On-site generation: gas turbines, fuel cells and storage

Data centres have always had on-site generation, but traditionally only diesel standby generators that run a few hours a year (DCE-340). Since 2024, with grid connections taking years, many developers have turned to prime power: generation that runs continuously to supply the campus, either until the grid connection arrives ("bridging") or permanently. Natural gas turbines and reciprocating engines, solid oxide fuel cells and large battery systems are now being installed at hundreds of megawatts per site. This lesson compares these technologies, explains how on-site plants are designed for data centre reliability, and examines their costs, emissions and regulatory treatment.

Configurations

On-site generation can relate to the grid in several ways:

• Standby only: grid supplies the load; generators run during outages and tests (the traditional model).
• Grid-parallel prime power: on-site generation runs continuously alongside a grid connection, which provides backup, absorbs surplus and supplies shortfalls.
• Bridging: on-site generation supplies the load (islanded or with a small grid connection) until full grid capacity arrives, after which it becomes backup or is relocated.
• Islanded (off-grid): the campus has no grid connection, so the on-site plant must provide all energy, reserves, frequency control and redundancy.
• Co-location: the data centre is built next to an existing or new power plant and connects to it directly, with or without a grid connection (also discussed in Lesson 10 for nuclear plants).

Gas turbines

Gas turbines (combustion turbines) compress air, burn fuel in it and expand the hot gases through a turbine that drives a generator. Three main types are used:

• Aeroderivative turbines (about 20 to 70 MW), derived from aircraft engines: light, fast-starting (under 10 minutes), good at following load, with simple-cycle efficiency of 38 to 42%.
• Industrial frame (heavy-duty) turbines (50 to over 500 MW): rugged and efficient at large size, but slower to start and best run at steady output.
• Combined cycle: the hot exhaust from one or more turbines makes steam in a heat recovery steam generator (HRSG) that drives a steam turbine, reaching 55 to 63% efficiency. This is the most efficient fossil option but takes longer to build and start.

Efficiency is often stated as heat rate, the fuel energy per unit of electricity. In Canada heat rates are usually given in GJ/MWh: 100% efficiency corresponds to 3.6 GJ/MWh, so a 40% efficient turbine has a heat rate of 9.0 GJ/MWh and a 60% combined cycle 6.0 GJ/MWh. Natural gas is priced per GJ, with the Alberta AECO hub as the main Canadian benchmark.

Turbine output falls in hot weather because air is less dense: an aeroderivative may lose 15 to 25% of its rated output between 15 °C and 40 °C, unless inlet cooling (evaporative or chilling) is used. On-site plants must be sized for the hottest conditions, which are also when data centre cooling loads peak.

Turbine supply became a major constraint in 2025: with orders surging from utilities and data centres, the three main manufacturers of large frame turbines reported order books extending to around 2030. This led developers to use smaller aeroderivative and industrial turbines, reciprocating engines, fuel cells and even refurbished or mobile units.

Reciprocating engines

Large gas reciprocating engines (about 4 to 20 MW each) are similar to the diesel engines used for standby but run on natural gas for continuous duty. They reach 44 to 50% efficiency, start in a few minutes, hold efficiency at part load, lose little output in hot weather and come in modular sizes that allow many units for redundancy. They have higher maintenance costs and emit more NOx and methane slip than turbines unless controlled. Plants of 100 to 400 MW with dozens of engines now serve data centres.

Fuel cells

Fuel cells convert fuel to electricity electrochemically without combustion. Solid oxide fuel cells (SOFCs) operate at 600 to 800 °C, reform natural gas internally, and achieve 50 to 60% electrical efficiency. They emit almost no NOx or particulates (because there is no flame), are quiet, and are built in modular units of a few hundred kilowatts that can be installed quickly in arrays of tens to hundreds of megawatts. Their CO₂ emissions per MWh are similar to or slightly lower than combined cycle, and they can run on hydrogen or biogas. Data centre operators and utilities have signed agreements for hundreds of megawatts to more than 1 GW of SOFCs. Disadvantages include higher capital cost per kW, stack replacement every few years, and limited ability to follow rapid load changes without batteries.

