Power purchase agreements for AI data centers

In short

A power purchase agreement, or PPA, is a contract for an AI data center to buy electricity. It can be for power from a solar farm, a nuclear plant, a natural gas plant, or a wind farm. The power can be delivered over the public grid or generated right on the AI data center’s own property. The Federal Power Act splits legal oversight. The Federal Energy Regulatory Commission, FERC, regulates wholesale sales of electricity and the transmission of electricity across state lines. States regulate retail sales, the last sale to the end user. In 2024, soaring electricity demand from AI data centers drove a record year for corporate clean energy procurement, pushing PPA prices up nearly 35 percent. In 2025 and early 2026, a wave of new state tariffs locked in long term contracts and required large minimum payments. At the same time, FERC is rewriting interconnection rules and, for the first time, ordering grid operators to create special transmission options for AI data centers that sit next to power plants.

What is a power purchase agreement for an AI data center

A power purchase agreement is a contract between two parties. The seller owns a power plant. The buyer is the AI data center operator. The agreement sets the price, how much electricity the buyer will purchase, and who pays for the wires that move the power.

Many PPAs also transfer renewable energy certificates, or RECs. A REC represents the environmental benefit of one megawatt hour of renewable generation. If a PPA includes RECs, the AI data center can use them to claim that its electricity use is carbon free.

PPAs come in a few basic shapes. Getting the shape right matters because it determines which laws apply and how the power physically gets to the server racks.

Physical and virtual PPAs

A physical PPA moves electricity over transmission and distribution lines to the AI data center’s location. The generator and the AI data center must be in the same region, usually the same balancing authority area. The Amazon Talen Susquehanna nuclear PPA is an example of a large physical PPA. That 17 year deal provides up to 1,920 megawatts from the Susquehanna nuclear plant in Pennsylvania to Amazon’s AI data centers across Pennsylvania. Talen press release, Power Magazine

A virtual PPA is a financial agreement. The AI data center does not physically take the power. The generator sells the power into the wholesale market at the market price. The AI data center and the generator settle the difference between the PPA price and the market price. The AI data center continues to buy electricity from its local utility as usual. The virtual PPA gives the AI data center the RECs and a fixed price hedge. This structure is common when a company wants to claim renewable energy but its AI data centers are spread across many locations. In 2024, Big Tech companies accounted for 43 percent of all clean energy PPAs signed globally. Brookings Institution

Behind the meter and co-location PPAs

A behind the meter PPA means the generator sits on the AI data center’s own property, on the customer side of the utility meter. The power does not flow over the public grid. The AI data center consumes most of it directly. One working example is the Clearway Elbow Creek wind farm in Texas. A 122 megawatt wind project hosts a new high performance computing AI data center behind the meter. The first phase was energized in 2025. Clearway press release, PUC Texas filing, Data Center Dynamics

Co-location is a close cousin. An AI data center is placed right next to an existing power plant. It can draw power through a dedicated line, sometimes without using the wider transmission grid. Co-location can be behind the meter if the connection is on the plant side of the utility meter, or it can be a front of the meter arrangement that uses transmission service. Amazon’s Susquehanna deal was first proposed as a behind the meter co-location. After FERC raised concerns about cost shifting to other customers, the parties restructured it as a front of the meter grid connected PPA. Power Magazine

Who regulates a power purchase agreement

The answer turns on whether the sale is wholesale or retail. That line comes from the Federal Power Act.

The wholesale and retail line

Under the Federal Power Act, FERC has jurisdiction over the transmission of electric energy in interstate commerce and the sale of electric energy at wholesale in interstate commerce. Wholesale means a sale for resale. A power plant selling to a utility that will resell to homes is a wholesale sale. States retain authority over retail sales of electricity and over local distribution. 16 U.S.C. § 824(b)(1)

So, if an AI data center signs a PPA with a power plant that is located in another state and the power moves through the grid, that is generally a wholesale sale. The seller must have FERC authority. If the seller is a public utility, it must file its rate schedule with FERC. If it wants to sell at market based rates rather than cost based rates, it needs FERC market based rate authorization. 16 U.S.C. § 824d, FERC market-based rate initial applications

If an AI data center buys power from its local utility under that utility’s state approved tariff, the sale is retail. The state utility commission regulates the rates and service terms. FERC does not step in.

