Does the UAE Have an Advantage in Building Data Centers?

Large-scale subsidies (e.g. subsidized electricity)

Anticipation of future cost decreases (e.g. due to power plant construction plans) or cost increases in the US

Belief that participation in data center projects will unlock future business opportunities (sales and investments) in the UAE or other countries

Desire to hedge against risk of delays in the US Possible policy implications if this analysis is correct include:

The US may be better positioned to compete internationally than it first seems. It does not appear that the US is a more costly place to build data centers than the UAE. Additionally, the US may have structural advantages in constructing and operating data centers compared to countries like the UAE, including relatively inexpensive energy, a robust domestic data center industry, and locations with more data center-compatible environmental conditions. The US’s greatest cost deficit compared to the UAE appears to be its permitting time. While a significant hurdle, this does not need to be a permanent disadvantage.

Rationales for supporting data center construction in the UAE – substantially beyond current plans – may be somewhat weaker than they initially appear. If the benefits rest on subsidies, they will not necessarily apply to all future projects or be maintained into the indefinite future. Anticipated cost decreases or business opportunities may also fail to fully materialize. As such, it’s possible that recent investments in UAE-based data centers are worthwhile for US companies, but future investments will not be.

In the future, a larger number of foreign countries may be similarly attractive for data center construction. If the benefits of the UAE rest on subsidies, for example, other countries could plausibly provide attractive investment packages. Similarly, if companies are primarily hedging against construction delays, even countries with higher costs could be valuable for ensuring reliable capacity growth. This also suggests that, to the extent the US wants to enable further international data center builds, there are more countries outside the UAE that it could target. The sections that follow provide background on data center construction in both countries, describe the modeling approach, present findings and their policy implications, and then walk through detailed cost estimates. Background Data Center Construction in the US and the UAE Hyperscalers need increasing amounts of training and inference compute. The amount of training compute needed to train frontier AI models has grown 4-5x per year since 2010.11 Keeping pace with the demands of AI training has required the construction of more and larger data centers: since 2020, investment in data centers has increased more than four-fold,12 and the largest AI data centers have gone from 100-200 MW in 2021 to over 500 MW today. To keep pace with user demand for AI models, inference compute has also grown significantly. For example, OpenAI alone spent $1.8 billion on inference compute in 2024 and $7 billion in 2025.14 The infrastructure for frontier AI training is typically provided by “hyperscalers” – companies like Microsoft and Amazon that construct large-scale data centers. To keep up with the growth in compute demand, these companies have invested heavily in rapid development of AI data centers. While frontier AI data centers have historically been built in the US,16 hyperscalers have begun to look internationally for additional locations that could support more AI infrastructure, especially as they have struggled to construct data centers quickly enough in the US.17 The UAE has a relatively modest domestic data center industry, but has begun to position itself as a potential data center hub. The UAE has stated that it views the buildout of AI data centers as a national priority, and is prepared to make significant investments to expand its data center capacity. The UAE government and Emirati companies have recently signed on to various deals with hyperscalers to build frontier data centers in the UAE. These have primarily been facilitated through G42, an Abu Dhabi-based and government-backed technology group,18 which has formed partnerships with companies like Microsoft19 and OpenAI.20 G42 is also part of the Stargate UAE project, an agreement with OpenAI, Oracle, NVIDIA, Cisco, and SoftBank to build large-scale AI data centers in the UAE.21 The scale of these planned investments, including Microsoft’s $15.2 billion22 and 22 Brad Smith, “Microsoft’s $15.2 billion USD Investment in the UAE,” Microsoft, Nov. 3, 2025. 21 G42, “G42 Provides Update on Construction of Stargate UAE AI Infrastructure Cluster,” PRNewswire, Oct. 16, 2025. 20 G42, “G42 and OpenAI Launch Partnership to Deploy Advanced AI Capabilities Optimized for the UAE and Broader Region,” G42, Oct. 18, 2025. 19 Judson Althoff, “Microsoft and G42 Partner to Accelerate AI Innovation in UAE and Beyond,” Microsoft, April 15, 2024. 18 G42 is backed by Mubadala (Mubadala, 2025), an Emirati sovereign wealth fund (Mubadala, 2025). 17 Zachary Skidmore, “US Data Centers Face Grid Bottlenecks as Regional Operators Delay Upgrades,” Data Center Dynamics, Feb. 3, 2025. 16 Epoch AI, “Frontier Data Centers,” Epoch AI, last updated Dec. 12, 2025. 15 Anson Ho et al., “What You Need to Know About AI Data Centers,” Epoch AI, Nov. 4, 2025. 14 Epoch AI, “AI Companies,” Epoch AI, last updated Dec. 6, 2025. 13 Epoch AI, “Frontier Data Centers,” Epoch AI, last updated Dec. 12, 2025. 12 Wilhelm Schmundt et al., “How Investors Can Gain Advantage as the Asset Class Matures,” Boston Consulting Group, March 2025. 11 Jaime Sevilla and Edu Roldán, “Training Compute of Frontier AI Models Grows by 4-5x per Year,” Epoch AI, May 28, 2024. 10 Training compute is the computational power needed to build an AI model (e.g. teaching it to predict text). Inference compute is the power needed to deploy the model to users (e.g. generating responses to queries). AWS’s $5 billion,23 is large by regional standards, although it remains modest compared to global investment levels: global data center capex totaled around $455 billion in 2024,24 and significantly larger planned investments in AI infrastructure in the US have been announced (including Google’s25 and Anthropic’s26 planned investments of $25 and $50 billion, respectively). Total Cost of Ownership Models Estimates are derived using a modified Total Cost of Ownership (TCO) model. TCO models aim to capture both ongoing operating expenses (like paying for electricity) and upfront capital expenses (like building the data center facility). To compare one-time capital expenses to annual operating expenses, upfront costs are converted to annual costs. These annualized capital expenses can then be added to annual operating expenses, so total annual costs can be compared between the US and the UAE. TCO models are a standard method to estimate lifetime costs of a data center. The specific cost components included in the analysis vary, but they typically include operating costs like utilities and labor, and upfront costs like the building construction and IT equipment (see: Koomey, 2006;27 Hamilton, 2010;28 Torell, 202329). TCO analyses of data centers have historically been applied to small, non-AI optimized data centers (e.g. 4.4 MW,30 8 MW,31 5 MW32). Additionally, TCO models have not been applied to the UAE, or to comparisons between countries more broadly. This is the first report to compare total cost of ownership for hyperscale AI data centers in the US and the UAE. Analysis Approach TCO Modeling Approach This report adopts a standard approach to TCO modeling, with one major tweak: in addition to standard capital and operating expenses, the model also accounts for differences in permitting and construction times. Longer permitting and construction processes are more costly, and, all else equal, hyperscalers prefer locations that allow them to build data centers more quickly. 32 Wendy Torell, “TCO Analysis of a Traditional Data Center vs. a Scalable, Prefabricated Data Center,” Schneider Electric Energy Management Research Center, Dec. 19, 2023. 31 James Hamilton, “Overall Data Center Costs,” Perspectives, 2010. 30 Jonathan Koomey et al., “A Simple Model for Determining Total True Cost of Ownership for Data Centers,” Uptime Institute, 2006. 29 Wendy Torell, “TCO Analysis of a Traditional Data Center vs. a Scalable, Prefabricated Data Center,” Schneider Electric Energy Management Research Center, Dec. 19, 2023. 28 James Hamilton, “Overall Data Center Costs,” Perspectives, 2010. 27 Jonathan Koomey et al., “A Simple Model for Determining Total True Cost of Ownership for Data Centers,” Uptime Institute, 2006. 26 Emma Roth, “Anthropic Will Invest $50 Billion in Building AI Data Centers in the US,” The Verge, Nov. 12, 2025. 25 Spencer Kimball, “Google to Invest $25 Billion in Data Centers and AI Infrastructure Across Largest U.S. Electric Grid,” CNBC, last updated July 15, 2025. 24 Dell’Oro Group, “Spending on AI Training Infrastructure Grew by 161 Percent,” Dell’Oro Group, March 19, 2025. 23 Amazon Web Services, “New Report: The Economic Impact of the New AWS Region in the Middle East,” Amazon Web Services, Sept. 20, 2022. In order to compare costs, this report simplifies a few features of a hypothetical data center. First, the size of the data center is fixed at 100 megawatts (MW) of IT power (the power used by the servers, excluding the additional power used to run supporting systems like the cooling infrastructure). This is large enough to capture the scale of training frontier models,33 which currently require data center capacity close to or exceeding 100-MW, while still being small enough that there are existing case studies on which to draw. The report also assumes that the hypothetical data centers use Nvidia H100 chips, for which there is comprehensive documentation and analysis. Additionally, the model only estimates the cost of grid-connected electricity; it does not consider the cost of constructing co-located energy on-site rather than connecting to the grid. Most data centers to date are grid-connected, with about a fifth of data centers in 2024 using some form of co-located energy;35 the share of data centers using only co-located energy (i.e., drawing no power from the grid) is likely significantly smaller. This narrower scope is primarily intended to improve the accuracy of the model, as there are less reliable estimations and greater heterogeneity for the costs of building co-located energy than for the cost of drawing energy from a grid. However, this omits a relevant cost factor that may differ between the UAE and the US: the permitting time for the construction of co-located energy sources. Given faster permitting speeds in the UAE, the exclusion of co-located energy potentially biases the model against the UAE, which would plausibly appear more cost-effective for data centers supported by co-located energy. While the report models the cost of drawing energy from the grid, it does not model the cost of connecting to the grid, a process known as interconnection. For large data centers, interconnection can take substantial time, as grid authorities must review whether the grid can support the project’s electricity demand. In some cases — particularly where grid capacity is constrained or review processes are slow — interconnection may contribute meaningful delays to data center construction timelines. However, data on large-load interconnection times is scarce in both the US and the UAE. Additionally, interconnection requests can be submitted speculatively before a project has begun, making it difficult to translate review timelines into genuine delays. For these reasons, the model does not include interconnection delays. Interconnection may nonetheless be an important barrier for large data center projects and could affect the relative attractiveness of different locations. It is unclear how interconnection timelines compare between the US and the UAE: the US’s larger grid may enable it to accommodate more data center capacity in some regions, but the US likely also has a more backlogged interconnection review process. The potential implications of interconnection delays are discussed at greater length below. 36 While off-grid hyperscale data centers are rare, there are some cases of data centers operating mostly independent of the grid. For example, xAI’s Colossus 2 data center in Memphis relies almost entirely on energy from gas turbines (Ontiveros et al., 2025), which are housed in a retired Mississippi power plant xAI purchased and connected to their data centers. 35 Drew Robb, “Data Centers Bypassing the Grid to Obtain the Power They Need,” Data Center Knowledge, May 1, 2025. 34 While training runs for frontier models will plausibly soon require power draws of several gigawatts (GW), the limited number of case studies of 100-MW+ data center costs would reduce the accuracy of estimates for GW-scale data centers (You and Owen, 2025). 33 Josh You and David Owen, “How Much Power Will Frontier AI Training Demand in 2030?,” Epoch AI, Aug. 11, 2025. Data Estimation Approach Estimates in this report are drawn from a wide range of data sources. When available, government or consortium databases are used (e.g. for electricity rates). However, there are frequently no publicly available academic or government databases on data center cost items, especially for the UAE. Instead, a variety of alternate sources are used, including the following:

