The Deliverable Megawatt

During an 82-second sequence of six transmission faults in July 2024, roughly 1.5 gigawatts of data-center-type load disconnected from the grid. The largest sustained simultaneous drop was approximately 1.26 GW. NERC's incident review documents how successive voltage depressions triggered the facilities' protection systems.

The servers did not vanish. Their protection systems disconnected them. But from the grid's perspective, the effect resembled a large nuclear station suddenly arriving as surplus power. Frequency and voltage rose because generation that had been serving those facilities no longer had an electrical load to meet.

The event is useful because it exposes how misleading the ordinary language of data-center development has become. A company asks a utility for one gigawatt as if it were ordering a quantity from a catalog. At that scale, the request is closer to a new piece of the power system.

At which substation? By what date? Will the facility remain flat during the grid's most difficult hours? Can part of its computing work move? Who pays for the transmission upgrades? What happens if only half the campus is completed? What security stands behind fifteen years of minimum payments? Does the announced number describe computing equipment or the whole facility after cooling and electrical losses?

The annual arithmetic is easy. A one-gigawatt campus operating at a 90% load factor consumes 7.884 terawatt-hours in a year.

The difficult question is whether the next megawatt can arrive when and where the campus needs it.

The central claim: Current evidence does not establish a permanent premium for electricity. It shows a time-bounded scarcity premium for a contract-backed firm-service entitlement supported by completed interconnection, deliverable transmission, and accredited supply at a specific place and date—and flexibility determines how much firm service a data center must buy.

That premium does not necessarily belong to the company producing the cheapest electricity. It may accrue to an operating power plant, a substation position, a permitted site, a transmission upgrade, a large transformer, a powered data-center shell, or software capable of moving computation away from constrained hours.

This is also not a disguised argument that AI needs nuclear power. Existing nuclear plants possess several scarce attributes at once, which can make them unusually valuable. New reactors still have to prove that they can arrive on the relevant schedule and at an acceptable cost. During the current construction cycle, natural gas, renewable generation, batteries, grid expansion, and flexible demand are moving faster.

The scarce commodity is not electricity in the abstract. It is certainty in place and time.

The wrong unit

Electricity discussions slide between energy and capacity as if they were interchangeable.

A megawatt-hour measures energy produced or consumed over time. A megawatt measures the rate of production or consumption at a particular moment. Nameplate capacity describes what equipment was designed to produce. Accredited capacity estimates how much that resource can be trusted to contribute during system stress. Deliverable capacity adds the network: can that contribution actually reach the relevant place?

A data center needs all of them. It needs enough annual energy to operate, enough instantaneous capacity during difficult hours, and a physical path connecting the two.

National capacity-factor averages help show why annual energy is an incomplete comparison. In 2025, the U.S. nuclear fleet operated at a 91.0% capacity factor, geothermal at 65.9%, combined-cycle gas at 58.4%, wind at 34.2%, and utility-scale solar photovoltaic generation at 24.4%, according to the Energy Information Administration's tables for non-fossil and fossil resources.

Using those averages, the annual energy of a one-gigawatt campus at a 90% load factor would be equivalent to approximately:

Resource	Nameplate needed for the same annual energy
Nuclear	0.99 GW
Geothermal	1.37 GW
Combined-cycle gas	1.54 GW
Wind	2.63 GW
Solar PV	3.69 GW

These are energy equivalents, not firm-power equivalents. Capacity factor is observed annual utilization, not a reliability guarantee. Gas plants may run below their physical capability because other resources are cheaper. Nuclear plants experience refueling and forced outages. Solar and wind output follow weather and time of day. Every large plant can also be separated from its customer by a transmission constraint.

A portfolio can generate 7.884 terawatt-hours over a year and still fail to serve a one-gigawatt load during a critical evening, winter storm, transmission outage, or multi-day period of weak renewable output.

Cost comparisons have the same limitation. The EIA warns that comparing technologies by levelized cost alone can be misleading because the calculation does not capture their different value to the grid. A cheap megawatt-hour produced in a low-value hour is not the same service as accredited capacity at the system peak. Neither is necessarily deliverable to a specific substation.

The last few megawatts and the last few hours can determine the value of the entire project.

Power has a zip code and a delivery date

The latest demand forecasts are large enough to test that distinction.