When a practice directly affects wholesale rates

The line is not always clean. In the 2016 case FERC v. EPSA, the Supreme Court held that FERC can regulate a practice that directly affects wholesale rates, even if it touches retail markets. Supreme Court That means a PPA’s terms can be subject to FERC review if they influence the wholesale price or grid operation, even when the sale at first looks like a retail transaction.

FERC applied this principle in its Tri State High Impact Load Tariff order in October 2025. Tri State, a utility, proposed a special tariff for large loads, like AI data centers. FERC rejected it because the tariff would have set terms for retail sales, an area reserved for states. FERC explained that it can set technical requirements for participation in wholesale markets, but it cannot directly regulate the rate a utility charges a retail customer. FERC Tri-State order

Behind the meter sales

A behind the meter arrangement, where the generation and the load sit on the same side of a single meter, often falls outside FERC wholesale jurisdiction when no power is sold across the grid to another party. But if the arrangement involves net metering (selling excess power back to the grid) or interconnection with the grid, FERC and state rules can still apply. The exact boundary is case specific.

What about PURPA

The Public Utility Regulatory Policies Act of 1978, or PURPA, requires electric utilities to buy power from certain smaller renewable and cogeneration facilities. These are called qualifying facilities, or QFs. A QF can be a small power production facility using solar, wind, or geothermal, or a cogeneration facility that produces both electricity and useful thermal energy. The utility must pay a rate not higher than the utility’s avoided cost, the cost the utility would pay to generate or buy the power somewhere else. 16 U.S.C. § 824a-3

Some AI data center projects might involve a cogeneration plant that provides both electricity and cooling. If that plant meets the operating and efficiency standards in FERC’s rules and files a self certification (Form No. 556), it can gain QF status. 18 C.F.R. § 292.203(b) The utility would then have an obligation to buy the plant’s excess power at avoided cost.

But after 2005, no electric utility is required to enter into a new purchase contract with a QF if the Commission finds the QF has nondiscriminatory access to competitive wholesale markets, such as those run by regional transmission organizations. 16 U.S.C. § 824a-3(m) Many of the regions where AI data centers are being built fall under these competitive market structures. For that reason, PURPA alone rarely frames the massive PPAs that hyperscalers sign today. It remains relevant for smaller on site projects.

Interconnection is the slowest step

Building the AI data center itself can take 12 to 18 months. Connecting a new power plant to the grid often takes over 4 years, with a median of 5 years for projects built in 2023. Lawrence Berkeley National Laboratory That gap drives many of the risks in a PPA.

How FERC Order No. 2023 changed the queue

Before a new generator can send electricity onto the grid, it must complete an interconnection study. For years, the process was first come, first served. A project filed a request and waited in a line with thousands of others, many of them speculative. Delays grew. By 2023, the median time from request to completion had stretched to about 5 years for projects that reached operation. LBNL

FERC Order No. 2023, issued in July 2023, replaced that serial queue with a cluster study process. Now, transmission providers study groups of projects together on a firm schedule. They face financial penalties if they miss deadlines. The order also raised the bar for developers. They must post study deposits from $55,000 up to $250,000 depending on project size. They must pay commercial readiness deposits at each study phase. They must show site control, 90 percent at the request and 100 percent at execution of the facilities study agreement. FERC Order No. 2023 overview These changes are meant to clear out speculative projects and speed the queue for projects that are ready to build.

A push to treat big loads like generators

In October 2025, the Secretary of Energy sent a letter to FERC under section 403 of the Department of Energy Organization Act. The letter directed FERC to start a rulemaking that would treat large load interconnections, like those for AI data centers, the same as generator interconnections. Today, an AI data center cannot enter the interconnection queue on its own. It must rely on a generator to do so. Under the DOE’s proposal, an AI data center could file as an interconnection customer, pay for the network upgrades it needs, and possibly get connected faster. The letter also recommended withdrawal penalties and full responsibility for network upgrade costs from large loads. DOE Section 403 letter FERC opened Docket No. RM26-4 and issued an advance notice of proposed rulemaking in October 2025, and it has said it will act by June 2026. FERC RM26-4

