US data center demand is surging — but where, when, and how much?

US data center development and load growth have become a hot topic, from industry conferences to newspaper headlines. There appears to be a wide consensus that electricity demand will increase. This is driven by several factors: The electrification of residential and commercial sectors; the addition of manufacturing that does not strictly rely on high-heat processes – and therefore depends on electricity; and the potential surge in binary loads from the tech sector. However, the range of estimates is wide, and the assumptions behind them vary significantly.

Data center demand drives a significant share of electricity demand forecasts. Announcements of potential technological advancements, like the efficiency of DeepSeek, or signals from the market like reports of data center lease cancellations could prompt re-evaluations of those forecasts. Also, while the increasing adoption of AI in everyday life and the rapid pace of technological innovation have left industry participants confident in electric power growth, there is little consensus on its magnitude (see figure 1).

As such, understanding the drivers behind that growth is key to better understanding what the parameters around that growth should look like. We cannot treat data centers as a homogenous category. Instead, we need to consider how data latency needs will drive siting decisions, how supply chains will support (or restrict) growth, and ultimately, what role evolving regulatory frameworks, both local and federal, will play in bringing new facilities online. All of these factors will influence the national and regional demand curves, with implications not only for basis and nodal pricing, but also for the types of investments that may be required in those regions.

Technological evolution contributes to the wide range of demand estimates in the sector. Advances in technology potentially allow for workload shifting between data centers, which in turn affects the required proximity between facilities and users. Software improvements can optimize compute tasks, minimizing energy demand. Similarly, hardware optimization tailored to usage type can do the same. Cooling needs, which can consume some 40% of energy in a data center, can also vary, depending not only on software and hardware choices but also on advancements in cooling technologies, as seen in current retrofits.

In the following sections, we will break down these key considerations and their potential impact on growth.

The conflation of the terms data center and AI data center underscores some of the complexities of the broader computing and data center market. While AI and high-performance computing (HPC) are becoming more widely used, many more common and less energy-intensive data tasks – like email, simple internet searches, and cloud-based video streaming – may not necessarily require that kind of compute-intensive power. These varied use cases span a wide spectrum of requirements in terms of latency, energy consumption, and facility size to support their function. Among these, latency requirements will be absolutely critical in shaping future demand growth, particularly as data center capacity becomes more regionally distributed than in the past (see figure 2 for an overview of data center types and their characteristics).

Hyperscalers and their large-scale power and site needs have taken center stage in the discussions about the computing market. However, treating that market segment as the prototype for all data center demand means ignoring the differences between hyperscalers and other types of data centers. Differences that ultimately impact the type and scale of computing being done, and thereby power sector demand and the usage patterns of that power.

Both data center square footage and nameplate capacity have increased with recent announcements. Indeed, there have been multiple project announcements for data centers that require more than 300 megawatt (MW) of electricity to operate. Some data center campuses could eventually consume several times that amount, depending on how many facilities are added over time. However, despite the trend, the range of data center types and use cases is still quite varied. Broadly speaking, data centers can be grouped into the categories displayed in figure 2. Each type generally has different total energy consumption. In some cases, the equipment within the facility is specific to its compute use case. For example, equipment for AI workloads will require different processing capabilities than a data center focusing on cloud services, crypto, or other block-chain related workloads. In some facilities, racks within a data center may serve different functions or tenants, each with distinct use cases.

The data center campus represents a democratization of costs in terms of shared maintenance and energy supply. Tenants within the campus may also have specific computing needs. However, not all users will want to share costs or facilities, and some will prefer their own facility. As a result, a campus or even a multi-tenant single center may not be appropriate for all data center players.

With the AI data center category, compute use may be for AI training, or AI inference. AI training is the prerequisite for AI inference and is typically considered more computationally and energy-intensive. In contrast, AI inference involves processing large amounts of data, where the AI model leverages its training to make predictions, create content, or perform similar tasks. While training may be location-agnostic, inference tends to have stricter latency requirements, as it moves from internal research and development into a business-use environment.

