On-Site Power vs Grid Connection for Data Centers | Turbine Economics

By Jackie Jameson · Published April 6, 2026 · 17 min read

By Jackie Jameson, Chief Energy Economist

Financial modeling fact-checked by: Green Gas Turbines Research Team

Last Updated: April 6, 2026

Methodology: This article draws on 2025–2026 data center market reports, U.S. grid interconnection research, OEM data-center power materials, and hydrogen-capable turbine documentation. It distinguishes carefully between temporary bridge power, permanent behind-the-meter prime power, and grid-parallel microgrid architectures.

Executive Summary / Key Takeaways

Power availability is now a commercial constraint, not just a utility issue. In primary data center markets, average grid connection waits now exceed four years, and broader U.S. interconnection data show median project timelines around five years from request to commercial operation.
Behind-the-meter generation is moving from backup to prime-power planning. Developers are increasingly using on-site gas turbines and battery storage not only to bridge grid delays, but to secure permanent operating flexibility.
The core economic question is not “grid or turbine?” in the abstract. It is whether the value of faster energization, avoided delay, and optional spark-spread upside outweighs the added CapEx, fuel risk, and operational complexity of owning generation.
Hydrogen-ready turbines reduce long-term asset-stranding risk. They do not make a project zero-carbon on day one, but they do create a pathway from natural gas today toward higher hydrogen blends or future dedicated hydrogen fuel when regional supply matures.
Capacity planning is unforgiving. An undersized on-site plant risks brownouts and cooling failures; an oversized plant destroys project returns. Turbine count, load steps, redundancy, PUE, and fuel logistics all have to be modeled together.

Introduction

In the AI buildout, compute may be the product, but power is the gating input. A site in a premium market is no longer valuable simply because it has land, fiber, and zoning. If it cannot be energized on a competitive timeline, it can become an expensive option on a market that has already moved on.

That is why the industry is rethinking the old default: wait for the utility, build the shell, and let backup generators handle rare outages. That model still works in some markets. But where interconnection timelines stretch into the late 2020s or beyond, developers are increasingly turning to behind-the-meter generation—not as a temporary emergency workaround, but as a deliberate prime-power strategy.

This is the economic case for on-site generation in 2026: not that every data center should abandon the grid, but that in many AI-driven projects the cost of waiting is becoming larger than the cost of owning power.

The Breaking Point: Why the Grid Is Failing AI

Unprecedented Gigawatt Load Requests

AI is changing the scale of data center development. JLL projects the global data center sector could add roughly 97 GW between 2025 and 2030 and reach about 200 GW of total capacity by 2030. Bloom Energy’s 2026 data center survey goes further, saying campuses are increasingly scaling toward gigawatt-size AI factories, with about one in five campuses expected to exceed 1 GW by 2030.

Utilities were not built to serve dozens of these requests at once—especially where substations, transmission, transformers, and switchgear were already constrained before the AI wave accelerated.

The Cost of Waiting

The economics of a delayed data center are brutal because the project’s capital stack is so large. JLL says average global shell-and-core construction cost reached $10.7 million per MW in 2025 and is forecast at $11.3 million per MW in 2026, while AI fit-out can add as much as $25 million per MW.

That means a 100 MW AI campus can represent up to roughly $3.6 billion of construction and fit-out value before financing, land, and working capital. When a utility connection slips by years, the developer is not just waiting on electrons. They are carrying land, engineering, entitlement, procurement, and capital-allocation risk while revenue is deferred.

This is the key strategic shift: in some AI projects, the opportunity cost of waiting for the grid can rival or exceed the premium of on-site generation.

The Economics of Behind-the-Meter Generation

What “Behind-the-Meter” Means at Hyperscale

At hyperscale, behind-the-meter (BTM) does not mean a few standby engines hidden behind a substation fence. It means generation and storage located on the customer side of the utility meter and designed to serve the facility directly—sometimes as bridge power, sometimes as permanent prime power, and sometimes as a hybrid architecture that starts islanded and later becomes grid-parallel.

Bloom’s 2026 survey found expectations for fully on-site-powered data centers increased sharply over the prior six months, and that more than one-third of data centers are expected to use 100% on-site power by 2030.

