The data-center power story is often told as a choice between the grid and a new nuclear plant. In practice, the nearer-term choice is frequently more ordinary: bring a gas turbine onto the campus while the utility connection catches up.

That option is gaining attention because a large AI campus can be ready to fill racks before a transmission upgrade is complete. The turbine is not a magic source of cheap electricity. It is a project-management tool that converts a fuel connection and a permitted generator into usable megawatts. The trade is that the operator takes on fuel-price risk, local air permitting, equipment maintenance, and a more complicated relationship with the grid.

What a campus turbine actually does

A behind-the-meter turbine generates electricity on the customer side of the utility meter. The data center can run it as its primary supply, use it during constrained hours, or combine it with grid service and batteries. The arrangement can also be structured as a microgrid, with controls that isolate the campus during a utility outage.

A simple-cycle combustion turbine is attractive when speed and flexibility matter. It can start relatively quickly and is useful as a peaking resource, but its fuel efficiency is weaker than a combined-cycle plant. A combined-cycle plant adds a heat-recovery steam generator and steam turbine. That produces more electricity from the same gas, but it adds cost, equipment, water or air-cooling decisions, and construction time.

The distinction matters because a data center has a high and unusually steady load. A peaker built for a few hundred hours per year may be a poor fit if it must run around the clock. Conversely, a fast-start turbine can be valuable as a bridge while the permanent supply is built.

The Energy Information Administration reported in June 2025 that developers planned 18.7 gigawatts of new U.S. combined-cycle capacity through 2028, including 4.3 gigawatts already under construction. EIA also noted that more than half of the 3.3 gigawatts planned for 2026 was already under construction, while much of the 2027 pipeline was not. Those figures describe utility-scale projects, not campus generators, but they show the difference between a named project and equipment that is moving toward operation.

The first constraint is often the gas system

A turbine needs a firm fuel supply, not just a nearby pipeline on a map. The project owner must confirm pressure, delivery capacity, interconnection terms, curtailment rules, and whether the pipeline can support the expected hourly profile during cold weather.

This is especially important in New England and other regions where gas demand rises for heating. The power plant may be able to run in normal conditions and still face fuel constraints on the coldest days. A data center that promises nonstop service cannot treat gas as an assumed input. It needs a supply contract, backup fuel policy where permitted, and a plan for what happens if the pipeline operator curtails deliveries.

Fuel cost is another variable. EIA reported that Henry Hub gas averaged $3.52 per million British thermal units in 2025, 56% above 2024. Northeast hub prices averaged roughly twice 2024 levels. A turbine with a low capital cost can therefore become expensive to run when local gas prices spike.

Efficiency is a financial variable, not a footnote

The heat rate measures how much fuel a generator needs to produce one kilowatt-hour. EIA has reported that newer combined-cycle plants generally have heat rates below 7,000 Btu per kilowatt-hour, while simple-cycle and steam units are typically above 10,000 Btu per kilowatt-hour.

Consider a simplified example. A 100-megawatt turbine running at a 7,000 Btu per kilowatt-hour heat rate consumes about 700 million Btu per hour at full output. At $4 per million Btu, the fuel cost is about $2,800 per hour, before operations, maintenance, emissions costs, and capital recovery. At a 10,000 Btu heat rate, the same output uses about 1,000 million Btu per hour, raising the fuel bill to $4,000 per hour.

These are illustrative calculations, not a forecast. They show why the nameplate megawatt number is not enough. The financial model needs hourly dispatch, expected load factor, heat rate at partial load, start costs, maintenance intervals, fuel transportation, emissions costs, and the cost of purchased grid power.

Permits and community acceptance can set the schedule

A campus turbine is still a power plant. It may need air permits for nitrogen oxides, carbon monoxide, particulate matter, greenhouse gases, and startup or emergency operation. Noise, visible exhaust, water use, and emergency fuel storage can all affect local review.

The fast path is not always the cheapest path. A permit that limits operating hours may undermine the reason for installing the turbine. A requirement for selective catalytic reduction can add cost and water or reagent logistics. Diesel backup generators may exist for reliability but are not automatically an acceptable substitute for a continuous primary supply.

Operators should publish a clear operating envelope: expected annual hours, normal dispatch, emergency dispatch, fuel type, emissions controls, and how the project will respond to grid emergencies. That turns a technical project into a reviewable public commitment.

The prudent model is usually hybrid

The strongest case for campus turbines is often not grid replacement. It is staged power.