Lawrence Berkeley National Laboratory's 2030 reference case projects 649 terawatt-hours of U.S. data-center electricity consumption, equal to 11.8% of national electricity use. Its range is unusually wide: 521 to 843 terawatt-hours, or 9.5% to 15.3%. The International Energy Agency expects global data-center electricity use to rise from approximately 485 terawatt-hours in 2025 to 950 terawatt-hours in 2030. LBNL and the IEA agree that the load is consequential. They also expose how sensitive it remains to hardware shipments, utilization, efficiency, and the pace of construction.

That uncertainty matters because the request queue is not the same as demand.

In June 2026, ERCOT reported 438 gigawatts of large-load connection requests, 89% associated with data centers. That is several times the entire Texas grid. ERCOT's new process screens for site control and creates a pathway for customers willing to curtail. FERC Commissioner David Rosner has separately warned that shopping the same project among utilities, duplicate requests, and speculative applications can distort forecasts. ERCOT's total and Rosner's warning describe an options market, not proof of 438 gigawatts of future consumption.

The supply queue is no more literal. At the end of 2025, 2,061 gigawatts of proposed generation and storage were seeking U.S. grid connections. Projects reaching operation now spend more than five years between their initial request and commercial service. Of the capacity requesting connection from 2000 through 2020, only about 13% had reached operation by the end of 2025; recent projects remain unresolved. LBNL's queue analysis describes a large pool of possibilities, not a warehouse of power ready for delivery.

The better evidence becomes progressively harder to fake:

• Site control
• Posted deposits and collateral
• Executed service agreements
• Customer-funded network upgrades
• Transformer and switchgear orders
• Construction milestones
• Actual energization and measured load

Even a credible project has to reconcile different clocks. NVIDIA describes an annual AI-platform cadence. Virginia's legislative auditors estimate 12 to 18 months for an individual data-center building and five or more years for a complete campus. Distribution-transformer lead times moved from three to six months in 2019 to one or two years or longer by 2024, while large substation and generator transformers can require three to four years. Generation interconnection commonly takes more than five. Major transmission and first-of-a-kind clean-firm projects can take longer still. DOE's transformer discussion shows how an unglamorous piece of electrical equipment can determine whether billions of dollars of computing hardware begin earning revenue on time.

A national surplus of proposed energy does not solve a local shortage on the customer's date.

The deliverability chain

A useful way to think about the market is to stop treating “one megawatt” as a complete product.

A deliverable megawatt is a series system:

Link	Question it answers
Credible demand	Is a real customer prepared to use and pay for it?
Accredited supply	Is capacity available during the relevant stress hours?
Transmission	Can it reach the correct grid zone and node?
Interconnection	Has the customer received an enforceable right to connect?
Equipment	Are the substation, transformer, switchgear, and protection systems complete?
Fuel and permits	Can the assets legally and continuously operate?
Flexibility	What load can be curtailed, shifted, or staged?
Date	Will every required link be ready together?
Contract	Who bears delay, cancellation, underuse, and dedicated-infrastructure costs?

The weakest link caps the project's size. The delivery date is the latest required item on the critical path. A generator without a completed transformer is not deliverable. A signed energy contract without transmission is not deliverable. A promise to shift workloads without telemetry, controls, and consequences is not capacity.

A contract does not create electricity. It makes the forecast bankable.

Co-location illustrates the point. A data center built next to a power plant may appear to bypass the grid. It rarely does. A hypothetical 1,000 MW data center beside a 900 MW generator has a 100 MW net grid withdrawal only while that generator produces 900 MW. If the plant trips, the campus can suddenly require as much as 1,000 MW unless its contract enforces simultaneous curtailment or it has other islanded supply. It still depends on ancillary services, backup arrangements, and upgrades that preserve reliability.

An existing generator with a current capacity obligation cannot abandon it during the delivery period. Afterward, it may retire or contract differently under the applicable rules, but redirecting previously available capacity is not system-neutral; replacement capacity or network work may be required. FERC's December 2025 direction to PJM called for clearer firm and nonfirm contract-demand options while preserving appropriate grid charges. FERC's order makes “behind the meter” a service arrangement, not a synonym for “without system consequences.”

This is why the premium can appear as real-estate value. A powered shell includes years of engineering, permitting, equipment procurement, and utility negotiation. The building can survive several generations of GPUs. Its price capitalizes an enforceable claim on electricity rather than land alone.

Which asset captures the rent depends on which link remains scarce.

The last 100 megawatts

The strongest counterargument to the power premium is that computation does not need to behave like an aluminum smelter.