As the industry continues to evolve, there may also be some shifts between the categories. The presence of more hyperscaler data centers may eventually preclude the need for some edge data centers, depending on their systems, network architecture, and actual latency requirements.

In some cases, facilities may also undergo retrofits to equip them for denser rack size, including changes to cooling systems to support more compute-intensive tasks while staying within the power capacity of the facility. Such retrofits may allow the facility to more efficiently cool, enabling power savings to be redirected to compute tasks.

Converting a previously non-AI data center to an AI-optimized facility may be possible but will involve additional capital expenditure for cooling, networking, and replacement compute equipment. However, given supply chain, time to power (the time it takes for a new site to receive power from the grid), and power availability considerations, such conversions may be an attractive option.

Publicly available forecasts of data center energy consumption for 2030 vary widely, ranging from 200 terawatt hours (TWh) to 1,100TWh. Beyond broad assumptions and macro trends, other ways of attempting to decode future growth – such as the pipeline of announced capacity, the pipeline of utility data center customers, and the increasing average square footage and capital expenditure of announced facilities – all point to potentially significant growth in electricity demand. However, some of these forecasts appear difficult to achieve due to constraints in component availability, equipment, or power supply.

This wide range of data center energy consumption estimates highlights several core issues. Foundational assumptions about the future of the industry along with data availability are crucial. Publicly available data on consumption and related parameters is not always easy to uncover or current. Some forecasts likely base their assumptions on the estimated power use from training workloads and scale the result, which may not accurately reflect the power use for inference work. Importantly, whether a forecast is demand-led (based on announced data centers multiplied by some power use metric) or supply-led (based on the availability of key equipment required to bring that data center online) will determine the extent to which constraints are placed on estimates.

Indeed, even those with only a casual knowledge of data centers have likely heard one of the most frequently cited comparisons: The power required for a single Chat GPT request is approximately 10 times that for a traditional Google query. While such statistics may be useful barometers for demand, their translation across a vast array of data center use cases may not be representative, given how differentiated demand can be across queries, models, and how highly dependent it is on context. Translating the requirements of a training workload to an inference setting – which would be more akin to a scalable business – may not reflect the power use once additional investments may be made into software and hardware to effectively support operational efficiencies. In addition, without more data, it is unclear how many parameters were used for the query and what type of output was requested (e.g., video versus text). If parsed further, not only are there likely to be differences between training versus inference workloads, but the energy consumption may also be very different when AI evolves toward more agentic use cases and becomes a business and profit-center for tech companies.

Assumptions around the efficiency of demand are critical, the transference of non-latency-sensitive workloads to other markets and the shifting of workloads would all factor into what demand ultimately is. The ability to ramp usage up and down – effectively acting as a flexible rather than static load – will also be critical determining overall power demand.

Indeed, that potential shifting of workloads has led to data center development expanding beyond original core markets. This is due to increasing challenges in securing sites, rising land values, and – more importantly – growing concerns about securing attendant power in a timely fashion.

If we dig a layer deeper, the implications for demand growth are tied to not only how the concentration of capacity within the data center archetypes may shift, but also the geographic siting of these facilities.

The geographic spread of data centers will largely be determined by key siting considerations (see figure 3). These include power availability to run the site, time to power, and land cost and availability – reflecting the crossover between real estate finance and traditional project finance for the data center category. Other considerations include those specific to the infrastructure. For example, while water availability for site cooling has historically been a key consideration, advances in cooling technology have minimized this importance to some extent for some project types. Labor availability – including electrical engineers and technicians for construction, as well as staff for operations and maintenance – is also a key factor, given some of these facilities are full-time staffed once operational. Additionally, sites with lower exposure to natural disasters are preferred to minimize project downtime. Fiber optic connection is essential for data transmission, and proximity to subsea cables and related infrastructure like cable landing stations is particularly attractive for data centers serving both domestic and international markets.

