Gas turbine vs reciprocating engine for data centers

Pick the machine that matches the job, not the one with the better brochure.

• Bridge power — you need to energize a campus in months while a grid connection or a permanent plant is still years out. Reciprocating engines and refurbished or mobile aeroderivative packages both work; recips are the shorter procurement path, aeroderivatives the shorter site footprint.
• Prime power — the generation is the facility’s permanent primary source, running most hours of the year. Reciprocating engines under roughly 100 MW; aeroderivative turbines above that, where the number of engines becomes unmanageable.
• Standby — a source that sits idle and picks up load in ten seconds when the utility fails. Neither gas turbines nor lean-burn gas engines do this. That job belongs to a diesel generator set, backed by UPS through the transition.

Most real campuses end up with two rails, not one: a gas prime-power or bridge plant carrying the load, and a separate diesel standby plant plus UPS behind it. Treating them as alternatives is the most common design error.

On simple-cycle fuel efficiency, the reciprocating engine wins outright, and it is not close.

Large lean-burn gas engines reach the high forties. The Wärtsilä 50SG is rated around 48.6% electrical efficiency at about 18.8 MW; the Jenbacher J920 FleXtra reaches roughly 48.7% at about 10.6 MW. That is the highest simple-cycle electrical efficiency available from any combustion prime mover.

Simple-cycle gas turbines land lower. Modern aeroderivatives run roughly 38–42%. Older heavy-frame machines sit in the low-to-mid thirties — a GE Frame 7EA is about 33.4% at ISO base load. Add a steam bottoming cycle and combined-cycle plants reach 55–64%, but that is a different project with a different schedule.

Expressed as heat rate on a lower-heating-value basis, the spread is stark. A turbine running at 40% burns roughly 20% more fuel per MWh than an engine at 48%. Over a plant that runs most hours of the year, that is the dominant lifetime cost line — which is exactly why prime-power projects gravitate to engines and short-duration bridge projects do not care.

Nameplate efficiency is measured at full load. Data centers rarely sit at full load.

A gas turbine’s efficiency falls off sharply below about 80% of rating — it is a single machine with a fixed compressor, and throttling it moves it away from its design point. A reciprocating plant is not one machine. Ten 10 MW engines serving a 70 MW load run seven units at their optimum and shut three off. Efficiency at the plant level stays nearly flat across a wide load range.

That matters more for AI infrastructure than for a traditional data hall, because AI load is not smooth. NVIDIA’s own engineering work on the GB300 NVL72 makes the point directly: thousands of GPUs operate in lockstep on the same computation, so the whole cluster transitions between idle and high-power states together rather than averaging out the way uncorrelated enterprise workloads do. NVIDIA reports that its power-smoothing features — capacitive energy storage, ramp-rate-limited power capping, and power-burn hardware on shutdown — cut peak power demand by about 30% on a Megatron LLM training run and substantially dampened the fluctuation seen at the AC input (NVIDIA, 2025).

Even with rack-level smoothing, the generation plant sees steps. A multi-unit engine plant absorbs them by adding and dropping units. A single large turbine absorbs them by running inefficiently.

These three get conflated constantly. They are different numbers.

• Start time is idle to synchronized and loaded. A large lean-burn gas engine reaches full load in roughly two to five minutes. An aeroderivative turbine takes about ten minutes — the GE Vernova TM2500, for example, reaches full power in around ten minutes. A heavy frame takes ten to thirty. A combined-cycle plant needs thirty minutes to several hours from cold, because the steam cycle has to come up.
• Ramp rate is how fast an already-running machine changes output. Engines ramp faster than turbines and hold their efficiency while doing it, which reduces how much battery or flywheel storage the site needs to smooth the difference.
• Load acceptance is how much load the machine takes in a single step. This is the one that decides whether a design works. Diesel gensets accept large steps and meet NFPA 110 Type 10 — full rated load in one step in under ten seconds. Lean-burn gas engines accept much smaller steps, because a lean air-fuel mixture has little margin before the engine misfires or knocks. Gas turbines sit in between.