Some inference workloads are latency-sensitive. Expensive accelerators also create a strong incentive to keep equipment utilized. But model training, preprocessing, software builds, media rendering, and other batch work can sometimes move across hours or regions. A data center with batteries, backup systems, spare fleet capacity, and schedulable workloads is not necessarily a passive flat load.

The arithmetic is powerful.

Suppose a one-gigawatt campus operating at a 90% annual load factor can curtail 100 MW during the 100 most constrained hours of the year. It foregoes or shifts up to 10 gigawatt-hours—only 0.127% of its annual consumption. That percentage describes affected work, not necessarily an energy saving: deferred computation may run later. Yet the commitment can reduce the grid's peak obligation by 100 MW.

Annual energy barely changes. The marginal capacity requirement may change substantially.

If the residual local shortfall is 100 MW, dependable flexibility can cover it and materially reduce the marginal scarcity rent while the other 900 MW remain fully served. Capacity demand curves, reserve requirements, zonal constraints, offer behavior, and price collars mean that covering the shortfall does not guarantee a zero capacity price. This is why apparently small operational changes can still matter more than their energy totals suggest.

Google has demonstrated temporal and geographic workload shifting and has begun incorporating demand response into utility agreements. Its operational account also makes the limitation clear: not every workload can move.

Flexibility deserves capacity credit only when it is:

• Dispatchable during the grid's actual critical intervals
• Metered and telemetered
• Automated or operationally credible
• Available for the required duration
• Backed by contractual penalties
• Modeled for rebound consumption and correlated failure

Without those conditions, “AI is flexible” is a marketing claim. With them, software can function like a power resource: it creates headroom without generating another megawatt-hour.

What can arrive when

The physical grid serving AI will be less elegant than most corporate energy announcements.

For the current 2026–2029 construction cycle, the practical stack is likely to include:

• Existing grid headroom where it still exists
• New solar and wind for low-cost annual energy
• Batteries for ramps, peaks, and short-duration backup
• Natural gas for dispatchable capacity
• Enforceable load flexibility
• Existing nuclear preservation, uprates, and a limited number of restarts
• Transmission upgrades, grid-enhancing technologies, and new substations

These resources solve different problems. Four-hour batteries can shift an evening peak; they do not cover a multi-day event without an energy source to recharge them. Solar and wind can arrive modularly, but their interconnection may not. Gas is dispatchable but now faces turbine queues, pipelines, air permits, transformers, fuel-price exposure, and carbon risk. Existing nuclear is low-carbon and dependable, but most of its output already serves somebody.

The IEA expects renewables to provide a large share of data-center supply growth through 2030 while gas supplies much of the additional firm requirement in the United States. Its 2026 update estimates 15 to 27 GW of onsite gas by 2030, while warning that reliability requirements can force 30% to 70% generation overbuild and that turbine shortages can erode the apparent speed advantage. Onsite generation can reduce transmission dependence, but it does not eliminate switchgear, fuel delivery, permitting, reserves, or grid-stability obligations. The IEA's updated analysis describes a portfolio, not a single winner.

The 2030s option set becomes more interesting.

Enhanced geothermal may be the sharper clean-firm challenger. Google's approved Nevada arrangement with NV Energy and Fervo is designed to add 115 MW of around-the-clock geothermal power. Meta and Sage Geosystems have announced a project of up to 150 MW. These are credible scaling signals, but they also show the distance between today's projects and a repeatable one-gigawatt campus solution. Google–Fervo and Meta–Sage still depend on drilling performance, subsurface conditions, financing, permits, and interconnection.

Advanced nuclear offers potentially larger clean-firm capacity on a less certain clock. DOE estimates first-of-a-kind overnight costs of roughly $6,000 to $10,000 per kilowatt and argues that repeated deployment could lower them materially. Google's first 50 MW Kairos project targets 2030. Amazon's X-energy plans and Meta's TerraPower and Oklo agreements principally target the early 2030s and beyond. Those are meaningful commercial options, not capacity available to clear the current queue. DOE's commercialization analysis makes the need for repeated successful builds explicit.

The winning resource is not the one with the best theoretical cost. It is the portfolio that assembles the complete delivery chain on the customer's date.

Autopsy the clean-firm headlines

Corporate energy announcements often combine several different economic acts:

• Preserving an existing plant
• Reallocating existing output through a contract
• Uprating an operating plant
• Restarting a closed plant
• Developing a future project
• Building genuinely new capacity

Those verbs should never be treated as synonyms.

