A research agenda for understanding infrastructure, technology, market, and policy responses to large load growth
From a sequence of cases to a research agenda
Our recent series began with a simple stress-test question: what happens if data-center server electricity demand rises well beyond the level represented in the U.S. Energy Information Administration’s Annual Energy Outlook 2026 High Electricity Demand case? The first article, Data Centers Are Different: What a High-Demand Stress Case Shows about AI Load Growth, examined how the modeled system responded through changes in electricity generation, generating capacity, retail prices, and carbon dioxide emissions.
The second article, How Technology Costs Shape the Emissions Impact of High Data-Center Load, asked what determines how the system serves that higher load. Combining approximately the same high server load with alternative assumptions about zero-carbon technology costs produced materially different capacity, generation, price, and emissions responses. The result was not that one pathway should be treated as a forecast. It was that load growth alone is not enough to explain the outcome.
The third article, Why Integrated Modeling Matters: From Power Demand to Gas Markets, extended the response beyond the power sector. Different generation choices altered natural gas consumption, which in turn affected domestic production, prices, and LNG exports. That cross-sector propagation is one reason integrated modeling matters.
Together, the cases answer some questions and raise others. The next step is not to select a single number for future data-center electricity use and declare the problem solved. It is to identify the assumptions that govern how the energy system responds, then turn those assumptions into research questions that can be tested, replicated, and improved.
This research agenda is particularly timely because DOE released its draft 2026 National Transmission Needs Study for public comment while this article was being developed. The current comment period provides a concrete example of a research topic with immediate relevant policy relevance, but it is not the only window for useful work. Federal law directs DOE to conduct the Needs Study at least once every three years, and its findings may inform subsequent DOE actions. Because rigorous academic research rarely fits within a short public-comment period, researchers can anticipate recurring assessment cycles, develop and test modeling frameworks in advance, and produce results that are available when future assessments and related policy processes begin.
How the load behaves
Annual electricity demand is a useful starting point, but power systems must continuously balance supply and demand. The same annual consumption can impose very different system requirements depending on its hourly profile, seasonal pattern, coincidence with system peaks, and ability to reduce or shift demand during constrained periods.
The AEO 2026 High Electricity Demand case assumes that data-center server electricity use is relatively constant throughout the day, while associated space-cooling demand is more pronounced during midday. Our higher-load sensitivities extend the modeled quantity of data-center server purchased electricity. That variable should not be confused with total facility electricity use. Cooling, power conditioning, communications, storage, and other supporting loads can change the amount and timing of electricity required at the site.
Future research could test a range of operating profiles: highly stable load, peak-coincident load, seasonal load, renewable-following operation, interruptible service, or partially shiftable computing. The purpose would not be to assume that all computing is flexible. It would be to determine how much flexibility is technically and economically credible, which workloads can provide it, and how utilities or system operators would value it.
This is also where empirical work becomes essential. Researchers need information on realized utilization, workload mix, rack density, equipment efficiency, cooling configuration, power usage effectiveness, and operating response to prices or grid conditions. Announced megawatts describe development ambitions. They do not, by themselves, describe the load that will actually appear on the system or how it will behave after energization.
Where the load appears
A national increase in electricity demand can conceal large regional differences. Data centers may locate where land, fiber, water, tax treatment, generation, or transmission access is favorable. But the electric-system consequences depend on the conditions at the specific point of interconnection and in the surrounding region.
EIA’s AEO 2026 analysis projects the fastest growth in data-center server electricity use in the South Atlantic and West South Central census divisions, which include Virginia and Texas. Those regions differ in generation mix, market structure, reserve margins, transmission topology, gas infrastructure, renewable resources, and the pace at which new facilities can be interconnected. A megawatt of new load therefore does not have a uniform national consequence.
Regional research should examine at least three related questions. First, how does the alternative regional distribution of load change the generation, capacity, price, fuel, and emissions response? Second, can available generation be delivered to the new load under normal and stressed conditions? Third, what new transmission, distribution, and interconnection investments are required, when can they be completed, and who bears the cost and risk if the expected load does not materialize in full?
These questions move naturally across modeling scales. NEMS can represent regional investment and market interactions within a consistent national framework. More detailed transmission-planning, power-flow, production-cost, and resource-adequacy models can test deliverability, congestion, contingencies, and hourly reliability in particular regions. The results can then be translated back into structured NEMS assumptions or sensitivities rather than treated as a separate analytical exercise.
