Gas Turbines vs. Reciprocating Engines for Data Center On-Site Power

When a data center plans on-site generation, whether as bridge power or as a permanent behind-the-meter plant, the gas-fired options fall into three families. They burn the same fuel but behave very differently on a site.

• Heavy-frame (industrial) gas turbines are the large machines used in utility peaking and combined-cycle plants. Compared with aeroderivatives they are heavier, larger, move more air and are slower to start.¹⁰
• Aeroderivative gas turbines are adapted from aircraft engines. They are lighter and more compact, and major maintenance can be done by swapping the gas generator, which keeps a plant down for less than three days instead of the multi-week outages typical of heavy-frame units.¹⁰
• Reciprocating engines are large spark-ignited gas engines. DOE notes that multiple engines can be integrated to deliver well over 10 MW in a single plant, and that they can run on natural gas, biogas, renewable natural gas and hydrogen.¹¹

A 30 MW enterprise campus, a 300 MW AI training site in West Texas and a 1 GW campus with a combined-cycle plant next door each point to a different mix, and many projects use more than one. The rest of this guide works through the factors that usually decide it. The broader question of whether to self-generate at all is covered in grid power vs. on-site generation.

Heat rate, the Btu of fuel needed to make one kilowatt-hour, is the clearest way to compare fuel cost. Lower is better. EIA’s tested heat rates for natural gas units show the ranking across the U.S. fleet: combined cycle lowest, internal combustion engines next and simple-cycle gas turbines highest.⁶

Fleet averages include decades-old units, so new equipment does better. Wärtsilä rates its 50SG engine at 50.0% electrical efficiency at 60 Hz, measured at the generator terminals under reference conditions.¹ Power Engineering puts aeroderivatives at about 45% in simple cycle, against 30–35% for heavy-duty industrial turbines.² Heavy-frame turbines close that gap when paired with a steam cycle, which is why combined-cycle plants post the lowest fleet heat rates. A combined-cycle plant is a much larger, slower project, and is usually a utility or co-located plant rather than something built inside a data center fence.

Two points sharpen the comparison. First, on-site plants rarely run at full load. A data center designed for redundancy may operate several units partly loaded, and a plant made of many small engines can shut some units off and keep the rest near their best operating point. Second, a few points of efficiency matter a great deal at data center load factors. A plant running around the clock burns fuel every hour of the year, so heat rate differences compound into large annual fuel bills. Ask vendors for heat rates at the expected site conditions and part loads, not only at nameplate.

Speed matters in two ways: how fast a unit reaches full load from standby, and how well the plant follows rapid swings in demand. Wärtsilä lists the 50SG at under five minutes to full load on a regular start and under two minutes on a fast start, with a hot ramp rate above 100% of load per minute.¹ Aeroderivatives also start in minutes, while heavy-frame machines are slower to start.¹⁰

AI workloads raise the bar. An industry executive quoted by Data Center Frontier described AI training loads that can drop from hundreds of megawatts to near zero within milliseconds, far faster than any generator can follow, and reliability designs are moving to layered architectures that pair fast-responding UPS and batteries with generation.¹² In practice that means the generator choice and the battery storage design go together. A microgrid that absorbs fast transients in batteries lets turbines or engines run steadily.

Unit size also shapes redundancy. Data center plants are usually designed so that the load can be carried with one or more units out of service, the on-site version of N+1. With 18–19 MW engines, one spare unit is a small fraction of a 200 MW plant.¹ With aeroderivatives of about 35 MW each, every spare is larger,¹³ and with a large frame turbine a single spare can mean a big share of plant cost sitting idle.

Site conditions can change the answer more than the brochure numbers do. Gas turbines take in a roughly fixed volume of air, so hot or thin air means less mass flow and less power. Wärtsilä, which sells engines and so has an interest in the comparison, states that for every 10°C above the 15°C ISO rating point, gas turbines lose about 1% of efficiency and 5–10% of output depending on the type.³ It also says the heat rate of a modern lean-burn engine does not start to derate until about 40°C.³ Turbine and engine ratings use different ISO reference conditions, so compare both at the same site temperature and elevation.³

Gas pressure is the other site factor that is easy to miss. Wärtsilä says its engines run at full power on as little as 6 bar (87 psi), while aeroderivative turbines need nearly 50 bar (725 psi) and so must connect to a main pipeline or add compression.⁴ Turbomachinery Magazine puts typical turbine fuel gas pressures at about 200–900 psig.⁷ Compressors add capital cost, power draw, maintenance and another single point of failure. A site served by a low-pressure distribution line may suit engines; a site on a high-pressure interstate line may suit turbines. Our guides to pipeline access and firm gas transportation cover the supply side.