FERC orders co-location reform in PJM

In December 2025, FERC issued an order specifically about co located loads in PJM, the grid operator for 13 eastern states. FERC found that PJM’s existing tariff was unjust and unreasonable because it lacked proper rates, terms, and conditions for generators that serve an AI data center next door. The order directed PJM to create three new transmission service types, including Firm Contract Demand, Non Firm Contract Demand, and Interim Non Firm. It also reformed rules for behind the meter generation, including a megawatt threshold for netting and a three year transition period for existing behind the meter customers that expires in December 2028. FERC PJM Co-Location Order, FERC Fact Sheet, Law firm analysis

This order matters for developers who are planning co-located projects in PJM. They will need to understand which new service type fits their PPA and how the transition timeline affects existing behind the meter setups.

State large load tariffs are the new normal

While FERC handles wholesale markets and interconnection, state utility commissions decide what a local utility can charge a retail customer. In 2025, state regulators approved 29 large load tariffs, more than double the 14 approved from 2018 through 2024. Utility Dive The pattern is driven by a single concern. When an AI data center requests hundreds of megawatts of new capacity, someone must pay for the new power plants and transmission lines. State regulators want to make sure that cost falls on the AI data center, not on residential and small business customers.

How these tariffs work

The new tariffs share several common features.

• Minimum contract length, often 10 to 20 years.
• Minimum monthly payment, called a take or pay. The customer must pay for 75 to 85 percent of its contracted demand even if it uses less power in a given month.
• Large collateral or security deposits, sometimes equal to two years of minimum bills.
• Steep exit fees if the customer terminates early, often the full remaining minimum bills.
• Direct assignment of network upgrade costs to the new load.

These terms shift the risk of overbuilding away from the utility’s other customers and onto the AI data center.

The table below shows a few approved and pending tariffs, drawn from recent state filings and analyses collected by the Electric Power Research Institute.

Sources, EPRI Digest for Kansas, Michigan, Oregon, Florida. GovTech on Pennsylvania model.

Delaware opened a docket in October 2025 to develop a large load tariff and paused new interconnections for loads of at least 25 megawatts until the tariff is established. EPRI Digest

Virginia, a voluntary companion rate

Not every state is imposing mandatory tariffs. In Virginia, Dominion Energy offers a voluntary companion rate, called Schedule CFG, approved in 2024. A non residential customer using more than 500 kilowatts can use Schedule CFG to either buy up to 100 percent of the energy and RECs from a new carbon free or renewable facility that Dominion builds on the customer’s behalf, or buy power through a utility arranged PPA with a third party. Under the utility-build option, the costs stay ring fenced, separate from Dominion’s overall cost of service, so they do not flow through to other ratepayers. Under the PPA option, the costs are passed directly to the participating customer through the Schedule CFG Agreement and likewise do not flow through to other ratepayers. SCC Virginia Docket, Schedule CFG tariff, SCC Virginia Docket

Virginia also enacted HB 2084 in 2025. The law directs the State Corporation Commission to determine whether utilities are using reasonable customer classifications and whether new or separate customer classes are reasonable. NASUCA presentation

Texas, ERCOT’s own path

The Electric Reliability Council of Texas, ERCOT, operates outside FERC jurisdiction for most purposes because it does not engage in interstate transmission. So Texas sets its own rules through the Public Utility Commission of Texas and state legislation.

ERCOT created the Large Flexible Load, or LFL, program in mid 2022. It classifies any facility with an expected peak demand of 75 megawatts or more as a large flexible load. As of September 2024, ERCOT had approved 5,479 megawatts of LFL capacity. EIA

Senate Bill 6, enacted in 2025, tightened the rules. Any large load of 75 megawatts or more that interconnects after December 31, 2025 must meet three new requirements. It must report backup power capacity equal to at least 50 percent of its on site load. It must pay a $100,000 interconnection study fee. And it must show financial commitment for transmission infrastructure if its total load at a single site would exceed 75 megawatts. Texas SB 6, enrolled bill The bill also gives ERCOT the authority to order curtailment of large loads during energy emergency alerts. This means the grid operator can tell an AI data center to reduce its power draw to keep the system stable.

SB 6 also regulates net metering arrangements. If a large load and a generation resource were registered as separate standalone generators before September 1, 2025, their net metering arrangement is limited after the law takes effect. This matters for behind the meter co-location projects in ERCOT.