Meta's twenty-year agreement supporting the Clinton nuclear plant covers 1,121 MW beginning in 2027. It helps preserve a plant that would otherwise lose a state support program, but adds only about 30 MW through an uprate. Meta's announcement is valuable precisely because it allows the existing total and the incremental addition to be separated.

Meta's later nuclear portfolio mixes more than 2.1 GW from Vistra's operating plants, 433 MW of planned uprates in the early 2030s, and prospective TerraPower and Oklo reactors. Microsoft and Constellation target an 835 MW restart of the former Three Mile Island Unit 1 in 2028, subject to NRC review and execution risk. Google's first 50 MW Kairos project targets 2030, while Amazon's X-energy plans principally target the 2030s. Every project may prove worthwhile. None should be counted as new operating capacity before it clears its actual milestones.

The same discipline applies to renewable claims. Annual clean-energy matching can finance valuable new projects and offset a buyer's yearly consumption. It does not prove that the local grid supplied carbon-free electricity during every hour the servers operated. Google itself distinguishes its annual renewable-energy match from its harder goal of 24/7 carbon-free energy because surplus purchases in one time or region do not physically fill deficits in another. Google's explanation is more precise than many descriptions of corporate procurement.

Three questions should be reported separately:

1. Did the buyer finance or contract for clean energy equal to its annual consumption?
2. Was clean electricity physically available in the relevant region during each hour?
3. Could the grid reliably serve the load regardless of the generation mix?

All three matter. They are not the same claim.

The contract is part of the grid

A gigawatt request can cause a utility to build generation, transmission, substations, and dedicated facilities years before the campus reaches full load.

If the facility opens late, completes only one phase, consumes less than forecast, or leaves early, the physical assets remain. A regulated utility may still need to recover their approved costs. Without protections, an optimistic AI forecast can become a long-lived obligation for households and smaller businesses.

This is why the rate case matters as much as the power plant.

Virginia's GS-5 class takes effect in January 2027 for qualifying new customers at or above 25 MW and a 75% load factor. It includes a 14-year obligation, minimum payments generally based on 85% of contracted transmission and distribution demand and 60% of generation demand, and collateral that can reach 60% of minimum charges for new customers without sufficient credit. The tariff contains defined exceptions, but its direction is clear: the Virginia State Corporation Commission is attempting to make the initiating customer—not the captive ratepayer—support infrastructure built around its forecast.

The strongest large-load tariffs use some combination of:

• Site-control and duplicate-request screening
• Application fees and construction deposits
• Phased milestones
• Customer payments as dedicated work proceeds
• Minimum-demand bills
• Long take-or-pay terms
• Collateral or letters of credit
• Exit fees and cancellation reimbursement
• Direct assignment of customer-specific facilities
• Curtailment rights backed by telemetry
• Transparent treatment of shared network costs

These provisions do not prove that every cost is isolated. Generation and regional transmission can benefit multiple users. Wholesale prices and opportunity costs can spill across a market. A dedicated substation may be difficult to repurpose even when a contract covers part of it. Confidential agreements can also make “paying its own way” difficult to verify.

But the existence of the protections reveals the underlying risk. Regulators would not need fourteen-year commitments and enormous collateral if every request were guaranteed to arrive.

The rate case decides whether an AI forecast becomes a shareholder wager or a household obligation.

A megawatt has local liabilities

A megawatt is not truly deliverable until its local consequences are measured and assigned.

National percentages are poor proxies for host-community burden. Virginia data centers represented less than 0.5% of statewide water withdrawals in 2023, yet accounted for between 2% and 21% of water use at six utilities examined by the state's legislative auditors after reclaimed water was excluded. Their statewide land footprint was small, but they represented 20% to 30% of land development in Loudoun and Prince William counties from 2013 through 2021. Virginia's JLARC study shows how a nationally modest industry can become locally dominant.

Water claims require particular care. Figures should distinguish withdrawal from consumption, direct cooling water from power-sector water, potable from reclaimed supply, and host-basin impacts from water consumed elsewhere. Dry or zero-evaporation heat rejection can reduce direct water consumption. Reclaimed water can reduce freshwater withdrawals. Closed-loop liquid cooling reduces consumption only when it is paired with non-evaporative heat rejection, sometimes with an energy-efficiency tradeoff. “Data centers drain towns” is too broad. “Water risk depends on the basin, cooling design, source, and season” is defensible.