Given the cost of these infrastructure projects, financial considerations are also paramount. States or localities that offer sales and use tax exemptions for equipment can be especially attractive, given some data center equipment is replaced every three to five years due to obsolescence or warranty terms. While some pieces of equipment related to the full infrastructure considered in a data center – like the facility shell and the underlying land – are long-life assets, the equipment within the facility, including servers and storage systems, require regular replacement, often involving significant capital expenditure.

Siting data centers at locations with preexisting high-voltage grid connection – often former industrial sites – has become a strategic way to minimize “time to power.” These sites were previously occupied by users with a high energy-consumption on a square-foot basis. In some cases, this siting strategy has been employed by a different class of data centers, particularly those servicing the cryptocurrency sector.

If we look at development to date, these factors have typically been present in what are considered “primary” and, in many cases, “secondary” data center markets[1]. However, with the rise of AI training workloads, there have been facility announcements outside of areas typically considered primary or secondary markets. This shift underscores the growing importance of time to power, power availability, and cost in siting decisions. As power-related considerations come to the fore, more facilities may choose to “follow the power.” Some of these facilities may be used for AI training and other workloads that are less latency-intensive. However, a wider geographic spread raises questions about workload shifting: Will we see assets with workloads requiring less latency (which usually require proximity to end users) being handled in these non-traditional, more remote locations? And will we see multiple generations of AI training occurring in parallel with the current ongoing cycle for AI models in place today? Latency requirements vary significantly depending on the application. For instance, healthcare applications may demand much lower latency than a standard email or AI inference tasks. As such, edge data centers may still play a critical role in supporting super-high-latency use cases.

[1] Cushman & Wakefield (2025) define the primary markets as Virginia, Atlanta, Phoenix, Chicago, Columbus, Portland, Eastern Oregon, and Silicon Valley. Secondary markets refer to Austin, Toronto, New Jersey and New York, plus Montreal, Sao Paolo, Bogota, Querétaro, and Santiago. https://cushwake.cld.bz/Americas-Data-Center-H2-2024-Update/2/

When combining use case considerations with geographic spread, a market emerges that may deploy different types of data centers in different locations. That eventual differentiation also assumes a continued demand for lower-latency data centers beyond what has been typical in primary and secondary markets, or a broader adoption of workload shifting.

As of this writing, there is no one-size-fits-all approach to regulating how data centers are connected to the power grid , nor to the tariff structures governing their energy use within regulated markets. As power availability and time to power take on an even greater role in siting decisions, so too does a clear understanding of the proverbial “rules of the game.” How large-load users should pay for power – and whether they should contribute to funding infrastructure to support their connections, especially given potential constraints in some regions – have been much discussed topics. Frameworks for colocation and behind-the-meter connections have also been raised in various forums, including at the Federal Energy Regulatory Commission and at various technical conferences.

In some cases, the pipeline of data center projects awaiting energization mirrors the length of the interconnection queue. It is often unclear which projects will ultimately come online and which are more speculative. In general terms, simply tabulating various pipelines is likely to result in a demand estimate that is untenable with today’s infrastructure. There may also be duplication across utility proposals, as projects position themselves to meet demand – prioritizing time to market and cost. This has knock-on effects not only for predicting demand accurately, but also in providing the certainty needed to justify infrastructure investments to serve both data center needs and the ongoing demands of existing sectors.

To qualify serious applicants and substantiate demand, some tariff structures include minimum demand charges related to a high guaranteed percentage payment for dedicated capacity – akin to a take-or-pay contract. These tariffs may also include exit fees or no right to reallocation, to discourage customers from withdrawing. In some cases, proposed tariffs require data centers to contribute to infrastructure investments in the grid. Such tariffs may be assigned on a case-by-case basis.

Supply chain dynamics – and how the broader macro environment may impact tariffs and other trade-related costs – are critical to projects.

Essential construction materials such as steel, aluminium, and copper wire are especially impacted. At the time of writing, tariffs on steel and aluminum are subject to increased import tariffs that are still being negotiated. Copper is so far excluded from reciprocal tariffs, but this may change.

