Electricity demand is once again becoming one of the central questions in U.S. energy planning. After years in which national electricity sales were relatively flat, new sources of load growth are emerging from electrification, manufacturing, and, most visibly, data centers supporting cloud computing and artificial intelligence.
The U.S. Energy Information Administration’s (EIA) Annual Energy Outlook (AEO) 2026 recognizes this shift directly. EIA now reports server electricity use in data centers as a distinct category and includes a High Electricity Demand case that reflects substantially faster growth in both server power draw and installed server stock. That is important work. It gives policymakers, utilities, researchers, and the public a transparent national benchmark for thinking about a rapidly changing part of the energy system.
But the data center question does not end with the standard AEO cases. The pace and shape of AI-driven electricity demand remain highly uncertain. If demand grows faster than EIA’s High Electricity Demand case, how would the U.S. energy system respond? Would the result be mainly more natural gas generation? More renewable generation? Higher capacity additions? Higher emissions? Higher prices?
To explore those questions, NEMS Connect ran a high data center demand stress case using the public AEO 2026 NEMS framework. The case starts from EIA’s High Electricity Demand case and pushes data center/server electricity demand higher in the later years of the projection. The purpose is not to forecast the most likely future, but to evaluate system response under an even higher load-growth assumption for data centers.
The Stress Case
Table 1. Data center demand stress case
| Results for the year 2050 | Units | EIA Reference | EIA High Electricity Demand[1] | NEMS Connect Stress | Stress minus EIA HED |
| Data center server purchased electricity | BkWh | 446 | 818 | 1,086 | 267 |
| Commercial electricity sales | BkWh | 2,298 | 2,774 | 3,047 | 273 |
| Total electricity generation | BkWh | 6,273 | 6,776 | 7,060 | 285 |
| Electric power-sector CO₂ | MMmt CO₂ | 837 | 936 | 986 | 50 |
| Total energy-related CO₂ | MMmt CO₂ | 3,791 | 3,895 | 3,946 | 51 |
| Average electricity price | 2025 ¢/kWh | 14.0 | 14.3 | 14.2 | -0.1 |
| Total electric power-sector capacity | GW | 1,989 | 2,178 | 2,268 | 91 |
EIA’s AEO 2026 High Electricity Demand case already represents a substantial increase in data center server electricity use. In EIA’s AEO 2026 discussion, server electricity consumption reaches roughly 818 billion kilowatt-hours (BkWh) by 2050 at the high end of the case range.
The NEMS Connect stress case increases the late-period data center/server electricity trajectory above that level. In the resulting sensitivity, data center server purchased electricity reaches approximately 1,086 BkWh in 2050. That is about 267 BkWh above EIA’s High Electricity Demand case, or roughly one-third higher for that specific data center server end use.
That is a large increment. For perspective, 267 BkWh is not a marginal modeling adjustment. It is enough additional electricity demand to materially affect commercial electricity sales, total generation, capacity expansion, fuel use, and emissions.
The stress case therefore asks a practical planning question: if data center demand is materially higher than EIA’s high case, how does the integrated U.S. energy system respond?
Demand Growth Appears Where Expected
The demand-side result is straightforward. The added electricity use appears primarily in the commercial sector, consistent with the data center/server assumption being tested.
By 2050, commercial electricity sales in the NEMS Connect stress case are about 273 BkWh higher than in EIA’s High Electricity Demand case. Total electricity sales are about 268 BkWh higher.
That near one-for-one relationship is important. It shows that the stress-case assumption is flowing through the model in the expected way: higher data center server electricity use becomes higher commercial electricity demand and higher total electricity sales.
But the more interesting question is not whether demand increases. It is how the power system responds.
Higher Load Does Not Translate One-for-One into Higher Emissions
A common shortcut in discussions of data center electricity growth is to assume that more load automatically means proportionally more fossil generation and emissions. The NEMS Connect stress case shows why that shortcut can be misleading.
