Split an LNG plant's nameplate into two identical trains and, on our synthetic bench, 32% of years contain at least one month under 60% of nameplate. Split the same nameplate into seven trains with the same per-train reliability and that share falls to 0.17%, about one year in 600. Both plants average 90.5% of nameplate over a year. A seller who prices a monthly delivery commitment is pricing that tail, and the train count changes the tail while leaving the average where it was.
The question is live on the US Gulf Coast, which now runs both kinds of plant. Sabine Pass has six large trains [3]. Corpus Christi has three large trains and, since the seventh and last Stage 3 train was handed over on 28 August 2026, seven midscale ones, ten in all [1][2][3]. Calcasieu Pass has 18 midscale trains in nine blocks of two [4], and Plaquemines is designed for up to 36 small trains in 18 blocks [5].
Why the mean is the per-train availability
Output on a day is the share of trains running. If each of n trains is up independently with the same probability p, the expected share running is p whatever n is, and its variance falls as 1 over n:
Seven trains instead of two divide the standard deviation by the square root of 3.5, which is 1.87. That is all the reliability argument gives you. It says nothing about p. If the midscale trains are a little less reliable than large ones, the mean drops and the smaller spread does not make it up.
The bench, computed exactly
Every input below is illustrative, and none is fitted to a terminal. Each train has a 20-day turnaround a year, with the trains' turnarounds staggered through the year. It trips four times a year, each trip lasting a lognormal two days on average, and has 0.3 major outages a year lasting 20 days on average. The trains fail independently of each other. Together these give a per-train availability of 90.5%.
The first version of the bench was a Monte Carlo. This one is exact arithmetic: a month's output is the convolution of n independent train distributions, and a year is the convolution of twelve months. To make that possible, an outage starts only on a day its train is running and its days are booked to the month it starts in. That concentrates long outages in one month, so the monthly tail here is somewhat heavier than in the event simulation it replaces, which lets outages run across month ends. Both give the same mean.
| Layout, same nameplate | Mean year | P10 year | P90 year | Std of a year | Years with a month under 60% |
|---|---|---|---|---|---|
| One train | 90.5% | 85.5% | 93.7% | 3.32 points | 100% |
| Two large trains | 90.5% | 87.3% | 93.0% | 2.35 points | 32% |
| Seven midscale trains | 90.5% | 88.8% | 92.0% | 1.25 points | 0.17% |
| Ten midscale trains | 90.5% | 89.1% | 91.8% | 1.05 points | 0.007% |
All figures in the table are model outputs. The standard deviation falls from 2.35 points to 1.25, a ratio of 1.874 against 1.871 for independent units; the staggered turnarounds make the difference. The one-train plant has a month under 60% every year because its own turnaround takes 20 days out of one month. With two trains, that month falls under 60% with five more days of outages on either train, and it does about one year in three. With seven, a turnaround takes under a tenth of one month's output and the plant almost never gets there.
Every input is illustrative and the arithmetic is exact: each year is a convolution of twelve independent months, each month a convolution of 7 trains. Outages start on running days and are booked to the month they start in.
The exhibit opens on seven trains against two. Drag the train marker and the year's distribution narrows and rises while the amber mean line stays where it is. The ladder on the right gives the tail for every count at once.
The advantage lives in the deep tail
A commitment close to the mean is missed by any plant. Drag the commitment in the exhibit to 85% of nameplate and the seven-train plant misses some month in 81% of years, about 15 months a decade, against 26 for the two-train plant. At 80% it is 44% of years. The months that miss are mostly the turnaround months, where one train in seven out for 20 days and a few trips are enough. More trains buy protection against the deep month, the one at 60%, and much less against a shortfall of a few points.
The bench's turnaround is also long. Twenty days a year is 5.5% of the year in scheduled losses, where Malaret's survey of LNG plants puts scheduled losses at 1.4% to 3%, with a mean of 2.1%, and total availability at 89.4% to 95.5%, with a mean of 92.6% [7]. With an 8-day turnaround, 2.2% of the year, the mean rises to 93.6%, and the two-train plant still has a month under 60% in 21% of years against 0.06% for seven. For comparison, Woodside reported 97.8% LNG reliability at its North West Shelf project in the second quarter of 2026 [8]. That is one operator's measure for one project and one quarter, and an operator's reliability figure need not be defined like the availability above.