News reports. When aggregated datasets of data centers or power plants are not available, news reports and press releases are used for collecting details like construction timelines and capacity. When possible, these are supplemented with other data sources, like satellite imagery. Data center-specific news sources, like Data Center Dynamics,37 are especially useful for tracking data center construction.

Industry databases and analyses. Industry databases like Data Center Map38 and Baxtel39 provide useful regional information on data center capacity, ownership, and location. Analyses from organizations like the Uptime Institute,40 which provides industry standards, regional survey data, and cost estimations, are used as well.

Company filings and reports. Some hyperscalers and data center companies publish reports that provide useful data. For example, some companies release sustainability reports that are useful for estimating water and energy usage.

Satellite imagery. In the absence of robust accounts of data center construction timelines or capacity, satellite imagery (accessed through Google Earth Pro) is used. This is primarily valuable for corroborating construction timelines (e.g. by determining when construction began on a facility). This is especially useful in the UAE, where publicly available reporting on construction timelines is not always available or reliable. This is by no means a comprehensive list. All sources are identified in the detailed cost estimates and supplementary spreadsheet. Findings Total Costs are Comparable in the US and the UAE This report finds that the annualized total costs of ownership (i.e., the sum of annual operating expenses and annualized capital expenses) in the US and the UAE are very similar – less than 1% higher in the UAE. The difference in absolute cost is smaller than the uncertainty on many of the individual components, and should not be treated as significant. Given the relatively small difference 40 Uptime Institute, Research & Reports. 39 Baxtel Advisory. 38 Data Center Map. 37 Data Center Dynamics. and the uncertainty around cost components, small changes in the cost of an individual component could flip which country appears to be slightly cheaper. Individual Cost Components Differ While total costs are similar, certain individual cost components differ significantly (but mostly cancel out). This report considers 12 categories of costs that may vary between the US and the UAE (see Table 1 for a detailed cost breakdown). Table 1 | Key components of a 100-MW data center’s annual total cost of ownership, and an estimate of evidence quality for each component. Total costs are similar between the US and the UAE, though individual cost components differ. The quality of evidence for cost estimates varies by component. This report estimates that the US’s greatest cost advantages are its lower energy prices, lower maintenance costs and faster construction times. Energy prices in the US are lower due both to lower electricity prices and to higher-efficiency data centers (the high heat in the UAE likely reduces efficiency). The heat and dust in the UAE also contribute to higher maintenance costs. The US also appears to construct large data centers faster than the UAE, though it is important to emphasize that this is based on limited data and that there is significant variation between projects. Delays before data centers are operational and earning revenue are costly, so longer construction times are a meaningful determinant of overall costs. This is plausibly due to the fact that the UAE has simply not constructed many large data centers compared to the US – the US41 has about 90 times the number of data centers as the UAE,42 with about 100 times the total capacity, and many more years of experience. The US’s greatest cost disadvantages appear to be its permitting speed and facility construction costs. The UAE benefits from significantly faster permitting speeds, based on estimates of the permitting times for major energy projects in the US and the UAE (permitting times for data centers in the UAE aren’t publicly available). For similar reasons to the time cost of construction, permitting delays are a significant share of the buildout cost. The cost of constructing the data center facility is also lower in the UAE, likely due in part to the availability of cheaper construction labor. Robustness The estimates in this analysis are uncertain and subject to various assumptions; they should be treated as a starting point for further research. Table 1 indicates my relative confidence in each estimate. All components of the TCO model include point estimates, which allow for comparability between components for each country. However, no 100-MW+ data centers have been fully completed in the UAE, and many operating cost estimates are proprietary, significantly limiting the accuracy of these estimates. As such, several estimates rely on a handful of case studies and are highly uncertain. Point estimates are useful for modeling, but should not be interpreted as conveying precision. Additionally, even if the point estimates of average costs by country are accurate, the estimates do not convey the wide range of possible costs a hyperscaler could face within a country. As an example, the model uses the average industrial electricity price in the US as the anchoring estimate for US energy costs, but the cost of grid-connected industrial electricity in the United States can vary by more than a factor of ten. In general, frontier AI data centers, which are not standardized and tend to be highly bespoke, likely vary significantly location-to-location, even within a country. 43 Chelan, WA offers industrial electricity at around 3 cents/kWh (Chelan County PUD, 2025), while certain consumers in Hawaii plausibly pay rates over 50 cents/kWh (the rates for small, medium, and large power use businesses in Lanai are 54.46, 52.19, and 48.48 cents/kWh, respectively) (Hawaiian Electric, 2025). 42 Baxtel Advisory, “United Arab Emirates (UAE) Data Center Market,” Baxtel Advisory, accessed Dec. 18, 2025. 41 Baxtel Advisory, “United States Data Center Market,” Baxtel Advisory, accessed Dec. 18, 2025. Why are Hyperscalers Investing in the UAE? The strong interest by hyperscalers in UAE data centers indicates that the UAE has additional advantages or apparent advantages not captured in this model. There are a few ways to reconcile the modeled lack of apparent cost advantage of the UAE with the strong apparent interest in the UAE.