Because ERCOT is a competitive wholesale market, PPAs there are often structured as financial hedges or direct bilateral contracts using ERCOT protocols. The new state rules add a layer of physical and financial readiness requirements that developers must plan for.

How AI demand is reshaping the PPA market

The numbers show a market under strain from a single powerful buyer class.

Global corporate clean energy procurement hit a record 68 gigawatts in 2024, a 29 percent jump from 2023. AI data centers accounted for more than 17 gigawatts of that total, more than a quarter of the global corporate volume. PV Magazine USA In the United States in 2024, AI data center operators took nearly 60 percent of all corporate clean energy deals. S&P Global Commodity Insights That concentration pushed average PPA prices up by 35 percent in 2024. Brookings analysis

Hyperscaler capital expenditures reflect the scale. Spending exceeded $200 billion in 2024 and is forecast to pass $600 billion in 2026, with about 75 percent tied directly to AI infrastructure. Energy IB Microsoft became the world’s largest corporate clean power buyer, with 40 gigawatts contracted by late 2025. Energy IB

Nuclear power PPAs take center step

AI data center operators are looking for 24 hour a day, carbon free power, and nuclear fills that need. In the past year through mid 2025, Big Tech companies contracted more than 10 gigawatts of possible new U.S. nuclear capacity. Introl

• Microsoft signed a 20 year PPA for the 835 megawatt Three Mile Island Unit 1, renamed the Crane Clean Energy Center. The $1.6 billion project targets a 2027 restart of the reactor, which had been shut down since 2019. Constellation, the plant’s owner, secured a $1 billion federal loan in November 2025. Constellation press release, CNBC
• Amazon and Talen Energy restructured their Susquehanna nuclear PPA as a 17 year, $18 billion front of the meter deal for up to 1,920 megawatts, with full volume by 2032. Amazon is investing $20 billion in Pennsylvania AI data centers which will be powered under its restructured Susquehanna nuclear PPA. Talen Energy press release, Power Magazine
• Meta committed to a 20 year PPA for 1.1 gigawatts from the Clinton Clean Energy Center in Illinois, with Constellation, and supported facility expansion. Trellis
• Google signed the first corporate fleet agreement for small modular reactors, up to 500 megawatts from Kairos Power, with the first reactor expected to connect to the TVA grid by 2030. Canary Media, CNBC

Renewable PPAs still dominate, but with higher prices and longer terms

Renewable deals remain the largest piece by volume, but they are getting more expensive. In 2026, Google signed a 15 year PPA for 1,000 megawatts of solar in Texas with TotalEnergies. TotalEnergies Meta signed multiple solar PPAs. One with Enbridge for a $900 million, 600 megawatt solar project in Texas, and another $900 million, 600 megawatt project with ENGIE for solar in Texas. Utility Dive, Power Systems Technology Clearway signed three PPAs with Google totaling 1.17 gigawatts of carbon free energy across the SPP, ERCOT, and PJM regions. Power Alliance

Behind the meter gas grows quickly

When grid interconnection is too slow, developers turn to on site generation. An estimated 50 gigawatts of behind the meter gas generation projects were announced in 2025 alone. BVP The Stargate and Crusoe project in Abilene, Texas includes 360 megawatts of new on site natural gas turbines for an AI campus, announced in January 2025. CRS Report R48762 Natural gas supplied over 40 percent of the electricity used by U.S. AI data centers in 2024. IEA

The sheer size of the pipeline, over 241 gigawatts of global AI data center capacity in development at the end of 2025, shows why the regulatory system is struggling to keep up. Energy IB

What to watch when you draft or finance a PPA

The rapid changes in law and market practice mean that a PPA signed today must account for risks that a PPA signed three years ago did not.

Pick the PPA type early and know who regulates it. The choice between a physical PPA, a virtual PPA, a behind the meter arrangement, or a co-located structure determines whether FERC, a state commission, or ERCOT protocols govern the transaction. For any wholesale sale by a public utility, confirm that the seller has market based rate authorization from FERC and that the PPA rate structure is just and reasonable under section 205 of the Federal Power Act.