Phase one uses temporary or bridge generation to energize an initial building. Phase two adds utility service and perhaps battery storage. Phase three moves the site toward a lower-carbon mix through renewable contracts, firm clean power, or a future nuclear or geothermal resource.

This structure also protects against stranded equipment. If the grid connection arrives early, the turbine can become a resilience and peak-management asset instead of an oversized primary plant. If the connection is delayed, it can keep the site moving.

The key diligence question is simple: what problem is the turbine solving? If it solves a documented interconnection delay and the fuel, permits, and operating costs are underwritten, it can be a practical bridge. If it is being used to make an uncertain site look powered on a slide deck, the risk is much higher.

The short version

Transformers rarely get the same attention as GPUs or transmission lines, but they are often the physical bottleneck between a power plan and an operating data hall. In March 2026, the U.S. Department of Energy said distribution-transformer demand was 41% higher than in 2019 and that lead times had moved from three to six months in 2019 to one to two years or longer in 2024. DOE also said larger transformers for substations and generators can take three to four years.

That is not a generic manufacturing delay. It changes the way investors should read a data-center pipeline. A site that says “power secured” may have a signed utility agreement but no delivered transformer, no energized substation and no revenue-producing load.

What the transformer does

A transformer changes voltage. The grid may deliver power at a high transmission or sub-transmission voltage, while a data center needs lower-voltage distribution and, inside the campus, additional step-down stages for switchgear, UPS systems and server power supplies. Large sites often need a chain of transformers rather than one box.

The distinction matters because each box can have a different specification, supplier and delivery date. A standard distribution transformer is not a substitute for a custom power transformer sized for a large campus substation. A project can therefore be constrained by the slowest item in the chain.

Why lead times stretched

DOE has pointed to several overlapping causes: aging grid equipment, post-pandemic demand, shortages of materials and components, limited domestic manufacturing, workforce constraints and excessive customization. The result is that a utility, developer or data-center operator may be competing for capacity with factories, housing, electrification projects and other large loads.

Copper, electrical steel, bushings, tap changers and specialized insulation are all part of the supply picture. A factory can expand final assembly and still be limited by one upstream component. That is why headline announcements about new transformer plants do not immediately translate into shorter customer schedules.

The practical schedule

A useful way to model a site is to break power readiness into separate gates:

1. Utility studies and an executed service agreement.
2. Substation design and long-lead equipment procurement.
3. Transformer manufacturing, testing and transport.
4. Substation construction and interconnection.
5. Energization and commissioning.
6. Customer load ramp.

The first gate is commercial. The third and fourth are physical. The sixth is operational. They should not be reported as if they were the same milestone.

For example, a 300-megawatt project could announce a power contract in 2026, order its major transformers in 2026, receive them in 2028 and energize an initial 100-megawatt phase later in 2028. The announcement is real, but the schedule is still exposed to manufacturing, permitting, delivery and testing risk.

How to read company language

When a developer says “power available,” ask what is actually available. Does the phrase mean the utility has accepted the load request? Does it mean the customer has signed a contract? Does it mean the substation is built? Does it mean power is flowing to the data hall?

The same caution applies to “transformers ordered.” An order may be cancellable, subject to a deposit, pending final engineering or scheduled for only one phase. A project with a firm purchase order and a factory slot has more evidence than a project with an informal supplier discussion, but neither proves energization.

The DOE data is also a reminder to separate distribution transformers from larger substation units. DOE’s March 2026 webinar described distribution-transformer orders at one to two years or longer in 2024 and large transformers at as much as three to four years. A data center may need both categories.

What could improve the picture

Standardization is one of the clearest levers. Reusing proven voltage classes, ratings and protection schemes can reduce engineering changes. Earlier procurement can reserve factory capacity before a site reaches final construction. Utilities and developers can also coordinate demand forecasts so suppliers see a credible multi-year order book.

DOE’s supply-chain work focuses on domestic production, materials and manufacturing productivity. Those efforts may help, but they do not erase the existing queue. The near-term question is not whether the United States wants more transformer capacity. It is whether a specific project has secured the equipment, slot and delivery path required for its next energization milestone.

BOSSTOX takeaway

The transformer line item deserves a place next to land, fiber and customer contracts in every data-center diligence file. Track the transformer type, supplier, purchase-order status, factory acceptance test, delivery date and installation milestone. Treat “contracted megawatts” as commercial visibility and “transformer delivered” as a physical checkpoint.