By 2050, the stress case produces roughly 285 BkWh more total electricity generation than EIA’s High Electricity Demand case. That incremental generation is supplied mainly by a combination of natural gas and renewables. Natural gas generation is about 141 BkWh higher while renewable generation is about 132 BkWh higher, and nuclear generation is about 12 BkWh higher. Coal, petroleum, hydrogen, and other generation sources change very little in the comparison.
The emissions result follows from that generation mix. Electric power-sector carbon dioxide emissions are higher in the NEMS Connect stress case than in EIA’s High Electricity Demand case, but the increase is much smaller than the increase in load might imply if one assumed that all incremental electricity was supplied by fossil generation. By 2050, electric power-sector CO₂ emissions are about 50 million metric tons higher than in EIA’s High Electricity Demand case, while total energy-related CO₂ emissions are about 51 million metric tons higher.
That is not a trivial increase. But it is also not a one-for-one emissions response to the added electricity demand.
The result illustrates a key point for policy analysis: the emissions consequences of AI and data center growth depends on the marginal generation mix, capacity expansion, regional constraints, fuel prices, renewable deployment, and system evolution over time. Annual electricity demand is only the starting point.
Capacity Expansion Is Part of the Story
The stress case also shows that additional data center load is not met through only dispatch changes. It leads to additional capacity.
By 2050, total electric power-sector capacity is about 94 gigawatts higher than in EIA’s High Electricity Demand case. The additional capacity includes renewables, combined-cycle natural gas, and combustion turbine capacity.
That result is important because it highlights the infrastructure dimension of data center growth. Data centers are not simply annual energy consumers. They are large, persistent loads. If they are treated as firm demand, the system must plan for both energy and capacity. That means generation additions, reserve needs, transmission considerations, fuel supply implications, and regional reliability planning all become part of the story.
This is especially relevant for utilities, state regulators, regional transmission organizations, and policymakers. The impact of data centers depends not only on how many kilowatt-hours they consume, but also on when and where they consume electricity, how firm that load is, how quickly new capacity can be built, and what resources are available in the regions where the load appears.
Load Shape and Server Economics Are Major Remaining Questions
This stress case retains EIA’s AEO 2026 representation of data center server electricity demand as essentially flat across all hours of the day. That is a reasonable planning assumption for high-availability digital infrastructure. Data centers are built around uptime, redundancy, and high utilization of expensive computing and cooling assets.
At the same time, load shape is a material uncertainty and an important area for further analysis. Some AI training workloads may be schedulable, geographically movable, or responsive to hourly electricity prices or clean-energy procurement signals. Other workloads, especially AI inference and customer-facing cloud services, may be much less flexible. Future data center load could therefore be flat, peak-coincident, partially flexible, renewable-following, or some combination of those patterns.
There are also broader server-economics questions that could affect the magnitude and timing of electricity demand. Server hardware costs, replacement cycles, utilization rates, and the possible repurposing of older servers after their initial economic life could all affect the installed server stock and the amount of electricity consumed. In that sense, server electricity demand is not only a load-shape question. It is also a capital-stock and utilization question.
Another important issue is whether some data center loads are served partly by behind-the-meter or dedicated generation. That could change the relationship between server electricity use, retail electricity sales, transmission and distribution losses, fuel use, emissions, and delivered power costs. These arrangements are already emerging in the market and could become an important part of how large digital loads are supplied.
For this initial stress case, NEMS Connect varied the magnitude of server electricity demand while retaining EIA’s load-shape assumption. That keeps the analysis focused and avoids changing too many assumptions at once.
These questions are well-suited for additional research by universities, policy analysts, utilities, technology companies, state energy offices, and other organizations interested in the energy-system implications of AI load growth. Future work could hold the annual server electricity demand constant while changing only the load shape. It could also test alternative assumptions about server costs, utilization, and behind-the-meter generation. Those sensitivities would help distinguish between the effects of higher annual electricity use, different hourly operating patterns, lower-cost server deployment, and alternative power-supply arrangements.