The estimates are inaccurate, or omit relevant cost factors. The point estimates are intended to represent plausible values for each cost component, but they may be inaccurate as averages and do not capture the wide range of possible costs a hyperscaler might face in a given country. Additionally, important direct costs may be missing from the model. As discussed, this report does not model the cost of interconnection delays as hyperscalers wait to be able to connect to the grid. If interconnection time favors the UAE, then hyperscalers building GW-scale data centers may be rationally investing in the UAE. (The extent to which time-to-power favors the UAE is somewhat unclear. This is discussed in the detailed cost estimates section).

The estimates are accurate today, but may be inaccurate for the future. Building in the US may become more expensive than in the UAE in the future. For example, while the UAE has historically relied on imported natural gas for electricity production, it has begun to transition44 towards nuclear and solar energy, with renewables accounting for around 30% of the UAE’s electricity generation in 2024. As low marginal cost energy sources make up a larger share of the UAE’s electricity production, their energy generation costs may decline. Hyperscalers may be accurately anticipating a relative cost decrease in the UAE in the future.

The “raw” costs, as reported in this model, might not actually reflect what hyperscalers pay. While the nature of contracts for US hyperscalers in the UAE are not public, one potential lever is subsidies. For example, some emirates have announced targeted electricity subsidies,45 such as a 2021 program46 in Abu Dhabi that would provide three tiers of lower rates to economically valuable facilities, with the most valuable facilities receiving rates of around 5 cents/kWh – reducing electricity costs in Abu Dhabi by nearly 40%.47 In addition to subsidizing electricity, the UAE has provided48 government land to data center construction companies and has announced49 significant infrastructure investments to support additional data center capacity. Additionally, while the specific structure of the deals is not known, the UAE’s agreements with companies like Microsoft,50 OpenAI,51 and other hyperscalers52 52 G42, “G42 Provides Update on Construction of Stargate UAE AI Infrastructure Cluster,” PRNewswire, Oct. 16, 2025. 51 G42, “G42 and OpenAI Launch Partnership to Deploy Advanced AI Capabilities Optimized for the UAE and Broader Region,” G42, Oct. 18, 2023. 50 Judson Althoff, “Microsoft and G42 Partner to Accelerate AI Innovation in UAE and Beyond,” Microsoft, April 15, 2024. 49 Yara Abi Farraj, “Abu Dhabi to Invest Over $10 billion to Power Data Center Industry, Says TAQA CEO,” Middle East Economy, May 30, 2025. 48 Abu Dhabi Investment Office, “ADIO Enables Khazna to Boost Abu Dhabi’s Data Economy,” Albawaba, May 30, 2025. 47 It’s unclear whether this program is still active and, if so, whether the same rates still apply. The original plan from 2021 indicated that rates might increase by 3% a year, but it’s unclear if this has occurred. 46 Industrial Development Bureau, “Electricity Tariff Incentive Program: User Guidelines-Existing Companies, v. 3.0” Industrial Development Bureau, accessed Dec. 18, 2025. 45 There is already a base level of subsidization for all electricity in the UAE. The electricity prices modeled for the UAE include this base level of subsidization (which applies to all electricity consumers in the UAE) but does not include the targeted subsidies. 44 Our World in Data, “Electricity Production by Source, United Arab Emirates,” Our World in Data, accessed Dec. 18, 2025. plausibly constitute significant financial support. While both the US and the UAE view data center buildout as a priority, it seems likely that the UAE is more willing to heavily invest government resources to incentivize domestic buildout.

Hyperscalers may value the UAE for other reasons. For example, its geographic proximity to Asia, Europe, and Africa may make it valuable for offering low-latency inference to other locations. Low latency may become particularly important in the future if, for example, advances in robotics require extremely fast connections for deployment. Additionally, investment in the UAE may open up other business opportunities or investments. For example, G42 committed53 to expanding its use of Microsoft Azure and Cloud when Microsoft invested in G42, and the UAE Stargate deal54 involved UAE investment in US Stargate infrastructure.