Air pollution needs the same discipline. Virginia found that actual 2023 emissions from data-center backup generators were about 7% of permitted thresholds. The maximum allowed by a permit is not the same as measured annual pollution. But turbines operating as primary or bridge generation are no longer merely emergency backup. A one-gigawatt campus supplied at a 90% load factor by gas generation emitting the EIA's 2023 U.S. utility-scale natural-gas fleet average of 0.96 pounds of CO₂ per kilowatt-hour would produce roughly 3.43 million metric tons of operational CO₂ annually, before upstream methane. EIA's emissions factors make the scale visible without pretending the rate describes a new combined-cycle plant, marginal emissions, or every campus's fuel mix.

The local bargain also includes noise, transmission corridors, tax incentives, construction work, permanent employment, and public revenue. Several operators told JLARC that a typical 250,000-square-foot facility may support roughly 50 full-time workers, around half of them contractors, compared with approximately 1,500 people onsite at peak construction. Those are industry estimates, not a universal staffing ratio. The jobs are real, but construction-years and induced employment should not be presented as permanent onsite work. Property-tax revenue can be transformative for a county while sales-tax exemptions and infrastructure commitments remain substantial.

The honest test is site-specific: audited taxes and local payroll against incentives, public infrastructure, water, air, noise, land conversion, and the risk that the promised load never arrives.

Who collects the premium?

The AI power trade is often described as a wager on utilities or electricity producers. The scarce rent can move through a much broader chain.

It may accrue to:

• Landowners controlling sites with power, fiber, permits, and cooling options
• Existing generators with useful interconnection rights
• Nuclear plants whose operating lives or output can be extended
• Gas-turbine, transformer, switchgear, and electrical-equipment manufacturers
• Transmission owners and regulated utilities building approved infrastructure
• Data-center developers controlling powered shells
• Hyperscalers able to sign long contracts and post collateral
• Software providers capable of turning flexible computation into a grid product

The International Energy Agency does not expect a generalized uplift across the entire energy sector from AI demand. It instead identifies more targeted opportunities in electrical equipment, turbines, nuclear supply chains, and constrained local markets. That distinction matters. A generator in the wrong location may capture no AI premium. A transformer supplier or site owner may capture more.

The skeptical question should follow every investment claim:

Which asset captures the rent, for how long, and compared with what alternative?

The cleanest measurement would compare two otherwise identical compute-ready sites—one with an enforceable firm-service entitlement and one without it—and then decompose the difference into capacity value, interconnection rights, avoided delay, permits, fiber, cooling, and contract risk transfer.

Without that counterfactual, “AI power premium” can become a tautology: scarce powered sites are valuable because valuable powered sites are scarce.

The honest countercase

There are serious reasons the premium could be smaller and shorter-lived than this argument suggests.

Demand forecasts may be wrong

LBNL's 2030 range spans more than 300 terawatt-hours. Chip efficiency, model architecture, utilization, inference economics, and project cancellations can all move demand. The 438 GW ERCOT queue demonstrates how little an unscreened request total proves.

Efficiency does not automatically mean lower electricity consumption; cheaper computation can expand use. But rebound should be measured, not assumed. The decisive evidence is realized facility load after efficiency, price, utilization, and demand interact.

Computation can move

Training and batch work can follow cheaper power across hours and regions. Better fiber and orchestration can reduce the value of a particular grid node. If enough workloads relocate whenever power prices diverge, regional scarcity rents will weaken.

The limits are empirical: latency, customer proximity, data rules, fiber, and sunk campuses. They should not be assumed away in either direction.

The grid can catch up

Faster interconnection studies, standardized equipment, new transmission, customer-funded upgrades, and better demand screening can bring the power clock closer to the data-center clock. The premium is not a law of nature. It is the product of a temporarily slow supply response.

Flexible service may be cheaper than firm construction

If a dependable 10% flexible block covers most stress-hour shortfalls, the grid may not need the final 10% of conventional capacity. A small sacrifice of compute timing can destroy the marginal rent without reducing annual AI use very much.

Clean portfolios may outperform new firm plants

Renewables, storage, transmission, and flexible demand may provide acceptable local reliability faster and more cheaply than new nuclear, gas, or geothermal. The fact that annual matching is insufficient does not imply that every facility needs a dedicated thermal generator.

Regulators may reject the bargain

Minimum bills, collateral, cancellation payments, emissions rules, or local opposition can make a speculative campus uneconomic. Strong protections can also entrench the wealthiest hyperscalers because smaller firms cannot support billion-dollar commitments.

