that itself, at its own scale, which rarely beats the utility on pure cost. Anyone selling on-site power as the cheaper option is selling the wrong benefit.

What it buys is speed and control, and on an AI project that is worth more than the cost difference. A campus that can be earning revenue years sooner because it did not wait in the queue can justify a higher cost of power against the value of the time. The economics turn on the cost of waiting, not the cost of the kilowatt-hour: if a delay means stranded capital, lost capacity, and a missed window, the premium for on-site power pays for itself. Where there is no urgency, the grid is the cheaper answer and usually the right one.

The numbers are project-specific and move with gas prices, equipment availability, the emissions controls a site needs, and the local power rate, so treat any single figure with suspicion. Run the comparison on the actual site: the capital and operating cost of the on-site plant against the grid rate and the demand and capacity charges, weighed against what the schedule is worth. Hedge every cost and timeline to the project and the utility, because the right answer flips depending on how constrained the grid is and how much the time is worth.

Reliability and redundancy of the on-site plant

An on-site plant becomes the primary source, which means its reliability now carries the mission the grid used to carry, and it needs the same redundancy thinking the rest of the power chain gets. A single generator, however large, is a single point of failure. The plant has to be built so that losing one source, for a fault or for scheduled service, does not drop the critical load, which is the N+1 or 2N logic the generator sizing guide applies to standby plants and that applies just as hard to a prime plant.

The redundancy comes from multiples and from the microgrid. Several engines or fuel cell modules, sized so the load fits on the surviving units when one is down, give the plant N+1 at the source level, and storage covers the seconds during a source transition. The microgrid controller is what makes that redundancy real, shedding nonessential load, dispatching the surviving sources, and holding the bus through the loss. A plant sized so the load only fits on every unit running has redundancy on the drawing and none in practice.

The grid changes role rather than disappearing. A behind-the-meter site can keep the utility connection as backup, running on-site as primary and falling to the grid if the plant fails, or it can run islanded and treat the on-site plant plus storage as the whole resilience story. Either is defensible; what is not defensible is assuming a brand-new on-site plant is more reliable than the grid it replaced without building in the redundancy and proving it. The availability target has to be met across the on-site sources, the storage, and whatever grid backup exists, read as one system.

Grid-parallel interconnection and anti-islanding

Even an on-site plant that makes its own power usually still ties to the grid, and that tie has to meet the utility's interconnection requirements, the same way any distributed generation does. IEEE 1547 is the standard that governs interconnecting distributed energy resources to the grid, setting how the on-site generation has to behave at the connection: the voltage and frequency it must ride through, how it responds to grid disturbances, and how it disconnects when it has to. The utility's own interconnection requirements sit on top of the standard and control the specifics at the point of common coupling.

Anti-islanding protection is the part that protects line workers and the grid. An unintentional island happens when the on-site generation keeps energizing a piece of the utility's system after the utility has been disconnected, which can endanger crews working a supposedly dead line and damage equipment on reconnection. IEEE 1547 and the associated UL 1741 listing require the interconnection to detect that condition and disconnect within a defined window, commonly within two seconds, using a mix of passive detection that watches voltage and frequency and active methods that probe for the grid's presence.

The distinction that trips people up is intentional versus unintentional islanding. A microgrid is designed to island on purpose, cleanly, with the controller managing the separation and the reconnection. Anti-islanding protection is there to prevent the unintentional case, where the generation backfeeds a dead utility line by accident. A data center microgrid has to do both: island intentionally for resilience while still protecting against the unintentional island that endangers the grid. Get the protection scheme and its settings coordinated with the utility, and verify them as set against the coordination study, not as the generic drawing shows.

The carbon angle and 24/7 clean power

On-site generation cuts both ways on carbon. A gas turbine or engine running prime power burns fossil fuel and emits carbon dioxide that a grid-fed site might have avoided through a clean power contract, so building your own gas plant can raise a campus's emissions even as it solves the schedule. Operators with carbon commitments feel that tension directly, which is part of why fuel cells, with their higher efficiency and a path to hydrogen, and the nuclear deals, with firm clean output, get the attention they do.

The cleaner end of the on-site story is the move toward 24/7 carbon-free energy, matching clean supply to demand hour by hour rather than buying enough renewable energy over a year to net out total use. The grid-utility guide covers the power purchase agreements and the carbon-free energy compact behind that shift. On-site, the same goal pushes toward firm clean sources, fuel cells on biogas or hydrogen, storage, and eventually SMRs, that can cover a flat load in every hour, which annual renewable matching never actually did.

Read the carbon claim for what it firms, not the megawatts it names. A site that runs a gas plant as primary power and buys annual renewable credits to offset it is not the same as one that runs firm clean generation in real time, even if both report carbon-free on paper. Where the carbon target is real, the on-site mix has to deliver clean power when the load is actually drawing it, which is a harder and more honest target than an annual average, and the technology choices on site are how it gets met or missed.

