How Electricity Is Produced: Comparing the Major Technologies, Costs, and the Best Power Mix for AI Data Centers

Electricity can be produced in many ways, but the economics of each technology are very different. Solar and wind can produce inexpensive electricity when conditions are favorable, while nuclear, natural gas, hydroelectricity and geothermal can provide electricity more consistently.

The most commonly used comparison is the Levelized Cost of Electricity, or LCOE. LCOE estimates the lifetime cost of building, financing, operating, maintaining and fueling a power plant divided by the electricity it produces.

However, LCOE does not tell the entire story. A megawatt-hour available whenever the grid needs it is not necessarily equivalent to a megawatt-hour produced only when the sun shines or wind blows.

This distinction is becoming increasingly important as artificial intelligence, cloud computing and hyperscale data centers create extremely large, concentrated electricity loads.

Approximate Cost of New Electricity Generation

Recent global and U.S. studies provide the following broad benchmarks:

These figures should be interpreted as ranges rather than fixed prices. Financing costs, fuel prices, geography, labor, permitting, transmission requirements, tax credits and utilization can materially change project economics.

IRENA reported that newly commissioned global onshore wind averaged approximately $34/MWh in 2024, solar PV about $43/MWh, hydroelectricity about $57/MWh, geothermal about $60/MWh, offshore wind about $79/MWh and bioenergy about $87/MWh.

1. Coal-Fired Electricity

Coal plants burn pulverized coal to heat water, producing high-pressure steam that turns a turbine connected to a generator.

The technology is mature and capable of continuous electricity production, but coal has several disadvantages:

• high carbon dioxide emissions
• sulfur, nitrogen oxide and particulate emissions
• fuel transportation requirements
• ash disposal
• significant water use
• increasingly expensive pollution-control systems

Existing coal plants can sometimes remain economical because their construction cost has already been paid. Building a completely new coal facility is considerably harder to justify economically in many developed markets.

Coal’s historical advantage was inexpensive fuel and around-the-clock operation. Its disadvantage today is increasingly the total cost of fuel, emissions controls and environmental compliance.

2. Natural Gas

Natural gas is one of the most important sources of electricity in the United States.

A modern combined-cycle gas turbine, or CCGT, first burns gas in a turbine. Instead of wasting the extremely hot exhaust gases, the plant uses that heat to produce steam and generate additional electricity through a second turbine.

This dramatically improves efficiency.

Natural gas plants have several advantages:

• relatively low construction cost
• fast construction compared with nuclear plants
• ability to increase or decrease output rapidly
• high reliability
• relatively small physical footprint
• useful for balancing wind and solar

The major economic risk is fuel cost.

Unlike wind, solar or nuclear plants, gas generators continually purchase substantial quantities of fuel. When natural-gas prices increase, electricity-generation costs increase as well.

3. Nuclear Power

Nuclear reactors produce heat through nuclear fission rather than combustion.

Uranium atoms split inside the reactor and release enormous amounts of heat. That heat ultimately produces steam, which drives a turbine and generator.

Nuclear’s major strengths are:

• extremely high capacity factor
• continuous 24-hour electricity production
• very low direct carbon emissions
• small land requirement per unit of electricity generated
• relatively small quantities of fuel

Its main disadvantages are:

• enormous initial construction cost
• long development and construction periods
• regulatory complexity
• financing costs
• radioactive-waste management
• expensive decommissioning

The economics of nuclear are therefore unusual.

Once an existing nuclear plant is operating, the incremental cost of generating electricity can be attractive. Building a completely new conventional nuclear plant, however, can be extremely capital intensive.

This is why small modular reactors, or SMRs, are attracting substantial attention.

The objective is to standardize reactor manufacturing, reduce construction risk and shorten project schedules.

Whether SMRs can achieve these cost reductions at commercial scale remains one of the industry’s biggest questions.

4. Solar Photovoltaic Power

Solar photovoltaic panels convert sunlight directly into electricity using semiconductor materials.

Solar offers several major advantages:

• no fuel cost
• relatively low operating cost
• modular construction
• short construction schedules
• installation from rooftops to gigawatt-scale facilities

Its fundamental limitation is straightforward:

Solar panels do not generate electricity at night.