Factor in the state large load tariff. If the AI data center will take retail service from a utility in a state that has adopted a large load tariff, the PPA must work within that tariff. Negotiate which party bears the risk of the minimum take or pay and the exit fees. If construction is delayed, who pays the monthly minimum. Also, watch for states that are still developing their tariffs, like Delaware, where interconnection may be paused until the tariff is final.

Managing interconnection risk. The single biggest timing risk for a new AI data center PPA is the generator’s interconnection queue. A PPA should tie key milestones, including the start date for physical delivery and the seller’s termination rights, to actual progress through the queue. If the generator fails to meet a cluster study deposit deadline or cannot achieve site control, the buyer needs a clear exit. If the DOE’s large load interconnection rulemaking moves forward, an AI data center could become an interconnection customer in its own right, which would change the risk profile.

Co-location and the FERC transition window. For projects in PJM, the new transmission service types for co-located loads are still being written. Existing behind the meter customers face a three year transition that ends in December 2028. Any PPA that relies on an existing behind the meter arrangement should address what happens when the transition window closes and the customer must move to one of the new FERC mandated services.

Texas specifics. In ERCOT, the SB 6 backup power requirement and the possibility of emergency curtailment should be reflected in the AI data center’s operational guarantees and in any force majeure or business interruption provisions. The $100,000 interconnection study fee and the financial viability demonstration are upfront costs that must be budgeted.

Lender considerations. Lenders should diligence the following. First, the queue position and study status of any generator whose output supports the PPA. Second, whether the interconnection agreement assigns cost responsibility for network upgrades to the generator or potentially later to the load. Third, the credit support structure, is the PPA backed by parent guarantees or letters of credit, and do those hold up under the large collateral demands of a state tariff. Fourth, for virtual PPAs, lenders need to understand the basis risk between the settlement hub price and the actual retail rate the AI data center pays.

Cost shifting and ratepayer pressure is a growing political risk. Several states have introduced legislation aimed at preventing AI data center energy and grid infrastructure costs from being shifted to residential ratepayers, including measures to create separate rate classes, require special tariffs, and in some cases mandate that AI data centers cover full transmission and distribution costs. NASUCA A PPA that locks in a low price for 15 years may face regulatory risk if the state later imposes additional charges or reclassifies the customer class. Keep an eye on dockets and legislation in the state where the AI data center operates.

Key takeaways

• A power purchase agreement for an AI data center can be physical, virtual, behind the meter, or co-located. That choice determines whether FERC or state regulators have primary oversight.
• The Federal Power Act gives FERC jurisdiction over wholesale sales and transmission. States regulate retail sales. The line is critical and sometimes blurred by arrangements that directly affect wholesale rates.
• PURPA’s mandatory purchase rules rarely apply to the large scale PPAs driving the current AI buildout, because most projects take place in competitive markets where the mandate does not apply.
• Interconnection remains the main bottleneck. FERC Order No. 2023 speeds the queue but demands more money and readiness upfront. The DOE is pushing to let AI data centers enter the queue as interconnection customers themselves.
• In December 2025, FERC ordered PJM to create new transmission service types for co-located loads and set a three year transition for existing behind the meter customers.
• Across the country, state regulators approved 29 large load tariffs in 2025 alone. These tariffs impose minimum 10 to 20 year contracts, take or pay obligations of 75 to 85 percent, large collateral, and stiff exit fees. They are designed to protect other ratepayers from cost shifting.
• In Texas, ERCOT and SB 6 impose a 75 megawatt threshold for large loads, require 50 percent backup generation, a $100,000 study fee, and allow emergency curtailment. ERCOT operates largely outside FERC jurisdiction.
• Global data center electricity demand could reach about 945 terawatt hours by 2030 and 1,200 terawatt hours by 2035, in the IEA base case. PPA prices rose 35 percent in 2024. Hyperscalers are contracting record volumes of nuclear and renewable power, with nuclear PPAs surging for 24/7 carbon free supply.
• When drafting or financing a PPA, allocate interconnection delay risk, account for state tariffs and political risk, and, for co-location projects in PJM, plan for the 2028 transition deadline.

Frequently asked questions

Q:What is the difference between a physical and a virtual PPA?