Comparing technologies

Designing for data centre reliability

A grid connection backed by standby generators is very reliable. An islanded on-site plant must achieve similar reliability on its own, which requires redundancy: more generating units than needed to carry the load, so that some can be out for maintenance or fail without interrupting supply. With \(n\) identical units, each available with probability \(a\), the probability that at least \(k\) are available is given by the binomial distribution:

from math import comb

def p_short(n, k, a):
    # probability that fewer than k of n units are available
    return sum(comb(n, i) * a**i * (1 - a)**(n - i) for i in range(k))

load, unit, a = 100, 20, 0.98
k = -(-load // unit)                      # units needed, rounded up
print("units needed:", k)
for n in range(k, k + 4):
    p = p_short(n, k, a)
    print(f"{n} units (N+{n-k}): shortfall {p:.5f} = {p*8760:7.1f} h/yr")

# with one unit on planned maintenance, does N+2 still cover a forced outage?
print("N+2 with one unit in maintenance:", round(p_short(k + 1, k, a) * 8760, 1), "h/yr")

units needed: 5
5 units (N+0): shortfall 0.09608 =   841.7 h/yr
6 units (N+1): shortfall 0.00569 =    49.8 h/yr
7 units (N+2): shortfall 0.00026 =     2.3 h/yr
8 units (N+3): shortfall 0.00001 =     0.1 h/yr
N+2 with one unit in maintenance: 49.8 h/yr

While one unit is down for planned maintenance, an N+2 plant is effectively N+1 and its exposure rises to about 50 hours a year, so maintenance is scheduled for periods when batteries, standby diesels or a partial grid connection can cover the gap, or the plant is built N+3.

Beyond unit count, islanded and grid-parallel plants must handle:

• Frequency and voltage control: without the grid, the on-site plant must hold frequency and voltage as loads change. AI training clusters can swing tens of megawatts within a second as thousands of GPUs start and stop computation together; turbines and engines cannot follow such steps alone, so batteries, supercapacitors or flywheels, and software that smooths GPU power draw, are used.
• Fuel supply: a single gas pipeline is a common-mode failure. Firm transportation contracts, a second pipeline, on-site liquefied natural gas or dual-fuel units with stored diesel provide backup. Pipeline capacity, not just gas price, often limits large on-site plants.
• Maintenance: major overhauls take units out for days or weeks, so redundancy must cover planned and forced outages together.
• Protection and grounding: the plant must provide fault current to operate protection and must coordinate with the data centre's electrical system (DCE-340).

Battery energy storage

Large battery energy storage systems (BESS), usually lithium iron phosphate, serve several roles at data centres:

• Fast response: covering the seconds to minutes before generators start, and absorbing rapid load swings.
• Peak shaving: reducing grid demand charges and peak-based allocations (Lesson 1).
• Grid services: providing frequency regulation or demand response for revenue when not needed on site.
• Shifting renewable output: improving hourly matching (Lesson 5).
• Enabling flexible connections: supplying load during grid curtailment periods (Lesson 3).

Emissions and permitting

Natural gas emits about 50 kg CO₂ per GJ (53 kg per MMBtu) of fuel. A combined-cycle plant emits about 0.33 to 0.38 t CO₂/MWh, a simple-cycle turbine or engine plant about 0.40 to 0.55 t/MWh. Methane leaks along the gas supply chain add to the climate impact. Large on-site plants therefore conflict with operators' climate commitments unless paired with carbon capture, renewable natural gas or hydrogen, or treated as a bridge to cleaner supply.

In Canada, on-site generation needs provincial environmental approvals for air emissions (for example an Environmental Compliance Approval in Ontario or an EPEA approval in Alberta), with federal Multi-Sector Air Pollutants Regulations limiting NOx from new gas engines and boilers and national guidelines for turbines. Large emitters are covered by industrial carbon pricing: Alberta's TIER system and Ontario's Emissions Performance Standards charge for emissions above a benchmark intensity. The federal Clean Electricity Regulations limit emissions from fossil units of 25 MW or more connected to the regulated grid from 2035, which affects the long-term use of gas plants connected to the grid. Alberta's 2026 Data Centre Regulation encourages self-supplied generation, and Alberta and other provinces are promoting gas generation with carbon capture.

Behind-the-meter and co-location rules

When on-site or co-located generation serves a load that is also connected to the grid, questions arise about who pays for the grid services the load still uses (backup, frequency regulation, black start) and whether the arrangement shifts costs to other customers.

In Canada, the rules are provincial. Alberta has long allowed industrial self-generation and "behind-the-fence" arrangements at oil sands and industrial sites, and its data centre rules build on this. In Ontario, embedded generation and load displacement arrangements are governed by the Transmission System Code and IESO market rules, with Global Adjustment and transmission charges structured around net load. In Crown utility provinces, self-generation generally requires utility and government agreement.

In all jurisdictions, developers must decide early whether on-site generation is temporary or permanent, because that choice affects equipment selection (mobile or permanent units), permits, fuel contracts, financing and how the plant will be used once the grid arrives: as backup, for grid services, or relocated to another site.

Lesson 8 covers the gas supply, transport and price risk behind on-site generation, and Lesson 11 how to compare its economics with grid supply and the value of time to power.