Output also varies with cloud cover, season and latitude.

Consequently, as solar becomes a larger percentage of an electrical system, its economics increasingly depend on:

• batteries
• transmission
• demand shifting
• natural-gas backup
• hydroelectric storage
• nuclear or other firm generation

5. Wind Power

Wind turbines convert the kinetic energy of moving air into rotational energy that drives a generator.

Modern onshore wind is among the cheapest sources of new electricity in locations with good wind resources.

Wind’s advantages include:

• no fuel cost
• very low operational emissions
• competitive generation cost
• relatively fast construction

Its limitations include:

• variable output
• dependence on suitable geography
• transmission requirements
• land-use and community concerns
• grid-balancing requirements

Offshore Wind

Offshore wind typically produces electricity more consistently than onshore installations because ocean winds are stronger and steadier.

But offshore projects are considerably more expensive because turbines must be installed and maintained in harsh marine environments.

6. Hydroelectric Power

Hydroelectric plants use moving water to turn turbines.

Large reservoirs have a major advantage over wind and solar: electricity production can often be controlled.

Water can be stored behind a dam and released when electricity demand rises.

Hydroelectricity therefore provides several services simultaneously:

• low-cost electricity
• grid balancing
• energy storage
• peak-demand generation
• frequency regulation

The biggest limitation is geography.

Many of the best dam locations have already been developed, and major new dams can have substantial environmental and social impacts.

7. Pumped-Storage Hydroelectricity

Pumped storage is better understood as an enormous battery.

When electricity is inexpensive, pumps move water uphill into a reservoir.

When electricity becomes valuable, the water flows downhill through turbines to regenerate electricity.

Some energy is lost during this cycle, but pumped storage can shift electricity from periods of excess generation to periods of high demand.

This capability becomes increasingly valuable as solar and wind penetration increases.

8. Geothermal Power

Geothermal plants use heat from inside the Earth.

Underground hot water or steam is brought to the surface and used directly or indirectly to operate turbines.

Its most important advantage is reliability.

Unlike solar and wind, geothermal plants can often operate continuously.

The limitation is geological.

Conventional geothermal works best where hot underground resources are relatively accessible.

New technologies such as enhanced geothermal systems and advanced drilling could potentially make geothermal available in many more locations.

9. Biomass

Biomass power plants burn or process materials such as:

• wood waste
• agricultural residues
• municipal organic waste
• specially grown energy crops

Like coal and gas, biomass plants can generally produce electricity when required.

However, fuel supply and transportation can be expensive.

10. Waste-to-Energy

Municipal solid waste can also be burned or processed to generate electricity.

Waste-to-energy plants provide two services:

1. electricity generation
2. waste disposal

This can make their economics different from conventional power plants because avoided landfill costs may constitute part of the project’s value.

However, emissions-control systems can be expensive, and projects frequently face local opposition.

11. Concentrated Solar Power

Concentrated solar power, or CSP, uses mirrors to concentrate sunlight and generate heat.

The heat can then produce steam and electricity.

One major advantage is that heat can be stored relatively inexpensively in materials such as molten salt.

A CSP plant can therefore continue generating electricity after sunset.

12. Diesel and Oil-Fired Generation

Diesel generators are easy to install and extremely reliable.

They are therefore widely used for:

• emergency backup
• remote communities
• hospitals
• military facilities
• temporary power
• islands

But diesel electricity is normally expensive because fuel costs are high.

Consequently, diesel generally makes little economic sense for large-scale continuous electricity production where a conventional grid is available.

13. Tidal and Wave Energy

Oceans contain enormous amounts of energy.

Tidal generators can exploit predictable water movements, while wave-energy systems attempt to convert ocean-wave movement into electricity.

These technologies have an attractive theoretical advantage: tidal cycles are extremely predictable.

However, the marine environment makes construction and maintenance difficult.

Corrosion, storms, underwater transmission and mechanical wear currently prevent tidal and wave power from competing economically with mature solar, wind, hydro and natural gas in most markets.

14. Hydrogen-Fired Generation

Hydrogen can be burned in turbines or converted into electricity using fuel cells.