Hyperscalers want to hedge their bets. Data centers construction projects are getting larger and more complex, with increasing pressure for fast completion times. This probably leads to a higher rate of unexpected hurdles and stumbling blocks than is typical of construction projects. Data center construction delays are costly to hyperscalers, and they may be willing to pay a large premium to reduce the risk of not having capacity ready when they need it. To do so, they may sign on to more projects than they strictly need to minimize the chance of overall delays, especially as it becomes increasingly difficult to find sites that enable large-scale data center projects. Implications The US appears to have certain structural advantages in building data centers, including cheaper electricity, a more hospitable natural environment, and a more robust domestic data center industry. This does not, however, guarantee that the US will maintain its edge. The US could lose its edge, for example, if energy procurement in the US does not speed up quickly enough, and interconnection processes grow prohibitively long. The UAE could also gain an edge on the US. For example, it seems possible that the UAE’s relatively slow construction times on large data centers – which this report estimates as one of the largest cost disadvantages of the UAE – is due to their relative newness to building large data centers. It seems likely that this is something that the UAE could get better at, by building up know-how and infrastructure simply through the process of constructing more data centers. This suggests that it is neither inevitable that the UAE will become a major hub, nor is it guaranteed that the US will continue to be one of the most attractive places to build data centers. If the US wanted to maintain its dominance in data center buildout, there doesn’t appear to be a structural reason that it couldn’t – but it would likely take significant political will to build the energy and infrastructure necessary for large-scale buildout. 54 OpenAI, “Introducing Stargate UAE,” OpenAI, May 22, 2025. 53 Judson Althoff, “Microsoft and G42 Partner to Accelerate AI Innovation in UAE and Beyond,” Microsoft, April 15, 2024. Additionally, these estimates suggest that the UAE’s position is not unique – plausibly, many other countries could be cost-competitive data center buildout locations. To the extent that the UAE has a real advantage, it is likely due to government investment levels, or the fact that hyperscalers expect them to continue to support significant capacity growth. Both of these features are plausibly replicable in other countries. Detailed Cost Estimates The discussion of cost estimates is split into analyses of annual operating expenditures and upfront capital expenditures, listed in descending order of total cost. Lastly, a comparison of operating and capital expenditures is discussed. Additional detail on costs can be found in the supplementary spreadsheet. Operating Expenditures Table 2 presents annual operating expenditures in the US and the UAE. Each component is discussed individually below. Energy constitutes the majority of operating costs, followed by maintenance expenses. Costs like operating labor, rent, water, and taxes are each less than 10% of annual operating costs. This report estimates that electricity and maintenance costs are both higher in the UAE, leading to higher overall operating costs. Table 2 | Annual operating expenditures for a 100-MW data center. Operating costs are higher in the UAE, mostly due to higher electricity and maintenance costs. Components Electricity US: $58 million/year UAE: $69 million/year Electricity expenditures are a function of (1) the price of electricity and (2) the total amount of electricity that is used. Both of these are somewhat higher in the UAE. Electricity price Perhaps counterintuitively, the average price of industrial55 electricity in the UAE is higher than in the US: based on data from their respective federal statistical agencies, industrial electricity costs about 8.5 cents/kWh in the UAE compared to 8.1 cents/kWh in the US.56,57 The UAE, like most Gulf states,58 is oil-rich, but scarce in natural gas. To generate electricity, it mostly relies on imported59 natural gas (oil-fired power plants are comparatively inefficient60). To keep electricity affordable for consumers, all electricity rates in the UAE are subsidized. As such, the values used in this report’s model underestimate the true cost of electricity in the UAE, which is higher; UAE electricity prices are comparable to US prices because of subsidization, not natural energy abundance. Nonetheless, because these subsidies do not specifically target data centers and are standard across the Emirati economy, the baseline subsidized rates are treated as representative values for the UAE in this report. A notable difference between the US and the UAE is the degree of heterogeneity in electricity prices. The UAE has standardized electricity tariffs set at the level of Emirates or groups of Emirates: Sharjah, the North Emirates, Dubai, and Abu Dhabi each have a single rate system. The US has significantly more variation in electricity pricing, with electricity set at the utility level, meaning state-level average prices can obscure the presence of local utilities with notable cheap (or expensive) electricity. In particular, there are several utilities in the US, including Oklahoma Gas & Electric,62 Chelan County Public Utility Department (WA),63 and Grant County Public Utility District 63 Chelan County PUD, “Electric Rate Schedules,” Chelan County Public Utility Department, accessed Dec. 18, 2025. 62 Oklahoma Gas & Electric, “Standard Pricing Schedule: LPL-1 State of Oklahoma”, Oklahoma Gas & Electric, Nov. 26, 2024. 61 Glnar Eskandar, “Missed Opportunities: The Billions Sacrificed Annually to Generate Electricity in the GCC,” Middle East Institute, March 27, 2023. 60 Buck Feng, “Power Plant Efficiency: Coal, Natural Gas, Nuclear, and More (Updated for 2025!),” PCI Energy Solutions, April 17, 2023. 59 U.S. Energy Information Administration, “Country Analysis Brief: United Arab Emirates (UAE),” U.S. Energy Information Administration, last updated Aug. 28, 2023. 58 Glnar Eskandar, “Missed Opportunities: The Billions Sacrificed Annually to Generate Electricity in the GCC,” Middle East Institute, March 27, 2023. 57 Because the UAE does not provide a nationwide average (as the US does), the regional prices (UAE Stat, 2023) are weighted by regional industrial electricity consumption (UAE Stat, 2023) to calculate the average. Additionally, UAEStat, the federal statistical agency, does not include rates for Abu Dhabi, so Abu Dhabi’s rates (ADDC, 2025) are based on a weighted average of the peak and base rates reported by Abu Dhabi’s regional utility. Further detail is in the supplementary spreadsheet. 56 U.S. Energy Information Administration, “Table 2.10. Average Price of Electricity to Ultimate Customers by End-Use Sector, by State, 2024 and 2023 (Cents per Kilowatthour),” U.S. Energy Information Administration, accessed Dec. 18, 2025. 55 The category of electricity rates that data centers fall in varies by utility. Data centers might often fall into business or commercial categories, or, increasingly frequently, their own rate class. Because industrial loads tend to be the largest loads on the grid (Thunder Said Energy, 2025), industrial rates are used for standardization. (WA)64 that offer industrial electricity at around 3 cents/kWh. This means that even though the average industrial rate in the US is only about 10% cheaper than the average rate in the UAE, there are some utilities in the US that provide electricity that is three times cheaper than the average rate in the UAE. See Figure 2. Electricity usage Data centers are not perfectly energy efficient: if the servers use 100 MW, the data center will require more than 100 MW, because additional energy is needed for supporting equipment like cooling systems. Power Usage Effectiveness (PUE), which measures the ratio of data center energy use to the energy used just by the servers, is the standard measure for this additional electricity usage. The lowest possible PUE is 1, indicating a maximally efficient data center that requires no extra energy to support the servers. According to a 2025 Uptime Institute survey,65 the average PUE in North America is 1.49. Due to environmental factors66 like heat, humidity, and dust, a typical PUE in the Middle East and Africa is higher – around 1.68. Larger data centers generally have lower PUEs,67,68 so both regional PUEs are adjusted down to account for efficiency gains from larger facilities. Additionally, data centers don’t typically run at full capacity all of the time; there’s downtime caused by factors like inefficient load distribution or hardware malfunctions. Estimates of the amount of time servers spend operational vary,70 and different workloads have different utilization rates – training loads spend about twice as much time operational as inference loads. Lawrence Berkeley National Lab71 estimates a typical utilization of 80% for training and 40% for inference; this report assumes an average utilization of 60%. The utilization rate is assumed to be the same in the US and the UAE.72 The following formula provides annual electricity cost: Annual cost = Server capacity (MW) x PUE x utilization x Hourly cost ($/MWh) x Hours in a year This produces an annual cost of $58 million in the US and $69 million in the UAE. 72 Specific data centers might be used more heavily for training or inference. Because of uncertainty about what workloads will be run in the US and the UAE, this report assumes an even distribution. Even if actual use skews toward training or inference, the relative costs between the US and UAE will likely remain roughly accurate. 71 Arman Shehabi et al., “2024 United States Data Center Energy Usage Report,” Lawrence Berkeley National Lab, Dec. 19, 2024. 70 Tyler Norris, “The Puzzle of Low Data Center Utilization Rates,” Power & Policy, Aug. 8, 2025. 69 Specifically, the regional values (1.49, 1.68) are respondent-weighted, meaning each data center is given the same weight in the survey, regardless of capacity. To adjust this value for larger capacities, this report applies the ratio of the Uptime Institute’s 2024 estimated global PUE for data centers over 30 MW (1.44) to their global respondent-weighted PUE (1.58) (Davis, 2024). 68 Larger data centers tend to be newer, with more efficient cooling and advanced control systems (Davis, 2024). 67 Jacqueline Davis, “Large Data Centers Are Mostly More Efficient, Analysis Confirms,” Feb. 27, 2024. 66 Joe McCaffrey, “An introduction to Enhancing Efficiencies in Data Centre Construction: A Middle Eastern Perspective,” DMC Global Partners, Nov. 2024. 65 Uptime Institute, “Uptime Institute Global Data Center Survey 2025: Regional View,” Uptime Institute, Aug. 2025. 64 Grant County Public Utility District, “Rate Schedule No. 15: Large Industrial Service,” Grant County Public Utility District, accessed Dec. 18, 2025. Figure 2 | Industrial electricity prices in the US and the UAE. The figure includes Emirate- and state-level average prices, as well as rates in three US utilities with notably cheap electricity rates. The average industrial electricity price is higher in the UAE than in the US. The US also has more heterogeneity in electricity prices. Maintenance US: $12 million/year UAE: $23 million/year Data centers require regular maintenance73 to replace consumable items like filters and batteries, test the functioning of critical systems like the backup power supply, and replace any broken equipment. Using estimates from a data center financial model developed by A.CRE,74 annual maintenance costs are $120,000/MW in North America. Maintenance costs are likely higher in the UAE due to the UAE’s environment. Particulate and gaseous contaminants degrade data center equipment,76 so the higher risk of dust storms in the Middle East increases maintenance requirements. Additionally, data centers in high heat environments tend to require more maintenance. Costly data center components, like uninterruptible power supply (UPS) systems, wear out more quickly when exposed to high heat or dust. Supporting infrastructure, like roadways,80 also degrades more quickly in the UAE due to high heat. It’s unclear exactly how much higher data center maintenance costs are in the UAE. However, according to analysis from the International Facility Maintenance Association, general facility maintenance costs are nearly twice as high in the Middle East as they are in North America. Additionally, for specific components, the recommended maintenance frequency in the UAE is meaningfully higher than that in the US:

Replacing the filters in air handling units (AHUs) is recommended every 1-3 months82 in the UAE, while AHU maintenance is recommended twice a year83 in the US.