What the Case Shows
This stress case suggests several preliminary conclusions.
First, data center/server electricity demand can be pushed substantially above EIA’s High Electricity Demand case while still producing a coherent integrated energy-system response.
Second, higher electricity demand does not automatically translate into a proportional increase in emissions. The model response includes additional renewables, additional natural gas generation, capacity expansion, and changes in the generation mix.
Third, data center planning is not only a question of annual electricity consumption. It is also a question of capacity, timing, geography, and firmness of load.
Fourth, EIA’s High Electricity Demand case provides a valuable national benchmark, but the public NEMS framework can support additional targeted sensitivities that go beyond the standard AEO case set.
That last point is important. The Annual Energy Outlook is not intended to answer every possible question. No standard case suite could. Its value is that it provides a transparent, internally consistent foundation from which additional analysis can proceed.
Why NEMS Matters
This is the kind of question NEMS is uniquely suited to examine. NEMS does not treat electricity demand, generation, fuel markets, capacity expansion, prices, and emissions as separate issues. It links them.
That is what makes the model so useful for analyzing data center growth. A higher data center load assumption is not just an electricity-demand assumption. It can affect generation, capacity, natural gas consumption, renewable deployment, regional power-sector outcomes, electricity prices, and CO₂ emissions. Those effects need to be evaluated together.
EIA deserves significant credit for making NEMS and its associated documentation publicly available. As EIA Administrator Tristan Abbey noted in a recent Stanford presentation, EIA’s methods, sources, and assumptions are disclosed, and “anybody can check our work with the right equipment and patience.”
That phrase captures both the promise and the practical challenge of NEMS.
The model is available. The documentation is available. The analytical framework is credible and nationally recognized. But turning the public NEMS release into an operational tool that can answer a specific question still requires equipment, configuration, experience, and patience.
Where NEMS Connect Fits
NEMS Connect exists to help organizations shorten the path to a more accessible NEMS.
Our purpose is not to replace EIA or compete with the Annual Energy Outlook. Quite the opposite. EIA built and maintains the national reference framework for U.S. energy analysis. NEMS Connect helps universities, NGOs, state agencies, companies, and policy researchers use that framework more effectively.
The role of NEMS Connect is to help move organizations from “NEMS is available” to “NEMS is operational and answering our question.”
That can include configuring the model, developing targeted sensitivities, running cases, comparing results, interpreting outputs, and translating model results into decision-relevant analysis. The goal is to make NEMS more usable for the broader energy-policy community while remaining grounded in EIA’s public framework.
The data center stress case is one example. Other organizations may want to examine industrial reshoring, hydrogen deployment, carbon capture, advanced nuclear, regional electricity constraints, fuel-price shocks, electrification, or policy sensitivities. Many of these questions are too specific to be included as standard AEO side cases, but they are exactly the kinds of questions an operational NEMS platform can help answer. For data centers specifically, NEMS Connect hopes this initial stress case encourages others to explore load shape, server utilization, repurposed hardware, dedicated generation, and other assumptions that could materially affect the energy-system response.
Data center growth is not a single-number problem. It is a system problem. NEMS is one of the few tools capable of treating it that way, and NEMS Connect is working to make that capability accessible to more analysts and institutions.
Methodological Note
The NEMS Connect stress case discussed here was run as a preliminary sensitivity using the public AEO 2026 NEMS framework and compared against EIA’s High Electricity Demand case. It is not an official EIA projection and should not be interpreted as a replacement for EIA’s published AEO cases. The objective was to demonstrate how a targeted data center demand assumption propagates through the integrated NEMS framework and to illustrate the kind of applied analysis that can be performed using NEMS.
Works Cited
U.S. Energy Information Administration, Annual Energy Outlook 2026.
[1] Values may not add exactly due to rounding.