The strongest project disclosure is not the biggest megawatt number. It is a dated chain of evidence showing how the project moves from a utility agreement to a tested, energized substation.

A project-level example shows why this matters. Suppose a utility confirms 300 MW of service in a 2026 study, but the first transformer delivery is scheduled for the fourth quarter of 2028. The commercial commitment can support site planning, yet the data center still cannot run its full load until the equipment is installed, tested and accepted. If the first phase needs only 100 MW, the operator may energize that portion while the remaining transformer bank stays on the critical path.

The diligence file should also record whether a transformer is new, refurbished or transferred from another project. A transfer can shorten the schedule but may create compatibility, warranty or inspection issues. Factory acceptance tests, transport permits and on-site oil processing are real schedule events. They deserve dates in the project model, just like a building permit or customer delivery.

AI is often discussed as if it lives inside a software product. In practice, every model request depends on a physical chain: servers, power distribution, cooling, networking, storage, and a grid connection. As models become larger and more people use them, electricity is turning into one of the main constraints on AI infrastructure.

The numbers behind the headlines

The basic issue is simple. A data center converts electricity into computation and heat. The servers perform the useful work, while cooling and power equipment support that work. The International Energy Agency estimates that data centers used about 415 terawatt-hours of electricity worldwide in 2024, or roughly 1.5 percent of global electricity consumption. Its base case projects that data-center consumption will approach 945 terawatt-hours by 2030. That is a large increase, even though it remains a minority share of total world demand.

The local impact can be much larger than the global percentage suggests. A data center is not spread evenly across a country. It is concentrated in a particular utility territory, substation, and transmission corridor. A conventional facility may draw tens of megawatts. AI-focused facilities can be much larger. If several projects request service in the same region, the utility must plan for generation, substations, transmission, transformers, and backup systems at the same time.

Why AI load is different

AI changes the shape of the demand. Traditional enterprise servers tend to have a more predictable and mixed workload. AI training can run many accelerators at high utilization for long periods. Inference may be less steady, but it can grow with every customer-facing application. The result is a demand profile with high power density and a strong incentive to keep expensive equipment busy.

Three ways the industry is responding

The first response is efficiency. Better chips can perform more work per watt. Software teams can use smaller models, quantization, batching, caching, and routing to reduce unnecessary computation. A model that answers a routine question with a small specialized model may avoid sending the request to a much larger general model. These choices matter because the cheapest megawatt is often the one the workload never needs.

The second response is better scheduling. Some training and batch workloads can move to hours when the grid has more capacity or when renewable generation is abundant. This does not solve every reliability problem, but it lets operators treat part of AI demand as flexible. Data centers can also combine grid power with storage, onsite generation, or long-term power contracts. Those tools have different costs and emissions profiles, so the important question is not whether a facility claims to use clean energy, but how its power is matched to its actual hourly demand.

The third response is better grid planning. Large facilities should not be treated as ordinary commercial customers if they require a new substation or transmission upgrade. Utilities need clear information about project size, construction timing, ramp rates, and expected utilization. Developers need a realistic interconnection process rather than an assumption that a requested load will appear on schedule. The U.S. Department of Energy has warned that data-center connection requests can reach hundreds of megawatts and may stretch local grid capacity.

Why geography still matters in the cloud

This is why geography still matters in a cloud industry. A company may buy compute from a provider with a global network, but the physical servers are somewhere. Regions with reliable power, available land, fiber, water or alternative cooling resources, and a predictable permitting process will attract more capacity. Regions with congested grids may see projects delayed even when demand is strong.

There is also a customer-side lesson. Businesses buying AI services should ask how the provider measures utilization, latency, and energy use. A low price per token does not tell the whole story if the workload creates large bursts, requires redundant capacity, or sends data across regions. A good architecture uses the smallest model and the closest service that can meet the business requirement.

The bottom line

The energy problem is therefore not an argument against AI. It is an argument for treating AI as infrastructure. The industry must build enough power and network capacity to support useful applications without assuming that every request deserves the largest available model. Efficient chips, flexible workloads, transparent reporting, and coordinated grid planning can make growth more manageable.

AI will continue to need electricity. The practical question is whether the industry can turn that demand into a planning problem that utilities, developers, and customers can measure and solve.

What BOSSTOX is watching

• Utility filings that separate firm data-center load from speculative requests.
• How much training work operators are willing to make genuinely flexible.
• Whether providers start publishing hourly, not annual, clean-energy matching.