Bridge power while waiting for the grid

A common use of on-site generation is as a bridge: temporary prime power that runs the campus now while the permanent grid connection works through the queue and the upgrades. The site builds gas turbines or engines, energizes on its own power, starts earning revenue, and then transitions to grid power, or to a cleaner permanent source, once the utility tie is ready. The bridge plant either stays as backup or peaking, moves to another site, or is sized from the start to be the long-term backup it becomes.

The bridge strategy makes sense because it separates the two timelines that no longer line up: the construction schedule, measured in quarters, and the interconnection schedule, measured in years. Rather than let the building sit dark waiting on the grid, the operator bridges the gap with on-site power and pulls the revenue forward by the length of the queue. The math is the same speed-versus-cost trade as the rest of on-site generation, just with an explicit end date when the grid arrives.

The risk is that the bridge becomes permanent by default. A temporary gas plant stood up fast under a standby or limited permit can run into emissions limits and equipment wear if it ends up running prime power for years because the grid connection slipped again. Plan the bridge as if it might have to last longer than promised: spec the machines and the permit for the duty they may actually serve, and confirm the grid connection timeline before betting the site on a bridge that has to come down on schedule.

Gigawatt AI campuses and behind-the-meter power

The on-site generation surge is driven by the gigawatt-scale AI campuses now being planned, which are loads the grid simply cannot serve on the timeline the projects need. A traditional enterprise hall drew a few megawatts. A large cloud campus draws tens to low hundreds. The AI campuses announced over the last few years reach into the hundreds of megawatts and past a gigawatt on a single site, a load that rivals a city and that no local grid was built to absorb all at once.

At that scale, behind-the-meter generation stops being a workaround and becomes a primary strategy. When a single campus needs more firm power than the regional grid can add in the years the project has, building the power on site is the only path that fits the schedule. This is why the announcements pair the AI campus with a gas plant, a fuel cell deployment, or a nuclear deal in the same breath: the power is no longer assumed to come from the grid, it is part of the project scope from the start.

The reading for anyone working these projects is that power moved to the front of the plan. The chips and the capital are available; the energized megawatts are the constraint. A campus that secures its own power, on a schedule it controls, gets built. One that assumes the grid will be ready because the building is ready waits, and the wait is measured in years. On-site generation is how the gigawatt campus gets powered when the grid cannot keep up, which is the defining condition of this build cycle.

The microgrid controller and the power monitoring system

The microgrid controller is the brain of an on-site plant, and it is the part that decides whether the whole thing works. It dispatches the sources, shares load among them, manages the battery's charge and discharge, runs the islanding and reconnection sequences, and holds voltage and frequency stable through every transition. A plant with good machines and a poor controller behaves like a plant with poor machines, because the failures show up in the transitions the controller is supposed to manage smoothly.

Above the real-time controller sits the monitoring and energy management layer, the EPMS or power management system, that watches the whole electrical plant, trends it, alarms it, and gives the operator the dispatch picture. On a data center it integrates with the building management and the critical-power monitoring so the operators see the on-site generation, the storage, the grid tie, and the load as one system. This is where an operator sees a source drifting, a battery aging, or a fuel issue building before it becomes an outage, which is the difference between running a plant and watching it run.

The lesson from commissioning is that the controls have to be proven, not assumed. The hardest failures on a microgrid are in the sequences: the island that does not separate cleanly, the reconnection that bumps the load, the source trip the controller does not ride. Those only surface under real test, which is why the islanding, the load-step, and the source-fail sequences belong in the commissioning script. A microgrid controller is engineered and commissioned, not configured and forgotten, and the time to find a control gap is on the test, not on the outage.

When on-site generation makes sense

On-site generation makes sense when the grid cannot deliver power on the schedule the project needs, and it makes less sense when the grid can. The clearest case is a grid-constrained market: a strong site stuck behind a multiyear interconnection queue, where on-site power is the only way to energize on time. The next is the speed case: a project where the value of starting years sooner outweighs the higher cost of making your own power. And the third is the scale case: a gigawatt campus the local grid cannot serve as a single firm load regardless of the queue.

It makes less sense where the grid has spare capacity and a near-term connection, where the carbon target rules out a fossil plant and no clean on-site source fits the schedule, or where the site cannot get a gas supply or an air permit on terms that work. In those cases the queue, the cost, and the permitting all favor waiting for the grid, and building a power plant solves a problem the site does not have. The honest read is that on-site generation is a tool for a constrained grid, not a default.

Most large projects now run both tracks: pursue the grid connection and build on-site power, using the on-site plant as the bridge and the backup while the grid catches up. That hedge fits the reality that neither timeline is certain. Decide based on the actual constraint, the grid availability and queue in the specific market, the schedule the project can accept, the fuel and permit the site can get, and the carbon target, rather than on a blanket rule. The constraint, not the trend, should drive the call.

What to document

Write down what the on-site plant is, what each source is rated and permitted to do, and how the microgrid is supposed to behave, because the operator who runs it and the next engineer who touches it both need to know the duty and the modes the plant was built for. An on-site plant whose sources, fuel arrangement, and islanding scheme were clear at turnover and quietly drifted with a later change is the common way a site ends up running a standby-rated machine prime, or islanding into a sequence nobody tested.