The major question is where the hydrogen comes from.

Hydrogen is an energy carrier rather than a primary energy source.

If electricity is first used to produce hydrogen through electrolysis and hydrogen is later converted back into electricity, substantial energy is lost during the round trip.

Hydrogen therefore may make greater economic sense for:

• long-duration energy storage
• industrial heat
• steel production
• chemical manufacturing
• backup electricity
• seasonal storage

rather than routine daily electricity production.

What Type of Electricity Is Best for AI Data Centers?

AI data centers fundamentally change the electricity discussion.

Traditional commercial buildings may use several megawatts.

Large AI campuses can require hundreds of megawatts, while future hyperscale developments can approach or potentially exceed 1 gigawatt of demand.

That is comparable to the electricity consumption of a major industrial facility or a small city.

The challenge is not simply producing cheap electricity.

An AI data center needs:

large quantities of electricity + extremely high reliability + predictable cost + rapid interconnection + continuous operation.

This makes the ideal energy strategy different from the strategy for an ordinary commercial building.

AI Data Centers Need 24/7 Power

AI training clusters may contain tens or hundreds of thousands of GPUs operating continuously.

Interruptions can affect:

• model-training jobs
• computing utilization
• customer workloads
• cooling systems
• network infrastructure
• storage systems
• expensive GPU assets

For this reason, the cheapest intermittent electricity source is not automatically the best primary source for an AI data center.

The most suitable architecture is likely a diversified power portfolio:

Nuclear + natural gas + renewables + battery storage + grid connection + microgrid controls.

Each component serves a different purpose.

Nuclear: Excellent for Large, Continuous AI Loads

Nuclear energy may be particularly well matched to hyperscale AI infrastructure because reactors can operate continuously at very high capacity factors.

Major advantages include:

• 24/7 generation
• extremely high reliability
• large output from a relatively small site
• relatively predictable fuel costs
• very low operational carbon emissions
• decades-long operating life

Microsoft, for example, has supported restarting the former Three Mile Island Unit 1—now called the Crane Clean Energy Center—to supply reliable carbon-free electricity associated with its data-center demand.

Existing nuclear plants may therefore become increasingly valuable assets in regions experiencing rapid data-center development.

Future SMRs could also eventually be located near large technology campuses, reducing dependence on distant transmission.

Natural Gas: The Practical Bridge for Rapid Expansion

Natural-gas generation may remain one of the most practical solutions for AI data centers needing electricity quickly.

A gas plant can generally be constructed faster than:

• a nuclear plant
• major transmission infrastructure
• large hydroelectric projects

Gas generation can also follow changes in demand rapidly.

For large AI campuses, combined-cycle gas plants could provide:

• baseload power
• backup generation
• peak-demand capacity
• renewable balancing
• microgrid generation

Their main disadvantages remain fuel-price exposure and emissions.

Nevertheless, natural gas is likely to play an important transitional role where grid capacity cannot expand quickly enough.

Solar and Wind: Excellent Low-Cost Contributors

Solar and wind can substantially reduce the average cost of electricity for data centers.

However, neither should generally be considered a standalone 24/7 power source for a hyperscale AI facility.

A 1-GW data center cannot simply depend on a 1-GW solar plant.

The solar plant may generate 1 GW around noon under ideal conditions and essentially zero electricity at night.

The better strategy may be:

solar/wind + storage + firm generation + grid access.

During periods of high renewable generation, renewable electricity can supply the data center and charge batteries.

During periods of low renewable generation, nuclear, natural gas, hydroelectricity, geothermal or the grid can take over.

Geothermal: Potentially Ideal

Geothermal deserves considerably more attention in the AI-power discussion.

It combines several desirable characteristics:

• continuous generation
• relatively small land footprint
• low emissions
• high capacity factor
• limited dependence on weather

Historically, geothermal development has been geographically constrained.

Advanced drilling and enhanced geothermal systems could change this.

If next-generation geothermal becomes commercially scalable, it could become one of the most attractive electricity sources for data centers.

