15 engines. 282 MW. 30 seconds to sync, under 2 minutes to full load.
On October 5, 2026, Wärtsilä said it will supply a 282 MW onsite power plant built from 15 Wärtsilä 50SG gas engines for a US data center developed by an unnamed independent power producer. Delivery is planned for 2028 and full operation for 2029. It is the company's seventh US data center order, pushing its total for that market past 3 GW. (Wärtsilä press release)
When the grid connection is years away, AI campuses are building their own power plants. But "282 MW of engines" doesn't mean "282 MW of firm power for GPUs." The real answer comes from four pieces of math: how many engines you can lose, how much gas you burn, how fast the fleet can follow a GPU load step, and how you run a plant whose load is never steady.
1. Fleet math: 15 × 18.8 MW, and what N+1 really leaves you
Wärtsilä rates one 18V50SG genset at 18,875 kW at 60 Hz (18,434 kW at 50 Hz), at the generator terminals under reference conditions of 25 °C intake air and 100 kPa (Wärtsilä 50SG data).
P_fleet = 15 × 18.875 MW = 283.1 MW (nameplate)
282 MW ÷ 15 = 18.8 MW per engine (the order's figure, which suggests a small site derate)
N+1 (one unit down): 14 × 18.8 = 263 MW
N+2 (two units down): 13 × 18.8 = 244 MW
Engines need scheduled maintenance, so a data center owner plans around N+1 or N+2, not nameplate. Losing two units costs about 38 MW, roughly 13% of the plant. That's the real price of the "modular" pitch: small units make each outage small, but you still have to pay for spare units.
2. Heat rate and fuel: compare on the same basis
The 50SG datasheet lists a heat rate of 7,207 kJ/kWh at 60 Hz, which is 50.0% electrical efficiency (Wärtsilä 50SG leaflet, PDF). Converted to US units:
HR = 7,207 kJ/kWh ÷ 1.05506 kJ/Btu ≈ 6,831 Btu/kWh
Fuel at full load = 282,000 kWh/h × 7,207 kJ/kWh ≈ 2,032 GJ/h ≈ 1,926 MMBtu/h
Per day at full load ≈ 46,200 MMBtu (on the datasheet's basis)
Here's the trap. The US EIA reports that the 2024 average operating heat rate of the US natural-gas power fleet was 7,754 Btu/kWh (EIA Electric Power Annual, Table 8.1). It looks like the engine is about 12% better, but EIA figures are normally on a higher-heating-value (HHV) basis, while European engine datasheets usually quote lower heating value (LHV). For natural gas, HHV/LHV is about 1.108:
6,831 Btu/kWh (LHV) × 1.108 ≈ 7,570 Btu/kWh (HHV)
→ about 2–3% better than the 2024 US gas-fleet average, not 12%
Fuel (HHV) ≈ 2,130 MMBtu/h ≈ 51,000 MMBtu/day ≈ 49 MMcf/day at ~1,037 Btu/scf
So a 282 MW engine plant running flat out needs a gas lateral sized for roughly 50 million cubic feet per day. The pipeline, not the engines, is often the long-lead item.
3. Ramp rate vs a GPU load step: E = ΔP² / 2R
Wärtsilä quotes a hot ramp rate of more than 100% load per minute, 30 seconds to connect, and under 2 minutes to full load on a fast start (Wärtsilä 50SG data). That assumes hot standby, with cooling water preheated and bearings prelubricated. An older Wärtsilä paper gives about 12 minutes to full load for a 50SG plant from a cold start (Wärtsilä, "Defining true flexibility", PDF).
That's fast for a power plant and slow for a GPU cluster. Here's an illustrative case:
Per-engine ramp R₁ ≈ 18.8 MW/min ≈ 0.31 MW/s
15 engines online: R ≈ 4.7 MW/s
Illustrative training-load step ΔP = 50 MW in about 1 s
Time for engines to catch up: t = ΔP / R ≈ 10.6 s
Energy a buffer must supply: E = ΔP² / (2R) = 50² / (2 × 4.7) ≈ 265 MJ ≈ 74 kWh
74 kWh is small in battery terms, but it has to come out at tens of megawatts within a fraction of a second. That's why engine-powered campuses still pair the plant with a BESS, UPS or supercapacitor layer. The engines carry the energy, and the buffer handles the fast changes.
Key point
An onsite engine plant has to be sized for three different numbers. Firm capacity after outages (N+2 ≈ 244 MW here), fuel delivery (≈ 50 MMcf/day), and the energy needed to bridge a load step (E = ΔP²/2R). Nameplate MW answers none of them.
4. Availability: why 15 small units beat 3 big ones, with numbers
If each engine is independently available with probability A, the chance that at least k of n engines are running is a binomial sum:
P(≥k of n) = Σi=k..n C(n,i) · Aⁱ · (1−A)ⁿ⁻ⁱ
Illustrative A = 0.95: P(all 15) = 46.3%, P(≥14) = 82.9%, P(≥13) = 96.4%
Illustrative A = 0.97: P(all 15) = 63.3%, P(≥14) = 92.7%, P(≥13) = 99.1%
Even with good engines, you'll rarely have all 15 running. Planning around 13 units (about 244 MW) is what gets you near 99%. The modular design also helps at part load. Wärtsilä lists 10% minimum load per unit but 1% at plant level, because you can shut engines down instead of throttling them all (Wärtsilä 50SG data). A 150 MW night load can run on 9 engines at about 89% instead of 15 at about 53%, which keeps each running unit near its best-efficiency point.
To verify
• The 50SG heat rate is assumed to be on an LHV basis (the leaflet specifies gas LHV, but the heat-rate basis isn't stated explicitly), and HHV/LHV ≈ 1.108 is a typical pipeline-gas value. Both still need checking.
• The 50 MW / 1 s load step and the A = 0.95–0.97 availability values are illustrative, not project data.
• The customer, site, grid tie and any BESS on this project haven't been disclosed. The per-minute ramp spec also doesn't describe sub-second governor step response.
What to take away
Onsite engines solve the schedule problem, with delivery in 2028 instead of waiting years in an interconnection queue. They don't remove the engineering problems, though. They move them to gas supply, redundancy margin and fast-transient buffering. When you see the next "X MW onsite" headline, divide by unit size, subtract two units, convert the heat rate to HHV and ask what handles the first ten seconds of a load step.
For a different way to take an AI campus off the grid, see our video on fuel cells and solid-state transformers: Ditching the Grid! Delta's SOFC Secret & SST for Perfect AI Microgrids.
Sources: Wärtsilä press release (Oct 5, 2026) · Wärtsilä 50SG technical data · Wärtsilä 50SG leaflet (PDF) · Wärtsilä flexibility paper (PDF) · EIA Table 8.1 heat rates