Footprint runs the other way. Turbines produce more megawatts per acre, so an engine plant of the same output needs more land, more stacks and more units to maintain. Both technologies need room for gas conditioning, emissions controls, step-up transformers and setbacks for sound. Engines and turbines have different sound profiles and enclosure options, so noise setbacks belong in the early layout.

A plant that runs continuously is a very different permitting case from emergency backup generators. Turbines and engines fall under different federal rules, and either can push a site into major-source review.

• Turbines: EPA’s final review of the combustion turbine new source performance standards, published January 15, 2026, created Subpart KKKKa for turbines constructed, modified or reconstructed after December 13, 2024.⁸ EPA set combustion controls as the best system of emission reduction (BSER) for NOx for most turbines and added selective catalytic reduction (SCR) for one subcategory, with subcategories based on size, 12-month capacity factor and efficiency, plus a new subcategory for temporary turbines.⁸
• Engines: new stationary spark-ignited gas engines are covered by NSPS Subpart JJJJ. For non-emergency natural gas engines of 500 hp or more manufactured after July 1, 2007, Table 1 sets limits of 2.0 g/hp-hr NOx, 4.0 g/hp-hr CO and 1.0 g/hp-hr VOC, with lower limits for some later engines.⁹ Hazardous air pollutants fall under the RICE NESHAP, Subpart ZZZZ.

Federal standards are a floor. A plant’s total potential emissions decide whether it needs a major-source PSD and Title V permit, and state and local agencies often require tighter controls such as SCR and oxidation catalysts on either technology. Permitting is also where public opposition concentrates. In Memphis, the Shelby County Health Department issued xAI a permit for 15 permanent gas turbines in July 2025.¹⁴ TechCrunch reported that, according to the Southern Environmental Law Center, which opposed the permit, xAI had been operating as many as 35 turbines without permits.¹⁵ Our guide to air permits for data center generators covers thresholds and timelines; confirm the path with the state agency and an air permitting consultant early.

Delivery dates often decide the technology before efficiency does. Verdantix describes the data center power squeeze as cascading from large frame turbines into other types of gas generation, as developers try to match equipment lead times to build schedules of about four to five years in primary markets.⁵ That has pulled engines and aeroderivatives into large deployments.

At OpenAI and Oracle’s Stargate campus in Abilene, Texas, reported plans call for 29 GE Vernova LM2500XPRESS units totaling about 1.015 GW.¹³ In January 2026, Wärtsilä announced an order for 24 of its 50SG engines, 429 MW in all, for a U.S. power plant serving a data center.¹⁶ Verdantix reports Caterpillar supplying 2 GW of G3516 engines to a West Virginia campus, with deliveries from September 2026 to August 2027, and an aftermarket supplier, FTAI Aviation, announcing a $1.5 billion order in July 2026 for refurbished aircraft-engine-based turbines with deliveries through November 2027.⁵

The generator set is only part of the critical path. Step-up transformers, medium-voltage switchgear, gas compression, emissions controls and the gas lateral all have their own lead times, and generator step-up transformers are among the tightest items in the market, as our guide to transformer and switchgear lead times explains. Treat any quoted date as provisional until a manufacturing slot and deposit terms are in writing.

A short list of questions usually narrows the choice quickly:

1. 01What is the load and how will it ramp? Small blocks suit phased campuses; large units suit a big, steady first phase.
2. 02Is the plant bridge power for a few years or permanent? Bridge plants favor fast delivery and redeployable units; permanent plants favor efficiency and a path to combined cycle or grid backup.
3. 03What are the design ambient temperature and elevation? Hot, high sites erode turbine output more than engine output.³
4. 04What gas pressure is available at the fence, and is the supply firm? Low pressure favors engines or adds compression for turbines.⁴
5. 05What will the air agency require, and how much opposition is likely? Emissions controls and permit timing can outweigh the efficiency difference.⁸
6. 06Which units can be delivered, installed and commissioned by the target date, including transformers and switchgear?⁵

The answers belong in the site screen, not after land is under contract, because they bear on acreage, gas routing, setbacks and the behind-the-meter vs. front-of-the-meter structure. When we screen a site for on-site power, we look at the gas supply, ambient conditions and permit path together; if you have a parcel in mind, you can get a site reviewed. For fuel cells as a third option, see our guide to fuel cells for data centers.