A:
In a physical PPA, the electricity moves over the grid to the data center’s location. The parties must be in the same grid region. In a virtual PPA, the AI data center does not physically take the power. Instead, the parties settle the difference between the PPA price and the wholesale market price. The AI data center buys electricity from its local utility as usual and receives the renewable energy certificates from the project.

Q:Can an AI data center build its own power plant and avoid utility regulation?

A:
Yes, a behind the meter generator that serves only the AI data center and does not export to the grid may avoid FERC wholesale jurisdiction and state retail tariff rules. However, interconnection rules, siting permits, and environmental laws still apply. If the AI data center ever sells excess power to the grid, it becomes a wholesale seller and must comply with FERC.

Q:What are the main federal laws governing PPAs for AI data centers?

A:
The Federal Power Act gives FERC jurisdiction over wholesale power sales and interstate transmission. PURPA requires utilities to buy from qualifying small renewable and cogeneration facilities at avoided cost, but its mandatory purchase rule does not apply in regions with competitive wholesale markets. FERC orders on interconnection and co-location set the rules for getting new power onto the grid.

Q:Why are state large load tariffs multiplying so quickly?

A:
Utility commissions worry that residential and small business customers will pay for the massive new infrastructure needed to serve AI data centers. Large load tariffs try to ensure that the AI data center itself bears that cost through longer contracts, higher minimum payments, and direct cost assignment.

Q:How do FERC’s interconnection reforms affect a PPA?

A:
Order No. 2023 speeds the queue by grouping projects into cluster studies, but it requires developers to post larger deposits earlier and to prove site control. For a PPA, this means the generator must be well financed and the contract should tie milestones to the generator’s progress through the queue. The DOE’s 2025 proposal to let AI data centers file their own interconnection requests could one day let an AI data center directly pay for the grid upgrades it needs.

Q:What is co-location and why did FERC order PJM to change its rules?

A:
Co-location places an AI data center next to a power plant so it can draw power directly, often without using the public transmission grid. FERC found that PJM’s old rules were unjust and unreasonable because they did not have proper rates, terms, and conditions for this setup, with utilities including Exelon and American Electric Power arguing the arrangements could shift up to $140 million in annual transmission costs to other customers. FERC, Utility Dive, Industry analysis FERC ordered PJM to create three new transmission service types for co-located loads.

Q:Is a PPA in Texas different from one in other states?

A:
Yes. ERCOT is not under FERC jurisdiction. Texas has its own interconnection rules, the Large Flexible Load program, and SB 6, which requires backup power, a study fee, and allows emergency curtailment. PPAs in ERCOT are governed by state law and ERCOT protocols, not federal rate filing.

Q:What are the biggest risks for an AI data center PPA right now?

A:
Interconnection delay is the top risk. It can take over 4 years, with a median of 5 years, to connect a generator, while an AI data center can be built in 12 to 18 months. Regulatory risk is high. New state tariffs, potential FERC large load rules, and cost shifting politics can change the economics after a deal is signed. Market risk, especially the rapid rise in PPA prices, makes locking in a long term price expensive but often necessary.

Q:Do PPAs for AI data centers require FERC approval?

A:
If the seller is a public utility and the PPA is for a wholesale sale, the seller must file the rate schedule with FERC and the rate must be just and reasonable. Sellers that use market based rates must have FERC authorization. Retail PPAs with a local utility are approved by the state utility commission instead.

Q:Can an AI data center use a PPA to claim it is carbon free?

A:
Yes. If the PPA includes renewable energy certificates and meets recognized carbon free criteria, the AI data center can use the RECs to match its electricity consumption. Virtual PPAs are especially popular for this purpose, because they allow the buyer to claim the green attributes even though the physical power mix at the AI data center site may be different.

Subscribe to The Compute Law Brief

The Compute Law Brief is a free weekday newsletter on the law of AI infrastructure across tax, real estate, construction, power, and deals. The big US build markets and federal law. Three minutes a morning. No paywall, and no email gate to read the blog. Subscribe if you want it in your inbox.

Junde Liu, JD, LL.M. (Taxation) candidate at UF Law. Originally published on Compute Law Blog. This article is general information and does not constitute legal advice. Reading it does not create an attorney client relationship. The reader should not act on the basis of any content here without first consulting a licensed attorney in the relevant state. Last reviewed for accuracy May 23, 2026.