Battery Storage: Essential, but Not the Primary Generator

Large battery installations can provide:

• instantaneous backup
• peak shaving
• frequency regulation
• renewable balancing
• demand management
• temporary operation during grid disturbances

Batteries are particularly valuable because they respond almost instantly.

However, batteries do not create electricity.

They store electricity produced by another source.

Therefore, batteries are best considered part of the data center’s power-management infrastructure, not its fundamental energy source.

The Data Center Could Become Part of the Power System

One of the most important future changes may be that large AI campuses cease being passive electricity consumers.

Instead, they could become:

power producers + consumers + storage operators + grid-balancing participants.

This is where government and private-sector participation could create significant benefits.

The U.S. Department of Energy has specifically highlighted microgrids as a potential method for serving large electricity loads such as data centers while reducing pressure on transmission and distribution infrastructure.

A Data Center Microgrid

Consider a large AI campus with:

• 500 MW data-center demand
• 300 MW natural-gas generation
• 300 MW solar
• 150 MW battery storage
• 100 MW geothermal
• utility-grid connection

The campus does not necessarily need every source operating at full output simultaneously.

An intelligent energy-management system could continuously optimize the portfolio.

At noon:

Solar production may be high.

The system could:

• operate the data center
• charge batteries
• reduce natural-gas generation
• export excess electricity into the public grid

At 8 PM:

Solar production disappears.

The system could:

• increase gas generation
• discharge batteries
• draw power from the grid
• use geothermal or nuclear generation

During a grid emergency:

The data center could temporarily reduce grid withdrawals and operate more heavily from its own generation and battery systems.

This effectively turns the AI campus into a grid-support asset rather than simply a grid burden.

Feed Excess Electricity Back Into the Grid

One particularly important opportunity is allowing private data-center generation to export surplus electricity.

Suppose a campus builds enough electricity infrastructure for peak demand.

Not all of that capacity will necessarily be needed every minute.

Rather than leaving excess generation unused, it could be sold into the regional electricity market.

The model becomes:

Generate → Consume → Store → Export surplus → Import when economical.

This would improve utilization of expensive generation assets.

For example, a 500-MW data center might build 700 MW of combined generation and storage capability to provide redundancy.

If the facility is only using 450 MW at a particular moment, some of the remaining capacity could potentially supply the grid, subject to interconnection and market rules.

This arrangement could benefit both parties.

Data-center operator benefits

The company could:

• monetize surplus generation
• reduce effective electricity cost
• improve return on energy infrastructure
• increase energy security
• hedge electricity prices

Grid benefits

The utility or regional transmission organization could gain:

• additional generation capacity
• emergency reserves
• frequency response
• demand flexibility
• battery-storage resources
• lower peak-load pressure

FERC has increasingly focused on exactly these issues as AI data centers and other very large loads seek grid connections.

In June 2026, FERC initiated actions involving all six regional grid operators under its jurisdiction aimed at improving the integration of large electricity users such as data centers while maintaining reliability and affordability.

Bring Your Own Generation

A potentially powerful model is what FERC has described as “Bring Your Own New Generation.”

Instead of a data center requesting 500 MW from an already constrained electricity system, the developer could simultaneously develop or contract for new generation.

For example:

500-MW data center

plus

500-MW new gas, nuclear, geothermal, solar/storage or other generation

could be evaluated together.

FERC has argued that coordinated development of large loads and associated generation may reduce transmission upgrades and accelerate interconnection.

This model could significantly change how America finances electricity infrastructure.

Rather than asking existing electricity customers to fund all of the grid expansion needed for AI, part of the investment can come directly from companies creating the new demand.

Government + Private Sector Partnership

A practical AI-energy strategy may therefore require cooperation among:

Federal government

State governments

Utilities

Grid operators

Technology companies

Independent power producers

Infrastructure investors

The objective should not necessarily be for government to directly generate all the electricity.

Government can instead create the infrastructure and market framework that makes private investment economical.

Government’s Potential Role

Government can help through:

• permitting reform
• transmission approvals
• standardized interconnection rules
• nuclear licensing modernization
• SMR demonstration programs
• geothermal research
• infrastructure financing
• grid modernization
• tax incentives
• energy-storage programs
• cybersecurity standards
• land-use planning

Government may also coordinate large regional transmission projects that would be difficult for one private company to develop independently.