Capture the generation sources and their ratings and duty, the microgrid one-line and operating modes, the fuel supply and its redundancy, the air permit and permitted hours, the interconnection and protection settings, the plant redundancy scheme, and the controller and monitoring configuration. Record the assumptions behind each, because the assumptions, the load served, the hours permitted, the gas firmness, are what change when the site does.

Common mistakes

• Running standby-rated gensets as prime power, so the machines wear out fast, void the rating, and lose the emissions exemption.
• Ignoring the fuel-supply risk and feeding the plant from a single interruptible gas pipe, so the site just moved its single point of failure.
• Building the plant with no air-permit or emissions plan, then discovering a prime source does not get the standby exemption and needs full aftertreatment.
• Tying the on-site generation to the grid with no proper interconnection protection or anti-islanding, endangering utility crews and the grid.
• Expecting on-site power to be cheaper than the grid, when the real benefit is speed and schedule control, not the cost per kilowatt-hour.
• Sizing the on-site plant with no redundancy, so a single source fault or a service outage drops the critical load.
• Treating the microgrid controller as configured-and-forgotten instead of engineering and testing the islanding and reconnection sequences.
• Claiming CHP efficiency on a site with no real heat sink, when a data center is trying to reject heat, not use it.
• Standing up a temporary bridge plant under a limited permit and then running it prime for years as the grid connection slips.

Field checklist

Standards and references

Several bodies govern different parts of on-site generation, and naming the right one for the point is the credibility. IEEE 1547 covers interconnecting distributed energy resources to the grid, including the voltage and frequency ride-through and the anti-islanding behavior the on-site plant has to meet at the point of common coupling, with UL 1741 the listing that demonstrates the inverter-based equipment complies. ISO 8528 is the engine-generator rating standard that defines the prime and continuous ratings the on-site machines have to be rated to for continuous duty, covered in the generator sizing guide.

On the installation and safety side, NFPA 70, the National Electrical Code, governs how the on-site generation is wired and classified, and NFPA 110 covers emergency and standby power systems where the plant also serves that role. NFPA 853, the standard for the installation of stationary fuel cell power systems, governs fuel-cell plants specifically. Emissions fall under the EPA Clean Air Act program: the New Source Performance Standards set the federal limits for turbines and engines, New Source Review and Prevention of Significant Deterioration apply to major sources, and the local or state air authority sets the best-available-control-technology limits and the permitted hours. The standby-engine emissions exemption does not extend to a prime plant, and regulators have been tightening how on-site data center generation is treated.

Above all of these sit the utility interconnection requirements, the equipment manufacturer's instructions, and the project basis of design, which set the actual numbers: the ratings, the protection settings, the fuel specifications, and the permit limits. The voltages, the lead times, the costs, and the technology mix in this guide vary by region, by utility, and by year, so confirm them against the specific utility, the adopted code editions, the air authority, and the actual agreements before relying on them. Cite the controlling document by topic, and let the utility requirements and the engineer of record override any rule of thumb.

Units, terms, and abbreviations

On-site generation borrows terms from power plants, the grid, and the data center, and the same word can mean different things in each. Pin the term to the context before acting on it. Power at the campus scale is measured in megawatts and now gigawatts, energy in megawatt-hours, and a generation source carries a rating, prime or continuous for a working source, standby for a backup, that sets how much of its nameplate is usable and for how long.

Behind-the-meter and on-site both describe power made at the site rather than bought from the grid. A microgrid is the local system that ties the sources, storage, and load together and can island or run grid-parallel. SOFC is the solid-oxide fuel cell used for clean on-site power, BESS is the battery energy storage system, and CHP is combined heat and power. Anti-islanding is the protection that stops on-site generation from backfeeding a disconnected utility line.

On-site / behind-the-meter (BTM)

Power generated at the data center, behind the utility meter, instead of bought from the grid

Prime / continuous power

A source rated to run as the normal supply for long or unlimited hours, not just standby backup

Standby power

A source rated to run only on a normal-supply failure, limited hours at a varying load

Microgrid

A local power system of sources, storage, and load that can island or run grid-parallel under one controller

Islanding

Running disconnected from the utility, the on-site plant holding voltage and frequency on its own

Anti-islanding

Protection that disconnects on-site generation from a de-energized utility line, per IEEE 1547 and UL 1741

SOFC

Solid-oxide fuel cell, converting natural gas or hydrogen to electricity without combustion

BESS

Battery energy storage system, used for ride-through, smoothing, and energy shifting

CHP

Combined heat and power, capturing a generator's waste heat to raise total efficiency

SMR

Small modular reactor, the long-horizon firm carbon-free option contracted through PPAs and co-location

Related tools

Calculators and readiness checks for this work

Compare your options

FAQ

People also ask

• How does a data center get power?
• How do you size a data center generator?
• How much fuel does a generator need?
• What megger voltage do I use on busway?
• What is data center power density?