Less reliable small trains lower the mean
Here the two large trains keep four trips a year and the seven midscale trains are made less reliable. These figures are also model outputs.
| Trips per midscale train a year | Seven-train mean | Seven-train P10 year |
|---|---|---|
| 4, the same as the large trains | 90.5% | 88.8% |
| 6 | 89.2% | 87.5% |
| 8 | 87.9% | 86.1% |
Two extra short trips a year per train cost the fleet 1.3 points of annual output, and its mean falls below the two-train plant's 90.5%. Its P10 year at six trips, 87.5%, is still above the two-train P10 of 87.3%, so the year is still smoother. At eight trips it is not. That is the trade: a modular plant may give up a little on each unit to get a smoother year, and whether it does depends on the per-unit availability of the actual equipment. No public source we found gives that figure.
Shared systems take every train down at once
A power failure, a feed-gas curtailment, a jetty problem or a hurricane takes every train down together, and for those events the plant behaves as one train whatever its count. Freeport LNG halted operations on 7 July 2024 before Hurricane Beryl reached the Texas coast, and restarted one of its three trains first [11]. A blocked export does the same thing from the shipping side, as we worked through for tank room when the ships stop [12].
The exhibit has a control for this. With one plant-wide outage a year, three days on average, the mean falls to 89.5%, and the share of years with a month under 60% rises to 42% for two trains and 2.5% for seven. The seven-train plant's advantage shrinks from about 190 times to about 17. So the bench without plant-wide outages is an upper bound on the modular advantage, and where a real plant sits depends on how many of its lost days were plant-wide. Monthly totals cannot tell a plant-wide outage from a train-level one. That needs daily or train-level records.
What EIA's monthly exports show
This section uses real data. EIA publishes each US terminal's monthly LNG exports to all countries [10]. We divided each month by its days and expressed it as a fraction of that terminal's own 95th-percentile month, a proxy for peak and not a measure of capacity. Nameplate is a definition before it is a number: EIA put the US terminals' 2024 output at 104% of nominal capacity and 86% of peak [9]. The window is April 2023 to July 2026; for Calcasieu Pass it starts in July 2023, after the plant had ramped up. Plaquemines is left out because its record is too short.
| Terminal | Trains | Months | Mean over peak | Spread of monthly volume | P10 month over peak | Months under 60% | Months under 80% |
|---|---|---|---|---|---|---|---|
| Sabine Pass | 6 large | 40 | 0.904 | 0.097 | 0.818 | 0 | 2 |
| Calcasieu Pass | 18 midscale | 37 | 0.886 | 0.104 | 0.785 | 0 | 10 |
| Cameron | 3 | 40 | 0.854 | 0.129 | 0.713 | 1 | 11 |
| Cove Point | 1 | 40 | 0.825 | 0.176 | 0.665 | 4 | 15 |
| Corpus Christi | 3 large, then 7 midscale added | 40 | 0.743 | 0.178 | 0.620 | 1 | 30 |
| Freeport | 3 | 40 | 0.844 | 0.216 | 0.720 | 3 | 10 |
The spread is the standard deviation of daily-average monthly volume divided by its mean. Train counts are from the references [3][4][6].
The terminals do not sort by train count. Sabine Pass, with six large trains, has a slightly tighter spread than Calcasieu Pass with 18 midscale ones, and neither had a month under 60% of its own peak. The single-train Cove Point had four such months, and its lowest months were October 2023, October 2024 and September 2025, which looks like an annual maintenance window, the one-train case on the bench; we have not confirmed the cause. Freeport, with three trains, has the widest spread, and its lowest months are March, April and July 2024, the last of them Beryl's. Cameron's lowest month is May in each of the four years. Corpus Christi's low mean over peak reflects a terminal adding trains through the window, so its own peak is its newest capacity.
The terminals mostly do not dip together. From July 2023, the correlation between one terminal's monthly fraction and another's runs from minus 0.21 to 0.63, with a mean of 0.24. That says nothing about plant-wide outages inside a terminal, which is what the modular case depends on, and monthly data cannot show it. Demand, feed gas, maintenance seasons and plant age all move monthly volume too. This data is a check on direction, and it does not test the bench.
Key takeaways
- With identical independent trains the mean is the per-train availability, 90.5% of nameplate on the bench at any train count, and seven trains instead of two cut the standard deviation of a year's output from 2.35 points to 1.25, a ratio of 1.874 against 1.871 for independent units.