Cleaning air conditioning coils is recommended monthly84 in commercial systems in the UAE due to the level of dust, while it’s recommended annually85 in the US. 85 Jackson Mechanical Service, “Best Practices: How Often to Clean AC Condenser Coils for Maximum Efficiency,” Jackson Mechanical Service, accessed Dec. 18, 2025. 84 AC Maintenance UAE, “AC Coil Cleaning Services in Dubai | Expert Technicians You Can Trust,” AC Maintenance UAE. 83 Ryan Waldron, “Air Handling Unit Maintenance That Gets Results,” BalCon, June 19, 2023. 82 Klima Global General Trading, “AHU Maintenance Checklist: UAE Guide,” Klima Global General Trading, accessed Dec. 18, 2025. 81 IFMA, “IFMA Releases First-ever Middle East O&M Benchmarks Report,” IFMA, April 19, 2023. 80 Public-Private Infrastructure Advisory, “UAE Sustainability Partnerships Reduce Waste and Improve Road Quality,” Public-Private Infrastructure Advisory, accessed Dec. 18, 2025. 79 DC Group, “Essential Maintenance Tips for Extending the Lifespan of Your UPS Systems,” DC Group, Dec. 11, 2024. 78 Vertiv, “Vertiv Issues Updated Guidance for Data Centers During Extreme Heat,” July 19, 2023. 77 Helen Morphew, “The Challenges of Designing Data Centres in the Middle East,” BSE 3D, Jan. 21, 2025. 76 ASHRAE, “Particulate and Gaseous Contamination in Datacom Environments, 2nd ed.,” ASHRAE, 2014. 75 Other sources put data center maintenance costs at $100-250,000/MW (CAE Lighting Company, 2025), and general facility maintenance costs are a bit under $2/sqft (IFMA, 2023), for a per-MW cost of around $62,000/MW data center in North America. Calculations are available in the supplementary spreadsheet. 74 Arturo Alejandro Parada Contreras, “Data Center Development Model,” Adventures in CRE, last updated Dec. 2025. 73 Server Room Environments, “Server Room and Data Centre Maintenance Services,” Server Room Environments, accessed Dec. 18, 2025. Given that the maintenance frequency for specific components is at least twice as high in the UAE, and that the UAE is a particularly harsh environment within the Middle East,86,87 applying the International Facility Maintenance Association’s facility maintenance cost ratio to data centers seems reasonable. This report therefore estimates that data center maintenance costs in the UAE are nearly88 twice those in the US. Labor US: $4 million/year UAE: $1 million/year While building a data center can require thousands of construction workers,89 operating a data center requires relatively few employees: BCG estimates that a 100-MW data center requires around 50 full-time employees90 during its operational phase. As of November 2025, postings on Glassdoor indicate that the yearly salary for a data center technician is about $87k91 in the US and $26k92 in the UAE. The labor costs are therefore around $4 million/year in the US, and about $1 million/year in the UAE. Land (rented) US: NA UAE: $3 million/year Based on 20 case studies (available in the supplementary data file), this report estimates that data centers typically require about 0.81 acres/critical MW,93 suggesting that a 100-MW data center would require 81 acres. It’s typically more economical94 to rent than buy land in the UAE, with industrial land frequently offered for lease rather than purchase.95,96 Using an average of advertised land rental rates in three economic “free zones”97,98 in the UAE (Khalifa Economic Zone,99 Jebel Ali 99 Middle East Free Zone, “Khalifa Economic Zone (KEZAD),” Middle East Free Zone, last updated Feb. 12, 2025. 98 About half of industrial rental spaces in Abu Dhabi and Dubai are in free zones (Knight Frank, 2025), which are designated economics areas that allow business full foreign ownership and streamlined regulatory processes (Filings Corporate Services, 2025). 97 Filings Corporate Services, “What Is the Difference Between Mainland and Freezone Company?,” Filings Corporate Services, accessed Nov. 19, 2025. 96 Jebel Ali Free Zone, “Open Plots of Land for Development,” Jebel Ali Free Zone, accessed Nov. 19, 2025. 95 National Industries Park, “Industrial Land for Lease at NIP,” National Industries Park, accessed Nov. 19, 2025. 94 Rajiv Pillai, “Strategic Property Insights: Rent vs. Buy in UAE’s Best‑Value Areas,” Gulf Business, July 8, 2025. 93 This aligns with the typical industry rule-of-thumb of around 1 acre/critical MW. 92 Glassdoor, “How Much Does a Data Center Technician Make in Dubai, United Arab Emirates?,” Glassdoor, accessed Nov. 19, 2025. 91 Glassdoor, “How Much Does a Datacenter Technician Make?,” Glassdoor, accessed Nov. 19, 2025. 90 Vivian Lee et al., “Breaking Barriers to Data Center Growth,” Boston Consulting Group, Jan. 22, 2025. 89 Anson Ho et al., “What You Need to Know About AI Data Centers,” Epoch AI, Nov. 4, 2025. 88 IFMA specifies that the annual maintenance cost in the Middle East “is almost double” the cost in North America. This report assumes that the cost in the Middle East is 190% of the cost in North America. 87 As an example, solar PV plants are cleaned 40-45 times a year in the UAE (PV Tech, 2019), compared to 25-30 times a year in milder Middle Eastern climates like Jordan. Solar PV plants are particularly sensitive to the accumulation of dust or sand, since any accumulation directly decreases performance, so this comparison may exaggerate the discrepancy for data center maintenance. 86 PV Tech, “PV Soiling in Dry Climates: Causes, Impacts and Solutions,” PV Tech, Nov. 2019. Free Zone,100 and UAQ Free Zone101), industrial land rent is roughly $40k/acre/year.102 Assuming 81 acres, a 100-MW data center would cost roughly $3.2 million/year. US land purchase costs are discussed in the capital expenditures section. Water US: $0.5 million/year UAE: $0.5 million/year Data centers use water to maintain optimal temperature and humidity levels in the facility. The standard measure of water consumption is Water Usage Effectiveness (WUE),103 which is measured in liters per kWh.104 A lower WUE (all else equal) is more desirable, as it indicates less water is needed to operate a data center of a given size. WUE varies by design and environment.105

Design: There are different types of data center cooling systems. Some data centers maintain cold temperatures using water-cooled chillers or evaporation-based cooling towers; others use exterior air or closed-loop systems (which recycle water) to maintain cold internal temperatures. Water usage varies by design. Newer models also tend to be more water efficient.