Time-to-Market Comparison

Power Path	Indicative Time to Power	Main Bottlenecks	Commercial Implication
Traditional utility interconnection	Often 4+ years in primary markets; U.S. queue data show a median of about 5 years from request to commercial operation for built projects	Queue studies, transformer lead times, transmission and substation upgrades, permitting	Lowest direct generation CapEx, but often the highest schedule risk
Fast-track modular on-site power	Project-dependent, but can be much faster than utility-led upgrades; some modular mobile packages are available within months	Fuel availability, permits, emissions controls, temporary or modular balance of plant	Strongest option for speed-to-initial-energization
Permanent behind-the-meter prime power	Project-dependent; typically engineered in parallel with the campus rather than after the utility queue clears	Gas interconnect, EPC scope, cooling, redundancy design, air permitting, O&M setup	Higher upfront CapEx, but much greater control over schedule and long-term operating strategy

Important nuance: the table above is why the market is shifting. Utility interconnection is often cheaper on a narrow electrical-supply basis, but on-site power can be dramatically better on a time-to-revenue basis.

Capital Expenditure (CapEx)

The upfront cost of BTM generation is real and should not be minimized. A serious prime-power plant usually includes:

Gas turbines or aeroderivative turbine packages
Fuel gas interconnection, metering, and compression as required
Electrical balance of plant, switchgear, transformers, controls, and protection systems
Emissions control systems, noise mitigation, and permitting support
Battery storage and/or UPS bridging layers
Cooling and water systems where required

But the “grid-only” option is not free either. Developers may still face utility contribution charges, private substations, land tied up while waiting, and later the cost of backup generation anyway.

Operational Expenditure (OpEx) & the Spark Spread

The spark spread is the classic starting point for evaluating gas-fired generation. EIA defines it as the difference between the electricity price and the cost of the natural gas needed to produce that electricity.

Formula:

Spark spread ($/MWh) = power price ($/MWh) – [natural gas price ($/mmBtu) × heat rate (mmBtu/MWh)]

Simple example: if power costs $90/MWh, gas costs $4/mmBtu, and the plant heat rate is 7 mmBtu/MWh, then the gross spark spread is:

$90 – ($4 × 7) = $62/MWh

That is only a starting point. EIA is explicit that spark spread is not a full profit metric because it excludes other variable and fixed costs. For a data center microgrid, a real project model also needs to include:

Variable O&M
Pipeline transport and fuel delivery charges
Maintenance reserves and outage planning
Demand-charge avoidance or standby charges
Carbon costs or emissions compliance
BESS and UPS amortization
The financial value of faster energization

That last item is why many AI projects move forward with BTM even when pure fuel-to-power cost parity is not perfect.

Future-Proofing the Investment: From Natural Gas to Hydrogen

Avoiding Stranded Assets

The central criticism of on-site gas generation is obvious: what if the facility wins on speed but loses on long-term carbon policy, corporate emissions targets, or future fuel rules?

That concern is real. It is also why many operators are not asking for “cheap gas turbines,” but for hydrogen-ready gas turbines that can secure near-term speed to power without freezing the site into a single-fuel architecture.

The Hydrogen-Ready Turbine Advantage

Hydrogen-ready does not mean carbon-free on day one. It means the turbine package and plant layout are configured so the site can transition later. GE Vernova says its gas turbine fleet has experience operating on hydrogen contents from 5% by volume up to 100% depending on model and configuration, but also notes that hydrogen retrofits can require modifications to fuel systems, fire systems, controls, package hardware, hazardous-area equipment, and purge systems.

That is the correct investment lens: a hydrogen-ready plant is not simply “buy a turbine and flip to hydrogen later.” It is a way to preserve optionality while using natural gas to solve the immediate problem of site energization.

Designing and Sizing Your On-Site Generation Facility

Matching Load Steps to Turbine Spin-Up

This is where many non-specialists underestimate the problem. AI loads are not only large; they are dynamic. GPU clusters, liquid-cooling equipment, and supporting electrical infrastructure create step changes that a standalone turbine fleet should not be expected to absorb instantly without support.

That is why serious BTM architectures typically pair turbines with UPS and BESS layers. Siemens’ data center power paper notes that a battery system can respond in under 1 second, hot-standby redundant units can reach full load in under 10 seconds, and cold-standby systems may require up to around 15 minutes to bridge, depending on configuration.

The right architecture depends on whether the plant is running as:

Prime power with turbines already synchronized
N+1 redundancy with one or more units in hot standby
Bridge power with batteries carrying the site until turbines stabilize

An islanded site can be made extremely robust, but only if the redundancy philosophy matches the real load profile.