Private Sector’s Potential Role

Technology companies and energy developers can invest directly in:

• power plants
• microgrids
• solar farms
• batteries
• advanced nuclear
• geothermal
• transmission connections
• backup generation
• efficiency technologies

Private companies can also enter into long-term Power Purchase Agreements, or PPAs, providing energy developers with predictable revenue that helps finance new generation.

AI Companies Could Help Finance America’s Next Power Plants

This is potentially one of the most important economic opportunities created by AI.

Data centers create enormous electricity demand.

That demand can become an anchor customer for new electricity infrastructure.

For example:

An energy company may hesitate to finance a multi-billion-dollar plant without knowing who will purchase its electricity.

A hyperscaler offering a 15- or 20-year electricity contract dramatically changes that calculation.

The result could be:

AI demand → long-term electricity contract → project financing → new power plant → additional national generating capacity.

Eventually, that generation asset may serve far more than the original data center.

Microsoft’s agreement supporting the restart of the Crane Clean Energy Center illustrates how data-center demand can help support investment in firm generating capacity.

Data Centers Can Also Reduce Demand When the Grid Is Stressed

Electricity management does not always require creating more generation.

Some computing workloads may be flexible.

Certain AI workloads could potentially be moved:

• to another hour
• to another data center
• to another geographic region

For example, a training job that does not need to finish immediately might be postponed from an electricity-demand peak at 6 PM to an off-peak period at 2 AM.

This creates a potential form of computational demand response.

A data center could tell the grid operator:

When electricity supply becomes constrained, we can temporarily reduce grid consumption by 50 MW.

That 50 MW could be extremely valuable during peak demand.

FERC has specifically recognized that large loads can have unusual flexibility and may be capable of changing electricity consumption rapidly.

Use Excess Electricity Rather Than Waste It

The growth of AI may also create an opportunity to use electricity that would otherwise have little economic value.

Renewable-heavy grids sometimes produce more electricity than customers need.

Wholesale electricity prices can consequently fall very low or even become negative.

Instead of curtailing generation, flexible data centers could increase workloads during those periods.

For example:

Excess solar at noon → run additional AI training

Excess wind overnight → run compute-intensive workloads

Low electricity prices → charge data-center batteries

This effectively converts computing into a flexible industrial load.

In the future, some AI workloads might migrate geographically depending on where electricity is inexpensive and abundant.

The Ideal AI Energy Campus

A future hyperscale AI campus might therefore look less like a traditional data center and more like an integrated industrial energy complex.

Imagine:

1 GW AI Campus

Power sources

• 400 MW nuclear or geothermal
• 300 MW combined-cycle natural gas
• 500 MW solar
• 300 MW wind contracts
• 400 MW / 1,600 MWh battery storage
• grid connection

Because these technologies do not all operate at maximum output simultaneously, their capacities can overlap.

The campus energy-management system could continuously select the most economical combination.

When electricity is abundant:

charge batteries + increase AI workloads + export surplus electricity

When electricity becomes scarce:

reduce flexible compute + discharge batteries + increase firm generation + reduce grid withdrawals

Such a system could effectively function as both:

an AI factory and a virtual power plant.

Why Public-Private Participation Matters

The largest risk is building two systems independently:

Data centers demanding enormous amounts of electricity

while

utilities separately attempt to build enough generation and transmission to serve them.

This approach may produce:

• long connection queues
• transmission bottlenecks
• higher electricity prices
• stranded infrastructure
• reliability concerns

A more integrated approach would plan:

new load + new generation + storage + transmission together.

FERC’s recent actions are moving toward clearer rules for large-load interconnection, co-location and flexible transmission service.

Protecting Existing Electricity Customers

Public-private cooperation should also address an important policy question:

Who pays for the new grid infrastructure required by enormous data centers?

If a private AI campus requires billions of dollars in new transmission infrastructure, automatically spreading those costs across residential consumers can become controversial.

One possible model is:

beneficiary pays.

The data-center operator finances a substantial portion of infrastructure specifically required by its load.