DOE’s draft 2026 Needs Study also provides a concrete example of how integrated modeling could complement more detailed transmission analysis. Because it synthesizes findings developed under different demand projections, regional conditions, policies, and modeling assumptions, selected findings could be translated into a common NEMS scenario framework. Researchers could test alternative regional load-growth, transmission-expansion, transfer-capability, cost, and timing assumptions; and then trace their effects across generation investment, electricity prices, capacity and planning reserve outcomes, fuel markets, emissions, and trade. Detailed transmission models would still be required to evaluate specific constraints and solutions, but NEMS could help reveal the national and cross-sector consequences of the alternatives.
What the system can build, and when
The technology-cost cases showed that long-run costs materially influence how the modeled system meets high load. In EIA’s Low Zero-carbon Technology Cost case, overnight capital costs and fixed operating and maintenance costs for specified technologies decline more rapidly, reaching 40 percent below their Counterfactual Baseline equivalents by 2050. In the High Zero-carbon Technology Cost case, overnight capital costs remain at their 2025 level throughout the projection period.
Cost, however, is not the same as availability. A technology can appear economically attractive in a long-run model and still face limits on how quickly it can be deployed. Turbine manufacturing capacity, transformer availability, interconnection equipment, skilled labor, permitting, financing, site control, and transmission construction can all affect the timing of additions. The DOE’s draft 2026 Needs Study identifies transmission limits as a critical bottleneck to interconnecting new generation and highlights substantial uncertainty in projected data-center load growth arising from supply-chain constraints, regional generation availability, load flexibility, and efficiency improvements.
A useful research design would therefore separate three ideas that are often combined: technology learning, delivered project cost, and deployable capacity. One sensitivity might vary long-run technology costs. A second might impose near-term construction or supply-chain limits. A third might test faster or slower transmission and interconnection expansion. Comparing those cases would show whether the modeled response is driven primarily by economics or by the pace at which physical infrastructure can be delivered.
This distinction also sharpens the resource-adequacy question. The model may add a resource to supply annual energy, meet peak demand, maintain planning reserves, provide flexibility, or compensate for the operating characteristics of other resources. Those functions are not interchangeable. Understanding why capacity is built is as important as reporting how much is built.
How the response propagates
Grid-facing and dedicated supply
Not all data-center electricity arrangements have the same physical meaning. A power purchase agreement is primarily a contractual arrangement and does not necessarily remove the facility’s demand from the grid. Dedicated or co-located generation is a physical supply arrangement, but it may still depend on the grid for backup, balancing, or delivery. Behind-the-meter generation or storage can offset some purchased electricity. Emergency generators ordinarily serve a reliability function and should not automatically be modeled as routine energy supply.
Future sensitivities could distinguish fully grid-facing load from partially self-supplied or dedicated arrangements. The relevant questions include how much electricity is purchased from the grid, when on-site resources operate, whether they can export, which party provides reserves and backup, and how emissions are accounted for. These choices can change utility planning needs even when the facility’s total electricity consumption is unchanged.
Natural gas markets and LNG exports
The earlier cases also demonstrated that the response does not stop at the power plant. Greater reliance on gas-fired generation raises power-sector natural gas demand. In an integrated model, that change can affect natural gas production, regional flows, prices, industrial consumption, and LNG exports. Lower-cost zero-carbon technologies can move those outcomes in the opposite direction.
Further work could identify which producing regions respond, whether pipeline constraints alter the result, how seasonal power demand interacts with storage, and whether changes in domestic gas use materially affect U.S. LNG competitiveness. These questions are particularly well suited to NEMS because the electricity, domestic fuel supply, and international trade responses are represented within a single framework. Regional gas-pipeline or market models could provide additional detail where infrastructure constraints are central.
Policy interactions
The current sensitivities isolate higher server load and alternative technology costs. Policy analysis could extend the sequence by testing clean electricity standards, emissions constraints, tax credit structures, transmission incentives, large-load tariffs, or regional resource adequacy rules. The analytical objective should be to evaluate policies within the system in which they operate, not simply add their expected effects independently. A policy that changes generation economics can also change fuel demand, infrastructure investment, customer costs, and trade.
Using the right combination of models
No single model should be expected to answer every question at every spatial and temporal scale. The important issue is not whether NEMS or a more detailed power-system model is the better tool in the abstract. It is which model is suited to each question and how the models can be used together without losing consistency.
NEMS provides the integrating framework. It is designed to trace linked responses across electricity generation and capacity, end-use demand, fuels, prices, investment, emissions, and trade over a long-term national horizon. That makes it useful for structured sensitivities in which a change in one sector can alter outcomes elsewhere.