Ultimately, the impact of tariffs will vary depending on the specifics of each supply chain and any warehousing that may have occurred to help developers minimize time to market for projects. However, not all developers – whether building data centers themselves or through third parties – will have the flexibility to diversify supply chains or stockpile materials in advance.

Component obsolescence is another key consideration, specifically for equipment housed within the data center shell – whether this be chips, servers, or server blades. In some cases, the replacements of obsolete components may be driven by warranties rather than obsolescence itself. Supply chain slowdowns linked to Covid-19, which previously affected these categories of equipment, have largely dissipated.

In the past several months, several retired nuclear plants – Palisades, Three Mile Island Unit 1, and Duane Arnold – have announced plans to restart operations in response to growing electricity demand, including demand from data centers. Nuclear license extensions in the US are not new. Current regulations allow the Nuclear Regulatory Commission (NRC) to issue up to two 20-year extensions to a plant’s operating license. Since 2000, more than 90 reactors have received such extensions. However, the NRC considers the restart of a previously shut-down plant a "first-of-a-kind (FOAK) effort,” and the restoration of a plant’s license to operational status requires the commission’s approval. As a result, a clear timeline for these restarts remains uncertain. The restoration process involves ensuring the safety of plant components, completing necessary facility upgrades and inspections, and conducting oversight activities to assess readiness for restart.

An important factor that may influence the prospects for a given facility to move through the NRC process is the decommissioning strategy it adopted after shutdown. There are essentially, although not formally, two decommissioning strategies possible with the NRC. In the safe storage (SAFSTOR) strategy – sometimes referred to as deferred dismantling – a facility is contained and monitored, allowing radioactive material to decay. This approach is sometimes chosen when a facility needs more time to build its decommissioning budget or is waiting for a disposal site to become available. The decontamination (DECON) strategy, on the other hand, involves removal of all fuel and equipment from the facility and is typically carried out on a more expedited timeline. In cases where SAFSTOR is used initially, DECON becomes the final step.

There has been significant discussion on the role of small modular reactors (SMRs) in helping to fulfill data center load. This includes partnerships and other agreements (like memoranda of understanding) between SMR companies and tech companies or data center developers. However, the NRC permitting process will ultimately determine the time to market.

There are currently two primary pathways for obtaining NRC permits: Part 50 and Part 52. Part 50 involves separate construction and operating permits and typically takes approximately six years to complete. This process is better suited for projects that may have ongoing design changes or refinements, making it generally more appropriate for FOAK projects. Part 52, by contrast, is a more streamlined approach best suited for standardized projects with relatively mature designs. It can potentially be completed in about three years. However, Part 52 can result in additional costs and potential changes in project timelines if design modifications require another NRC review process.

A third permitting process, Part 53, has been proposed as a performance-based approach, with a final rule expected by the end 2027. As such, even with the two approved permitting pathways and the proposed Part 53, the deployment of new assets will still be multiple years from now.

Notably, in April, the Canadian Nuclear Safety Commission granted a construction license for the first commercial SMR in North America. Ontario Power Generation received a license for the first of four planned nuclear reactors at the existing Darlington nuclear site.

The data center industry occupies a unique position due to the intersection of the US’s global competitiveness in AI, national security, and data sovereignty. Indeed, executive orders issued under both the current and previous administrations have addressed the siting of data centers on federal lands. While there are some key differences between the executive orders, they underscore that the executive branch views the data center industry as closely linked to national security. While a cat meme or a late-night leg up on a term paper may seem far removed from matters of national security, the infrastructure supporting AI and advanced computing are key priorities. The extent to which AI will also be leveraged for use within critical systems highlights both its obvious ties to tech-related infrastructure and its potential role within broader critical infrastructure. This has important implications for data sovereignty, as well as tangible repercussions for demand growth.

If you have any questions about this report, please contact Susan Hansen.