At the same time, if the new infrastructure improves the regional grid and serves millions of other customers, some broader cost sharing may be appropriate.

The exact allocation depends on regulatory structures and regional grid rules.

FERC has emphasized both accelerating large-load connections and protecting consumers from inappropriate cost shifting.

A New Electricity Business Model

Historically the electricity system largely followed this structure:

Power plant → transmission grid → utility → customer

AI data centers may help create a more dynamic structure:

Power plant ↔ grid ↔ data center ↔ battery ↔ private generation

Electricity will move in both directions.

The data center may:

• purchase electricity
• generate electricity
• store electricity
• sell electricity
• reduce demand
• provide reserve capacity
• stabilize the grid

This is potentially more efficient than treating a gigawatt-scale data center as simply another electric bill.

Which Technologies Appear Best Suited to AI Data Centers?

The most practical solution is therefore unlikely to be one technology.

It is a hybrid system.

The Cheapest Electricity Is Not Necessarily the Cheapest Grid

This remains perhaps the most important concept in electricity economics.

Suppose a solar plant produces power for:

$35/MWh

and a nuclear plant produces power for:

$90/MWh.

Solar initially appears far cheaper.

But imagine a data center requires 500 MW continuously at 2:00 AM.

The nuclear plant can continue operating.

The solar plant cannot.

Solar may therefore require some combination of:

• battery storage
• transmission
• backup natural gas
• demand response
• excess installed capacity

Those additional costs are not always fully reflected in plant-level LCOE.

Likewise, nuclear generation may appear expensive based purely on LCOE but simultaneously provide:

• firm capacity
• inertia
• frequency stability
• around-the-clock generation
• very high capacity factor

Power-system planning therefore requires more than comparing individual generation costs.

Capacity Factor Matters

A power plant’s rated capacity does not tell us how much electricity it actually generates.

A 1-GW nuclear plant operating at a 90% capacity factor could produce roughly:

7.9 terawatt-hours annually.

A 1-GW solar installation operating at a 25% capacity factor would produce roughly:

2.2 terawatt-hours annually.

Therefore, approximately:

3.6 GW of 25%-capacity-factor solar would be needed to generate the same annual energy as 1 GW of nuclear operating at 90%.

Even then, the electricity would not arrive at the same time.

This illustrates why data centers and electric utilities must consider more than installed megawatts.

The Best Future Electricity System

There is unlikely to be a single winner.

A resilient electricity system could combine:

Solar for inexpensive daytime electricity.

Wind for low-cost generation during favorable conditions.

Natural gas for flexible and rapidly deployable firm capacity.

Nuclear for continuous high-capacity-factor generation.

Hydroelectricity for dispatchable generation and storage.

Geothermal for continuous renewable baseload.

Battery storage for shifting electricity across hours.

Transmission for moving electricity among regions.

AI data centers themselves could become flexible loads, power producers and storage operators.

The economics increasingly favor a portfolio of complementary technologies rather than dependence on one source.

The Bigger Opportunity

AI’s electricity consumption is often presented only as a problem.

It may also become an infrastructure opportunity.

If designed correctly, enormous private-sector electricity demand could help finance:

• new nuclear plants
• natural-gas plants
• geothermal development
• renewable generation
• battery storage
• transmission
• microgrids
• grid modernization

Rather than simply asking:

“How will the grid supply enough electricity for AI?”

a more productive question may be:

“How can AI investment help build the next generation of the electric grid?”

A partnership in which private companies finance new generation, consume what they need, store electricity when it is abundant, reduce consumption when the grid is constrained and sell excess electricity back into the system could substantially increase overall grid efficiency.

Under that model, AI data centers would not merely become some of America’s largest electricity consumers.

They could become important participants in building, financing and stabilizing America’s future electricity infrastructure.

Fuel-Cell Electricity: Cost and Role in Data Centers

Fuel cells generate electricity electrochemically rather than by combustion. Depending on the system, they can use hydrogen, natural gas, biogas, or other fuels.

The electricity cost varies significantly according to the fuel used, system size, utilization, capital cost and whether the waste heat is also recovered.