Complementary models can add resolution where the research question requires it. Hourly production-cost models can test dispatch, ramping, storage operation, and curtailment. Resource-adequacy models can examine loss-of-load risk and the contribution of different resources during stressed conditions. Transmission planning and power-flow models can assess congestion, deliverability, voltage, stability, and contingencies. Engineering models can represent the electrical behavior of large data-center loads and on-site equipment. Pacific Northwest National Laboratory’s data-center model library illustrates both the value and limitations of generic models: site-specific grid studies may require additional technical data and subject-matter expertise.
The most productive approach is iterative. NEMS can identify system responses and regions that warrant closer examination. Detailed models can test those responses under more specific conditions. Their findings can then inform NEMS inputs, constraints, or scenario design. This preserves the value of national integration while improving the representation of the mechanisms that matter most.
From a research agenda to researcher-owned projects
The high-load cases should be read as a starting point for research, not as a claim that NEMS Connect has resolved the future of data-center electricity demand. They identify a sequence of testable questions:
- How large is the load, and how much of it is represented by data-center server purchased electricity rather than total facility electricity demand?
- When is electricity required, and what operational flexibility is technically and economically credible?
- Where does the load connect, and what generation, transmission, distribution, and gas infrastructure is available there?
- Which technologies are economically attractive, and which can actually be manufactured, permitted, financed, interconnected, and built on the required schedule?
- How do supply choices propagate through electricity prices, reliability, emissions, natural gas markets, and LNG exports?
- How do alternative policies change those responses, including the allocation of infrastructure cost and risk?
- Which questions require national integration, and which require hourly, regional, or engineering detail?
Each question is large enough to support independently led work by a university, research institute, policy organization, or sponsor. A regional study might connect NEMS scenarios to a production-cost and resource-adequacy model. An engineering study might characterize data-center load behavior and translate it into hourly profiles. A fuel-market study might examine the pipeline and LNG consequences of alternative power-sector responses. A policy study might evaluate how large-load tariffs or clean-energy requirements change investment and customer outcomes.
Independent ownership is important. Researchers should be able to frame the question, select methods, document assumptions, test alternatives, and publish conclusions as their own work. NEMS Connect can support access to the public NEMS framework, replication of benchmark cases, sensitivity analysis, and interpretation of integrated results. The substantive research question and its conclusions should remain with the researchers undertaking the work.
The broader lesson
Uncertainty about future data-center electricity use is real, but it should not force analysis to choose between forecasts. A better response is to ask which uncertainties change decisions, represent those uncertainties transparently, and trace their consequences through the energy system.
High load creates the challenge. The system response determines the outcome. That response depends on load behavior, location, infrastructure, technology availability, fuel markets, policy, and analytical resolution. The research opportunity is to separate those elements, study them with the tools best suited to each question, and reconnect the findings within an integrated framework.
The sensitivity cases in this series provide the beginning of that process. The next stage belongs to the broader research community.
Sources
NEMS Connect, “Data Centers Are Different: What a High-Demand Stress Case Shows about AI Load Growth.” https://nemsconnect.com/data-centers-are-different-what-a-high-demand-stress-case-shows-about-ai-load-growth/
NEMS Connect, “How Technology Costs Shape the Emissions Impact of High Data-Center Load.” https://nemsconnect.com/how-technology-costs-shape-the-emissions-impact-of-high-data-center-load/
NEMS Connect, “Why Integrated Modeling Matters: From Power Demand to Gas Markets.” https://nemsconnect.com/why-integrated-modeling-matters-from-power-demand-to-gas-markets/
U.S. Energy Information Administration, Annual Energy Outlook 2026 narrative. https://www.eia.gov/outlooks/aeo/narrative
U.S. Energy Information Administration, Annual Energy Outlook 2026 Case Descriptions. https://www.eia.gov/outlooks/aeo/assumptions/pdf/case_descriptions.pdf
U.S. Department of Energy, Draft National Transmission Needs Study, July 2026. https://www.energy.gov/documents/national-transmission-needs-study-draft-july-2026
North American Electric Reliability Corporation, 2025 Long-Term Reliability Assessment. https://www.nerc.com/globalassets/our-work/assessments/nerc_ltra_2025.pdf
Pacific Northwest National Laboratory, Electromagnetic Transient Modeling of Large Data Centers for Grid-Level Studies, December 2025. https://www.energy.gov/sites/default/files/2026-01/Data_Center_EMT_Models.pdf