Environment: Hot outside temperatures generally make it more difficult to keep internal temperature levels optimal, generally requiring more water, more energy use, or higher capital expenditures. An analysis106 from the Lawrence Berkeley National Lab found that the average WUE across AI-specialized data centers in the US was 0.61 L/kWh, though it’s worth noting WUE varies107 significantly.108 Data on the average WUE in the UAE or the Middle East is slim, but news reports on data center projects in the UAE indicate that they primarily rely on non-evaporative cooling methods to accord 108 WUE varies, even within companies, due to geographical differences and the use of different generations of cooling systems. For example, Microsoft (Microsoft, 2025) has WUEs ranging from 0.02 L/kWh in Singapore (which uses reused water and mechanical cooling, which is low water-intensity but tends to require greater capital expenditures [Ambros, 2025]) to 1.52 in Arizona (which uses evaporation-based cooling [Microsoft, 2024], which is typically more energy-efficient but less water-efficient [Lawrence Berkeley National Laboratory, 2024]). Average WUE also varies between companies. Amazon has a WUE of 0.15 (Amazon 2025), Microsoft has an average WUE of 0.3, and Google – which trades off its unusually low PUE for a higher WUE – has a WUE above 1 (Patel et al., 2025). 107 EPRI, “Technology Innovation Spotlight: Water Usage in Data Centers,” EPRI, May 2025. 106 Arman Shehabi et al., “2024 United States Data Center Energy Usage Report,” Lawrence Berkeley National Lab, Dec. 19, 2024. 105 Dylan Patel et al., “Datacenter Anatomy Part 2 – Cooling Systems,” SemiAnalysis, Feb. 13, 2025. 104 Unlike PUE, which is a ratio of energy usage, there is nothing special about a WUE of 1, since the numerator (liters) and denominator (kWh) are not measured in the same unit. 103 Dylan Patel et al., “Datacenter Anatomy Part 2 – Cooling Systems,” SemiAnalysis, Feb. 13, 2025. 102 These free zone rental rates probably slightly underestimate actual rental costs, for two reasons: 1) free zones may be less expensive than mainland areas 2) the rents cited here are the advertised “starting at” rents, which are presumably a lower bound on rental rates. However, absent publicly available data on UAE data center land costs specifically, these rates provide a reasonable baseline. 101 Umm Al Quwain Free Trade Zone Authority, “Industrial Land for Lease in UAE: Umm Al Quwain Free Trade Zone,” Umm Al Quwain Free Trade Zone Authority, accessed Dec. 18, 2025. 100 JAFZA, “Plots of Land,” Jebel Ali Free Zone Authority, accessed Dec. 18, 2025. with the country’s hot, dry climate.109 Using estimates from the Lawrence Berkeley National Lab110 on the WUE of different methods, this report estimates that the average WUE in the UAE is around 0.3 L/kWh. While this WUE lowers water costs, it likely comes at the expense of less energy efficiency, which translates to the UAE’s higher PUE. Based on federal data sources, water costs about 0.15 cents/liter in the US111 and about 0.33 cents/liter in the UAE.112 The following formula provides annual water cost: Annual cost = Server capacity (KW) x PUE x utilization x WUE (L/kWh) x Hourly cost ($/kWh) x Hours in a year Using WUEs of 0.61 and 0.3, respectively, indicates that the annual water cost is about $0.5 million in both the US and the UAE (the higher water cost and lower water usage in the UAE roughly cancel out). Taxes US: $0.3 million/year UAE: $0.2 million/year Property taxes The percentage rates of property taxes or fees are likely slightly higher in the UAE, but are offset by the lower cost of buying or renting land. In the US, the average commercial property tax113,114 is 1.755%. The UAE does not levy property taxes,115 but Emirates levy municipal fees on rented properties. This report estimates that hyperscalers would face an average municipality fee of 7.5% in the UAE.116 Using US property purchase prices and UAE rental prices, annual property taxes/fees are estimated at around $0.3 million in the US and around $0.4 million in the UAE. Sales taxes Because of exemptions, hyperscalers likely pay little in sales tax in either the US or the UAE. According to analysis from Husch Blackwell, most states in the US exempt data center equipment from sales taxes.117 Although some states levy taxes or fees on electricity consumption, these are 117 Husch Blackwell, “Tax Incentives for Data Centers: 50 State Survey,” Husch Blackwell, accessed Dec. 18, 2025. 116 The vast majority of Emirati data centers are built in Dubai or Abu Dhabi (Data Center Map, 2025). As such, this report takes a simple average of municipality fees in Dubai (10% for commercial properties) (CRC Property, 2025) and Abu Dhabi (5%) (Property Finder, 2025). Further detail is in the supplementary spreadsheet. 115 Global Property Guide, “Guide to Property Taxes in United Arab Emirates,” Global Property Guide, last updated Sept. 4, 2025. 114 Based on an average of the largest cities in each state for a commercial property worth $1 million. 113 Lincoln Institute of Land Policy and Minnesota Center for Fiscal Excellence, “50-State Property Tax Comparison Study: For Taxes Paid in 2022,” Lincoln Institute of Land Policy and Minnesota Center for Fiscal Excellence, August 2023. 112 UAE Stat, “Water Tariff by Authority, Use Slab and Sector,” UAE Stat, accessed Dec. 18, 2025. 111 United States Environmental Protection Agency, “Data and Information Used by WaterSense,” United States Environmental Protection Agency, last updated July 3, 2025. 110 Arman Shehabi et al., “2024 United States Data Center Energy Usage Report,” Lawrence Berkeley National Lab, Dec. 19, 2024. 109 Further detail on these estimates is available in the supplementary spreadsheet. already accounted for in the retail electricity prices reported by the US, so are not included here.118 In the UAE, although the 5% VAT applies to equipment as well as electricity and water119 consumption, businesses are usually able to recover these “input tax” charges,120 meaning businesses can get refunds for VAT charges on their expenses (or offset other taxes).121 As such, sales taxes are likely close to zero for data centers in both the US and the UAE. Other Utilities & Services Bandwidth The largest external networking cost is likely the installation of fiber-optic cables, which is discussed under capital expenditures below. Data on hyperscalers’ annual bandwidth costs is scarce. However, annual bandwidth expenses are likely small and so are assumed to be 0. This probably slightly understates annual operating costs, but it’s unlikely that recurring bandwidth costs would vary sufficiently between the US and the UAE to significantly change the difference in total costs. Software licensing Software licensing is bundled with the upfront price of chips for H100s,122 so is assumed to be 0 as well. Other software costs are likely marginal. Capital Expenditures Before a data center is operational, significant expenditures are required to build the facility. Unlike operating expenses, which are recurring, capital expenditures are one-time, upfront costs. To compare these costs to operating costs, capital expenditures are annualized — converted into annual costs. The annualization of the capital expenditures is discussed at the end of this section; all costs referenced in individual component sections are non-annualized. Table 3 presents upfront (non-annualized) capital expenditures in the US and the UAE. Each component is discussed individually below. 122 NVIDIA, “NVIDIA Enterprise Licensing Guide,” NVIDIA, accessed Dec. 18, 2025. 121 Federal Tax Authority, “Taxable Person Guide For Value Added Tax, issue 2,” Federal Tax Authority, June 2018. 120 Tushar Agarwal, “How to Recover VAT on Business Expenses in the UAE,” ProTax Advisors, Feb. 24, 2025. 119 Ed Clowes, “UAE to Tax Water and Electricity at 5 Per Cent,” Gulf News, last updated March 11, 2020. 118 U.S. Energy Information Administration, “Does EIA Publish Electricity Sales and Price Data by State and by Utility?,” U.S. Energy Information Administration, last updated Feb. 6, 2024. Table 3 | Capital expenditures for a 100-MW data center. Capital costs are similar between the US and the UAE, largely due to the fact that servers and network infrastructure account for the majority of capital costs and do not vary significantly by region. Facility construction appears to be more costly in the US. Components Servers US: $2,638 million UAE: $2,637 million Servers constitute the largest capital expenditure. Assuming the data center uses Nvidia’s DGX H100 systems,123 which each have a power density of 10.2 kW,124 a 100-MW data center could support around 9,800 DGX H100 systems. According to a SemiAnalysis estimate, the cost of a DGX H100 system is about $269,000.125,126 The total base cost is therefore around $2.637 billion. This cost is 126 This estimate is from 2023, and prices may have increased. Nvidia does not directly sell DGX systems; DGX systems are sold by Nvidia partners like Dell or HPE (Nvidia, 2025). As such, current prices cannot be directly verified with Nvidia. However, a few third-party resellers list current prices (these, as well as other estimates of DGX system costs, are available in the supplementary spreadsheet). The average listing price of these sellers is around $380,000. Hyperscalers probably would get some bulk discount, and would not pay the listed price on individually sold units. As such, the SemiAnalysis estimate may still be roughly accurate. Regardless, prices will not differ between the US and the UAE, so errors in this value will not significantly affect the overall comparison. 