Approximate Fuel-Cell Electricity Cost

A U.S. Department of Energy technology assessment estimated stationary fuel-cell electricity at approximately 15¢/kWh, equivalent to roughly $150/MWh. That figure should be viewed as a historical benchmark rather than a universal current price because fuel-cell economics vary substantially by system and fuel.

DOE data for larger 100-kW to 3-MW stationary fuel-cell systems show electrical efficiencies of roughly 42–47% historically, with technical targets above 50% and approximately 60% possible for solid-oxide fuel cells. Combined heat-and-power configurations can reach total energy utilization of roughly 70–90% because useful heat is recovered rather than discarded.

Why the Fuel Cost Matters So Much

With hydrogen fuel cells, the cost of the hydrogen often dominates electricity economics.

A useful approximation is:

Electricity cost = hydrogen consumption × hydrogen price + capital + maintenance

A fuel cell operating near 50–60% efficiency may require approximately 0.05–0.06 kg of hydrogen per kWh of electricity.

That means:

• hydrogen at $2/kg contributes roughly 10–12¢/kWh
• hydrogen at $4/kg contributes roughly 20–24¢/kWh
• hydrogen at $8/kg contributes roughly 40–48¢/kWh

And those figures are before adding the cost of the fuel-cell equipment, financing, maintenance and hydrogen storage.

This explains why hydrogen fuel cells can be attractive where very inexpensive hydrogen is available, but can be quite expensive where hydrogen must be produced, compressed, transported and stored.

DOE’s long-term hydrogen program is working toward much lower hydrogen production and delivery costs, including a goal of approximately $1/kg hydrogen production by 2032 and delivery/dispensing costs below approximately $2/kg in targeted applications.

Natural-Gas Fuel Cells May Be More Relevant to AI Data Centers Today

For large data centers, solid-oxide fuel cells, or SOFCs, operating on pipeline natural gas can be particularly interesting.

Rather than burning natural gas in a turbine, an SOFC converts the chemical energy of the fuel directly into electricity.

Potential advantages include:

• high electrical efficiency
• continuous 24/7 operation
• relatively small physical footprint
• modular deployment
• low local air-pollutant emissions compared with conventional combustion
• reduced dependence on large transmission upgrades
• ability to locate generation close to the data center

DOE notes that stationary fuel-cell systems can operate on natural gas today, while direct hydrogen and renewable fuels offer longer-term options.

For an AI data center that cannot wait several years for a new transmission line, a modular fuel-cell installation could therefore serve as behind-the-meter firm generation.

Fuel Cells vs. Other Data-Center Power Sources

Fuel cells therefore are not necessarily the cheapest source of electricity.

Their value comes from a different combination:

reliability + modularity + proximity to load + smaller transmission requirement + relatively fast deployment.

For AI data centers, those characteristics can sometimes justify paying more per megawatt-hour.

Fuel Cells Can Also Strengthen the Data-Center Microgrid

A large AI campus could combine:

Grid + solar + battery + natural-gas fuel cells + backup generators

or, over the longer term:

Grid + solar/wind + batteries + hydrogen production + hydrogen storage + fuel cells

During periods of abundant inexpensive electricity, excess renewable power could potentially be used to produce hydrogen through electrolysis.

That hydrogen could be stored and later converted back into electricity through fuel cells.

This provides long-duration energy storage, although round-trip efficiency is considerably lower than battery storage. DOE research notes that hydrogen storage becomes more attractive for very long storage durations because the marginal cost of storing additional hydrogen can be relatively low even though the conversion process loses substantial energy.

Practical Role for AI Data Centers

For today’s AI data centers, I would classify fuel cells this way:

Natural-gas SOFC: potentially attractive for continuous distributed generation.

Hydrogen fuel cell: currently more attractive for backup, resilience and long-duration storage than for the cheapest continuous generation.

Fuel-cell CHP: particularly attractive where the heat produced can also be used.

Future green-hydrogen fuel cells: potentially important if hydrogen production approaches $1–$2/kg at scale.

Fuel cells therefore should be viewed as another component in the broader data-center energy portfolio—not necessarily a replacement for nuclear, gas turbines, solar, wind or batteries, but as a technology that can provide high-reliability, distributed, behind-the-meter power.