Prime Power vs Combined Cycle vs Modular Aeroderivatives

There is no one “best” turbine architecture for data centers. The right choice depends on load factor, climate, redundancy philosophy, noise limits, water constraints, and whether the plant will eventually run grid-parallel.

Aeroderivatives are attractive where fast starts, modularity, and phased growth matter most.
Combined-cycle plants can deliver stronger efficiency where a stable, high-load prime-power case exists and the site can justify the added complexity.
Modular multi-unit fleets often make the most sense for AI campuses because they allow staged buildout and better maintenance flexibility.

Siemens explicitly models scenarios where multiple smaller gas turbines plus a redundant unit are used to power data centers with high availability requirements, and where optional battery storage reduces the burden on UPS sizing.

Calculate Your Microgrid Requirements

When bypassing the grid, there is zero room for error in capacity planning. An undersized turbine fleet leads to brownouts or cooling instability, while an oversized fleet can destroy ROI. To precisely engineer your prime-power requirements—factoring in compute loads, cooling overhead (PUE), and desired redundancy—use our Data Center Power Architecture Sizer. The tool helps project developers visualize the turbine capacity required for an islanded or behind-the-meter facility.

Grid-Parallel Operations: The Best of Both Worlds

Why “On-Site” Does Not Have to Mean “Off-Grid Forever”

One of the most underappreciated advantages of BTM generation is optionality. A project can begin life as an islanded or partially islanded site to bypass grid delays, then later become a grid-parallel microgrid once utility infrastructure catches up.

Siemens notes that even in on-site generation scenarios, a grid connection may still be planned in case of power surplus or other operational needs. That means a data center owner does not necessarily have to choose one model forever.

Peak Shaving, Self-Supply, and Limited Market Participation

Once grid-connected, the economics change again. The owner may choose to:

Self-supply during peak-price periods
Use the turbines to reduce demand charges or grid purchases
Hold the fleet primarily for resilience
Export excess power or participate in services where interconnection agreements and market rules allow

This is why the right financial model is not a one-time LCOE comparison. It is a phased business case: bridge power, prime power, resilience value, and later grid-parallel optionality all need to be modeled together.

Conclusion

The 2026 verdict is not that the grid is obsolete. It is that in many AI-driven projects, waiting passively for utility power is no longer the economically neutral choice.

Where grid timelines stretch beyond competitive development windows, behind-the-meter generation with gas turbines can create a decisive speed-to-market advantage. The economics become even stronger when the plant is designed intelligently: staged modular capacity, UPS and BESS for load steps, and hydrogen-ready hardware to preserve future fuel flexibility.

The developers who win this cycle will not be the ones who merely secure land. They will be the ones who secure credible time to power.

Use the Data Center Power Architecture Sizer to model your on-site generation requirements, or contact Green Gas Turbines to review whether a grid-only, hybrid, or fully behind-the-meter architecture makes sense for your site.

Frequently Asked Questions

What is an on-site power data center?

It is a data center that generates some or all of its electricity locally—often with gas turbines, engines, batteries, or a microgrid—rather than relying only on utility supply.

What does behind-the-meter generation mean?

Behind-the-meter generation is power produced on the customer side of the utility meter and consumed directly by the site, rather than delivered solely through the public grid.

Are gas turbines cost-effective for data centers?

Often yes, but not automatically. The case depends on fuel cost, heat rate, O&M, permitting, and—most importantly—the economic value of faster energization versus waiting on the grid.

How long does grid interconnection usually take for large data centers?

It varies by market, but current research points to multi-year waits. JLL says primary-market grid connections now average more than four years, and broader U.S. queue data show about five years for many built projects.

Do on-site gas turbines eliminate the need for batteries or UPS?

No. In most serious designs, turbines are paired with UPS and often BESS to manage fast load steps, power quality, and turbine startup or fault-response intervals.

Can a behind-the-meter plant later connect to the grid?

Yes. Many architectures can evolve into grid-parallel operation later, allowing the site to self-supply, peak shave, or use the grid as backup once interconnection becomes available.

Can on-site turbines later run on hydrogen?

Some modern turbines can, but the answer is model-specific. Hydrogen-ready operation may require upgrades to fuel systems, controls, fire protection, purging, and balance-of-plant equipment.