125 Dylan Patel and Gerald Wong, “AI Server Cost Analysis – Memory Is The Biggest Loser,” SemiAnalysis, May 29, 2023. 124 NVIDIA, “NVIDIA DGX H100/H200 User Guide,” NVIDIA, accessed Dec. 18, 2025. 123 One DGX system contains 8 GPUs. constant regardless of the location of the data center. However, there are associated costs that may vary between locations, including shipping and import duties. Shipping Freight rates from Taiwan to the US127 are about 60% more expensive than rates from Taiwan to the UAE.128 The size129 of a single DGX H100 indicates that about 46 shipping containers would be required to ship 9,800 DGX H100s. This would cost about $190,000 to ship to the US, and $120,000 to ship to the UAE. Shipping costs, while significantly higher for the US compared to the UAE, are a rounding error on the total cost of chips. Import duties Neither the US nor the UAE currently have import duties on semiconductors or servers.130 Under the GCC Unified Customs Tariff system, semiconductor servers have a 0% tariff rate.131 Similarly, the US currently exempts132 semiconductors and servers from reciprocal tariffs. However, the US has indicated133 that it may implement tariffs on certain semiconductor products, so the cost of import duties may change going forward. Facility US: $1,054 million UAE: $924 million After the servers and network infrastructure, the largest data center construction costs come from the building itself, plus cooling systems, uninterruptible power supplies, and other mechanical and electrical support components. Based on data from Turner and Townsend’s 2025 Data Centre Construction Cost Index,134 the cost of constructing a data center facility is about $9.2/watt in the UAE.135 The cost in different US markets varies, from around $9.5/watt in Charlotte or Dallas to over $13/watt in Silicon Valley, with an average of around $10.5/watt.136 See Figure 3. Based on these estimates, the construction of a 100-MW data center would cost $1.05 billion in the US and $0.92 billion in the UAE. 136 The average US cost is calculated by weighting regional costs by CBRE’s data on the distribution of data center capacity under construction (CBRE, 2025). 135 Turner and Townsend’s estimate includes the cost of “shell and core, architectural fit-out and finishes, mechanical and electrical fit-out, general contractor preliminaries, general contractor margin, general contractor contingency, and mechanical and electrical equipment” (Turner & Townsend, 2025). This includes the cost of the uninterruptible power supply (UPS) and the cooling system. The cost estimates exclude “client direct costs, land purchase costs, utility works, abnormal groundworks, site works, active IT equipment, fibre cabling to support office fit-outs or professional services fees.” Their estimates are restricted to data centers between 30 and 50 MW of IT capacity. While smaller than the 100-MW facilities considered in this report, the per-watt estimates are likely still applicable. 134 Paul Barry et al., “2025-2026 Data Centre Construction Cost Index,” Turner & Townsend, accessed Dec. 18, 2025. 133 Jeff Mason and Kanishka Singh, “Trump to Impose Tariffs on Semiconductor Imports from Firms not Moving Production to US,” Reuters, Sept. 5, 2025. 132 U.S. Customs and Border Protection, “CSMS # 64724565 - UPDATED GUIDANCE – Reciprocal Tariff Exclusion for Specified Products,” U.S. Customs and Border Protection, April 11, 2025. 131 Federal Authority for Identity, Citizenship, Customs & Port Security, “2022 ﺔﻓرﻌﺗﻟاﺔﯾﻛرﻣﺟﻟادةﺣوﻣﻟادولﻟسﻠﺟﻣونﺎﻌﺗﻟا ” Federal Authority for Identity, Citizenship, Customs & Port Security, 2022. 130 Semiconductors and servers are classified under codes 8471 or 8542. 129 NVIDIA, “NVIDIA DGX H100/H200 User Guide,” NVIDIA, accessed Dec. 18, 2025. 128 Freightos, “Shipping from Taichung City to Dubai: Air, Sea & Container Freight,” Freightos, accessed Dec. 18, 2025. 127 Freightos, “Shipping from Taichung City to Los Angeles, CA: Air, Sea & Container Freight,” Freightos, accessed Dec. 18, 2025. The cost discrepancy is plausibly due to differences in construction labor costs in the US and the UAE. McKinsey estimates137 that around 35% of data center facility construction costs over the next five years will be associated with labor costs.138 Construction labor costs are much higher in the US: Turner and Townsend’s 2025 Global Construction Market Intelligence report139 estimates that average hourly construction wages140 are about $76/hour in the US and only $6/hour in the UAE.141 This significant difference in construction labor costs may account for a large part of the discrepancy in facility construction cost. Figure 3 | Data center facility construction cost by location. The per-watt construction cost is higher in the US than the UAE. 141 The cost discrepancy may be smaller for data center construction labor, which likely requires more skilled labor and may require the UAE to pay more internationally competitive wages. 140 Turner and Townsend’s estimate of wages accounts for overhead and associated costs, including: “allowances, taxes, annual leave cost and where paid by the employer, workers’ compensation and health insurance, pensions and travel costs and fares” (Turner & Townsend, 2025). 139 Neil Bullen, “2025 Global Construction Market Intelligence,” Turner & Townsend, July 2025. 138 Labor + Shell and Site + Electrical and Mechanical Equipment = $0.6 + $0.3 + $0.8 = $1.7 trillion. $0.6 trillion/$1.7 trillion = 35%. According to Engineers and Architects of America, labor costs typically account for 20-40% of construction costs, so McKinsey’s estimate accords with typical construction labor costs (EAA, 2025). 137 Adam Barth et al., “The Data Center Balance: How US States Can Navigate the Opportunities and Challenges,” McKinsey & Company, Aug. 8, 2025. Network Infrastructure US: $442 million UAE: $446 million GPUs in a data center need to be able to communicate with each other. They are connected through a vast network of cables, switches, transceivers, and power distribution units (PDUs). Using Nvidia specifications142 for DGX networking requirements, this report estimates that networking capital expenses of a 100-MW data center are around $442 million. The complete estimation is discussed in the supplementary spreadsheet. Import duties Neither the US nor the UAE have import duties on switches or transceivers. While the US also exempts cables and PDUs, the UAE subjects PDUs and cables to a 5% import duty.143 PDUs and cables constitute a small share of the total networking cost, so the total cost in the UAE is only about $4 million higher than in the US. Land (purchased) US: $20 million UAE: NA While land is typically rented in the UAE, it’s more likely to be purchased in the US. Cushman and Wakefield estimate that data center land in the US sold for an average of $244,000/acre in 2024.144 Using the 0.81 acre/critical MW conversion (discussed above), this suggests that the total land cost for a data center is around $20 million. Utility works US: $10 million UAE: $33 million Substations Data centers use substations to take high-voltage power from the grid and step it down to a lower voltage for use in the data center. Substation needs vary depending on transmission voltage and the redundancy level of the data center: a data center connected to medium voltage distribution lines145 might not need a substation at all, while some data centers connected to high voltage transmission lines may have multiple substations (including for redundancy).146 This report assumes that the data 146 David Rupert, “Data Center Uptime Requires Substation Redundancy,” Data Center Dynamics, June 20, 2024. 145 Dylan Patel et al., “Datacenter Anatomy Part 1: Electrical Systems,” SemiAnalysis, Oct. 14, 2024. 144 Brian Ungles et al., “2025 Data Center Development Cost Guide: United States,” Cushman & Wakefield, 2025. 143 Federal Authority for Identity, Citizenship, Customs & Port Security, “2022 ﺔﻓرﻌﺗﻟاﺔﯾﻛرﻣﺟﻟادةﺣوﻣﻟادولﻟسﻠﺟﻣونﺎﻌﺗﻟا ” Federal Authority for Identity, Citizenship, Customs & Port Security, 2022. 