The premium exists only if someone is willing and able to bear it.

What would prove this wrong

The argument should weaken or be rejected if several observable conditions emerge:

• Screened and collateralized demand falls below existing headroom, accredited supply, deliverable imports, and qualified flexibility.
• Firm electric service becomes routinely available within the normal 18-to-36-month construction period for large campuses.
• Transformer, turbine, and interconnection timelines return to ordinary infrastructure schedules.
• Utilities accredit enough penalty-backed data-center flexibility to cover most local capacity shortfalls.
• Large workloads relocate in response to regional power-price differences without material latency or commercial penalties.
• Powered-site rent premiums and congestion prices collapse as new infrastructure enters service.
• Actual data-center electricity use settles near or below the low end of current forecasts.
• New generation, transmission, advanced nuclear, geothermal, or other substitutes repeatedly arrive on schedule and close shortages before powered-site premiums become durable.
• Renewable generation, storage, transmission, and flexible demand consistently provide equivalent local service sooner and at lower cost.

PJM offers a present benchmark. Its July 2026 capacity auction cleared at the capped price of $325 per megawatt-day while procuring 6,831 MW less than its reliability requirement. Only 525 MW of new generation and uprates cleared. PJM's report does not prove that data centers caused the shortfall or that the price will persist. It does demonstrate that accredited capacity has a value distinct from annual energy.

At $325 per megawatt-day, an isolated block of one gigawatt of accredited capacity has a gross auction value of approximately $118.6 million per year. Dividing that block by the energy of a perfectly flat one-gigawatt load yields $13.54 per megawatt-hour, but that is not a customer's capacity bill. PJM's 14.7% reserve margin, peak-load allocation, zonal rules, self-supply, bilateral hedges, and other adjustments separate the simple illustration from what any particular customer pays.

The number is not a forecast. It is an illustration of what one market is already charging for one piece of certainty.

The price of certainty

AI companies are moving into a part of the economy where a software request cannot immediately produce more supply.

The bottleneck is a sequence of physical and contractual permissions: a plant capable of producing, a network capable of delivering, equipment capable of transforming, a site allowed to connect, a customer willing to curtail or pay, and a date on which all those pieces become real.

Existing nuclear plants can benefit because they already control much of that sequence. Gas plants are benefiting because they can still be financed and dispatched on a nearer clock. Renewable generation will supply a large share of the additional energy. Geothermal and advanced nuclear are attempting to become the next sources of clean firmness. Flexible computing may reduce how much firm capacity needs to be built at all.

No single technology owns the thesis.

The durable insight is that energy, capacity, location, time, and risk are different economic goods. A project offering only cheap annual energy has not solved the data center's full problem. A powered site in the correct market may be more valuable than a lower-cost plant in the wrong one. A customer willing to disappear during 100 difficult hours may be more useful than another generator that takes five years to connect.

The premium will move as each constraint is relieved. It may eventually disappear.

Until then, the most valuable megawatt is not the one that looks cheapest in a national forecast. It is the one that can actually arrive.

Primary sources

1. United States Data Center Energy Usage Report, 2025 edition — Lawrence Berkeley National Laboratory
2. Key questions on energy and AI — International Energy Agency
3. 2025 Long-Term Reliability Assessment — North American Electric Reliability Corporation
4. Comments on large-load reliability disturbances — NERC
5. Queued Up: 2026 edition — Lawrence Berkeley National Laboratory
6. Transformer supply-chain convening — U.S. Department of Energy
7. PJM 2028/2029 capacity auction results — PJM Interconnection
8. FERC action on large-load integration — Federal Energy Regulatory Commission
9. Electricity rate designs for large loads — U.S. Department of Energy
10. Data Centers in Virginia — Virginia Joint Legislative Audit and Review Commission
11. U.S. data-center energy and water analysis — Lawrence Berkeley National Laboratory
12. Meta and Constellation's Clinton nuclear agreement — Meta
13. Commercializing advanced nuclear reactors — U.S. Department of Energy
14. Google data-center demand response — Google
15. EIA generation capacity factors and carbon-dioxide factors — U.S. Energy Information Administration
16. Meta's nuclear portfolio — Meta
17. Microsoft and Constellation's nuclear restart agreement — Microsoft
18. Google, Kairos Power, and TVA — Google
19. Virginia GS-5 large-load tariff order — Virginia State Corporation Commission