The consequence is architectural. A gas plant is not a standby plant. No gas turbine or lean-burn gas engine picks up a data hall from cold in ten seconds. If the gas plant is the primary source, you still need UPS to bridge, and normally a diesel standby tier behind that.

Turbines win the physical and operational arguments.

Power density. GE Vernova puts an aeroderivative at roughly 22 times the output per unit of a reciprocating engine while occupying about 30% less floor space (GE Vernova). Concretely: a 500 MW engine plant means 200-plus engines, each with its own foundation, exhaust, cooling, gas train, lube system, and switchgear cell — and the gas piping, cabling, and land to connect them all. The same 500 MW is a handful of turbines.

Maintenance cadence. GE reports aeroderivatives requiring maintenance about once a year against many times that for reciprocating engines, and quotes fleet availability of 98.2% for aeroderivatives versus 93% for recips. Those are OEM figures and should be read as such, but the underlying physics is real: a turbine has one rotating assembly, an engine has dozens of reciprocating parts per cylinder. Lube-oil consumption differs by roughly two orders of magnitude.

The counter-argument for engines is redundancy shape. Modern gas engines run 40,000–80,000 hours before a major overhaul, and — decisively — service happens one unit at a time while the rest of the plant runs. An N+1 engine plant loses one unit of capacity for a service outage. A single-turbine plant loses everything. Maintenance frequency is worse for engines; maintenance impact is much better.

That trade explains the sizing rule: engines up to the point where unit count becomes an operations burden, turbines beyond it.

Permitting decides more projects than efficiency does.

NOx. Gas turbines with dry low-emissions combustors reach 15–25 ppm NOx without selective catalytic reduction. Lean-burn gas engines generally need an oxidation catalyst, and often SCR with urea, to meet the same permit limits — extra equipment, extra reagent logistics, extra backpressure.

Methane slip. Lean-burn reciprocating engines pass some unburned methane straight through to the stack. GE puts engine slip at close to 150 times turbine levels. Methane is a potent greenhouse gas, so this shows up both in air permits and in corporate emissions reporting — and it partially offsets the engine’s efficiency advantage on a CO2-equivalent basis.

Fuel flexibility. Turbines are the more omnivorous machines: natural gas, LPG, distillate liquid, and increasingly hydrogen blends, with dual-fuel packages able to switch to stored liquid when pipeline gas is curtailed. For a behind-the-meter site where gas supply is the single point of failure, that matters.

The diesel trap. If diesel gensets are in the mix, understand the EPA line. Emergency standby engines are exempt from Tier 4 Final and may be certified at Tier 2 — but that exemption is bought with a hard cap of 100 hours per year of non-emergency running (and 20–50 hours in parts of California). Any non-emergency use — peak shaving, demand response, prime power, or running the site because the grid connection has not arrived — reclassifies the engine and pulls it into Tier 4 Final, which means DPF plus SCR plus DEF and a physically larger package that cannot be bypassed. Designing a "standby" fleet that is actually going to run is the most expensive mistake in behind-the-meter power.

The right machine you cannot buy is the wrong machine.

Reciprocating engine plants typically run about 12 to 18 months from order to commissioning. New aeroderivative turbines are on multi-year lead times, with manufacturer order books booked toward the end of the decade — the reason gas turbines are effectively sold out. Combined cycle is a three-to-five-year project before civil works are counted.

That asymmetry is why phased plans have become standard: engines or used and refurbished turbine packages to energize the site, new turbines added as they arrive, and a permanent plant or grid connection completing the picture years later. Used and surplus aeroderivative and industrial units are available through Pantheon in the turbines catalog, and reciprocating sets in the gensets catalog.

Whichever prime mover you pick, the generation is only half the schedule. Paralleling switchgear, unit substations, and medium-voltage distribution carry their own multi-year lead times, and an energized turbine with nowhere to send its output is not power.