- The share of years with a month under 60% of nameplate falls from 32% with two trains to 0.17% with seven; at an 85% commitment the seven-train plant still misses some month in 81% of years.
- One three-day plant-wide outage a year lifts the seven-train plant's 0.17% to 2.5%, so the bench without shared-system failures is an upper bound on the modular advantage.
- Two more trips a year per midscale train put the seven-train mean below the two-train plant's, and at eight trips its P10 year is lower too.
- In EIA's monthly exports, Sabine Pass's six large trains are as steady as Calcasieu Pass's 18 midscale ones, and the single-train Cove Point and three-train Freeport have the most deep months.
Limitations
The bench is synthetic. Its trip and outage rates, durations and turnaround length are illustrative, the trains are identical and independent, and none of it is fitted to a terminal or an operator's outage log.
Booking every outage to the month it starts in makes the twelve months independent and the arithmetic exact, and it concentrates long outages. The event simulation it replaces, which lets outages straddle month ends, put the share of years with a month under 60% at 27% for two trains and 0.05% for seven over 20,000 simulated years, against 32% and 0.17% here. The annual mean, P10 and P90 agree within a few tenths of a point.
Real turnarounds are usually taken every few years and last longer than 20 days, so a turnaround year is worse than the bench's average year and the years between are better. A turnaround spread evenly over every year understates how often the deep month happens in the turnaround year.
Outages on one train add up rather than overlap, and a train's outages cannot start while it is in turnaround. Both are bookkeeping choices, and we have not measured their effect on the tail separately.
Plant-wide outages are a single rate with a three-day duration. Hurricanes are seasonal and last longer, and feed-gas curtailments can take part of a plant rather than all of it.
The EIA comparison normalises each terminal by its own 95th-percentile month, which removes ramp-up and expansion only in part and flatters a terminal with a stable plateau. Monthly exports also include cargo timing, so a month's volume is not the month's production.
References
- US EIA, Today in Energy. Corpus Christi LNG expansion makes facility the second-largest in the United States. 15 September 2026, for the seven midscale Stage 3 trains and 3.1 Bcf/d nominal capacity: eia.gov.
- Oil & Gas Journal. Cheniere reaches substantial completion at CCL Stage 3. 1 September 2026, for the seventh and last Stage 3 train handed over on 28 August: ogj.com.
- Cheniere Energy Partners. Cheniere announces substantial completion of Train 6 at the Sabine Pass Liquefaction Project. 7 February 2022, for all six Sabine Pass trains and nine trains across Sabine Pass and Corpus Christi: lngir.cheniere.com.
- US EIA, Today in Energy. Calcasieu Pass, the seventh U.S. liquefied natural gas export terminal, begins production. 29 April 2022, for 18 midscale trains in nine blocks of two and 1.3 Bcf/d baseload: eia.gov.
- Oil & Gas Journal. Venture Global signs long-term LNG supply agreements with China Gas. 24 February 2023, for up to 36 0.626 Mtpa trains in 18 two-train blocks at the 20 Mtpa Plaquemines plant: ogj.com.
- US EIA, Today in Energy. U.S. liquefied natural gas exports have increased as new facilities come online. 7 December 2017, for one train at Cove Point and three each at Freeport and Cameron: eia.gov.
- F. J. Malaret (Technip). Understanding RAM analysis and typical availability values for LNG plants. Gas Processing & LNG, April 2015: gasprocessingnews.com.
- Offshore Technology. Woodside's Q2 2026 operating revenue $4.2bn. 30 July 2026, for North West Shelf quarterly LNG reliability of 97.8%: offshore-technology.com.
- US EIA, Today in Energy. The United States remained the world's largest liquefied natural gas exporter in 2024. 27 March 2025: eia.gov.
- US EIA. U.S. Liquefied Natural Gas Exports by Point of Exit, monthly, with each terminal's exports to all countries; release of 30 September 2026, data to July 2026. Public domain: eia.gov.
- Pipeline & Gas Journal. Freeport LNG to restart train following Hurricane Beryl damage. 15 July 2024: pgjonline.com.
- EarthScan. An LNG Plant 80% Full When the Ships Stop Has 1.8 Days of Tank Room. 22 October 2026.
- EarthScan availability bench, synthetic throughout, the source of every simulated figure above; available on request.