142 NVIDIA, “NVIDIA DGX SuperPOD: Next Generation Scalable Infrastructure for AI Leadership Reference Architecture Featuring NVIDIA DGX 100: Major Components,” NVIDIA, accessed Dec. 18, 2025. center has a single substation. Substation requirements are unlikely to vary significantly between the US and the UAE, since the voltage levels are likely similar.147 Data centers in the UAE likely use substations with gas-insulated systems (GIS), which cost more than the air-insulated systems (AIS) that are more common in the US. Gas-insulated systems148 are enclosed against harsh environmental conditions, making them better suited for desert environments. While the frequency of GIS use in the UAE isn’t public, press releases indicate the use of GIS substations rather than AIS systems in the UAE.149,150,151,152,153 In the US, by contrast, less than 5% of substations are GIS; the rest are AIS.154 Figure 4. A gas-insulated substation at a Khazna data center campus in Ibn Battuta (left),155 and an air-insulated substation at Meta’s Prometheus data center in Ohio (right). Satellite imagery is from Google Earth Pro, annotated by the author. Meta’s substation was identified with the aid of Epoch AI’s satellite analysis.156 Based on several cases where contract prices were published,157 the average cost of building a GIS substation in the UAE is around $32 million.158 In the US, the Midcontinent Independent System Operator (MISO) estimates that the construction of new substations costs an average of $7.9 million 158 In the rare case where GIS substations are built in the US, the price seems comparable: a 2015 construction of a GIS substation in the US cost $39.5 million (Burns & McDonnell, 2025). 157 (Siemens, 2012), (Mechanical Electrical Plumbing, 2021), (Dubai Electricity & Water Authority, 2020). As discussed in the previous note, a majority of Dubai’s substations are 132 kV, so these cases are all of 132 kV substations. 156 Epoch AI, “Meta Prometheus: Frontier Data Centers Satellite Explorer,” Epoch AI, Oct. 14, 2025. 155 Khazna, “Our Data Center Network,” Khazna, accessed Nov. 1, 2025. 154 Donald Parnell, “An Introduction to Gas Insulated Electrical Substations,” CED Engineering, 2019. 153 MEED Editorial, “DEWA Awards Substation Contracts,” Middle East Business Intelligence, Dec. 10, 2004. 152 WinStar Group, “DEWA Project 132kV ESKYN Substation GIS High Voltage & PD Measurement Work Completed,” LinkedIn, December 2024. 151 Siemens AG, “Siemens Energy wins €26 million contract to build high voltage Substation for DEWA,” Siemens AG, March 13, 2012. 150 Linxon, “Linxon Commissions Dubai’s Abraj Substation,” Linxon, Jan. 24, 2024. 149 Dubai Electricity & Water Authority, “ABB Briefs DEWA on Progress in 400 kV Substation at Mohammed bin Rashid Al Maktoum Solar Park,” Dubai Electricity & Water Authority, Dec. 19, 2015. 148 Donald Parnell, “An Introduction to Gas Insulated Electrical Substations,” CED Engineering, 2019. 147 Though transmission networks in both countries use a variety of voltages (and thus a variety of substations types), voltages seem generally comparable. A majority of substations in Dubai operate at 132 kV (meaning their input is 132 kV and they step it down to a lower voltage) (Dubai Electricity & Water Authority, 2023), and a plurality of transmission lines owned by American Electric Power (the owner of the largest transmission system in the US) are 138 kV (meaning data centers connecting to these lines would need to interconnect at 138 kV) (Grady, accessed Dec. 18, 2025). for substations in the 115kV to 161kV range.159 As such, this report estimates that substation cost $32 million in the UAE and $8 million in the US. Bandwidth This report estimates that the capital expenditures associated with bandwidth are higher in the US, though small in both countries. The cost of networking between GPUs within a data center is covered by upfront IT costs. However, data also flows into and out of data centers 1) to interact with the internet – for example, to service users; and 2) to connect to other data centers – for example, to save model weights.160 Most traffic leaving or entering a data center falls in the second category.161 As such, this report focuses solely on estimating inter-data center bandwidth costs. Inter-data center connectivity typically involves the construction of fiber-optic cables between data centers. The construction costs depend on both the distance between data centers and the conditions in which the cable is being installed (including environmental factors and labor costs). The following estimates attempt to capture both values for the US and the UAE. According to the Fiber Broadband Association’s annual cost report, installing one kilometer of fiber-optic cable costs around $60,000 in the US, with about 60-80% of the cost of fiber installation associated with labor costs.162 To estimate installation costs in the UAE, this report uses Turner and Townsend’s estimate that average hourly construction wages are about $76/hour in the US and $6/hour in the UAE (as discussed above).163 Assuming the remaining 20-40% of costs are the same, this produces an estimated per-kilometer installation cost in the UAE of around $20,000. Estimating the length of cable that a hypothetical data center will build is somewhat arbitrary. Since data centers may attempt to locate close to large power plants such that inter-data center distance may mirror the distance between plants,164 this report uses the average distance between large power plants in a country as an empirical basis for cable length. Based on a rough estimate, the average distance between power plants appears to be similar between the US and UAE: about 32 km in the contiguous US, and 27 km in the UAE. Further detail on this estimation is in Appendix B. Using these distance and installation cost estimates produces a total fiber installation cost estimate of $1.9 million in the US and $0.5 million in the UAE. 164 Jaime Sevilla and Anton Troynikov, “Could Decentralized Training Solve AI’s Power Problem?,” Epoch AI, Oct. 8, 2025. 163 Neil Bullen, “2025 Global Construction Market Intelligence,” Turner & Townsend, July 2025. 162 Fiber Broadband Association and Cartesian, “Fiber Deployment Cost Annual Report 2024,” Fiber Broadband Association and Cartesian, 2024. 161 While specific volumes are not publicly available, Meta (McKillop and Mediavilla, 2025) and Google (Hong et al., 2018) have reported much greater traffic volumes and faster growth in their data center-to-data center traffic compared to data center-to-internet traffic. Meta’s engineering team has noted that the emerging connectivity demands of AI training continue to be dominated by inter-data center traffic (McKillop and Mediavilla, 2025). Additionally, data center-to-data center traffic volume may grow at a faster rate as hyperscalers try to avoid power constraints: it’s generally cheaper to be farther from a connection hub but closer to energy (Thunder Said Energy, 2024), making it plausible that data centers will increasingly locate in remote areas close to energy but requiring longer fiber-optic cables to connect data centers. Epoch AI has demonstrated that it’s feasible to have decentralized training in order to circumvent power constraints, which would require significant fiber-optic cable buildout (Epoch AI, 2025). 160 Mikel Jimenez Fernandez and Henry Kwok, “Building Express Backbone: Facebook’s New Long-haul Network,” Meta, May 1, 2017. 159 Midcontinent Independent System Operator, “Transmission and Substation Project Cost Estimation Guide For MTEP 2018,” Midcontinent Independent System Operator (MISO), 2018. Other site works Data center construction may require additional on-site works, like fire suppression systems or storm drainage systems, but the need for these is site-specific and the cost is usually small compared to other components.165 As such, this report does not consider other site works. Annualized capital expenditures It’s useful to be able to compare the one-time capital expenses discussed in this section to the ongoing operating expenses discussed in the previous section. To do so, upfront capital expenses are converted into annual expenses by modeling upfront capital costs as being evenly distributed over the lifetime of the capital. This is done by calculating the capital recovery factor (CRF), which converts an upfront cost into an annual cost depending on the lifetime of the capital and the discount rate. The CRF is higher when the lifetime of the capital is lower, because the upfront capital is spread out over fewer years. The CRF is also higher when the discount rate is higher, because this makes capital more expensive, leading to a higher annual cost. This report assumes that facilities have a longer lifespan than servers and networking infrastructure, and so have a lower CRF. See Appendix G for the precise values and derivation. Upfront and annualized capital expenses are presented in Table 4. Table 4 | Upfront and annualized capital expenditures. Servers and network infrastructure depreciate more quickly than the data center facility (meaning their upfront cost is spread over fewer years), so they account for a larger share of annualized capital costs than upfront capital costs. 165 4Specs, “Division 33: Site Utilities,” 4Specs, accessed Dec. 18, 2025. Time There are several processes that happen before a data center is operational, including permitting, construction, and energy procurement. It's costlier for hyperscalers if these processes take longer, for at least two reasons:

Financing costs during delay: Between the time data centers finance their investments and the time they begin earning revenue on their product, they are losing money while their investors expect returns. Delays increase the period over which these financing costs accrue.

Slower time-to-market: Delaying when a data center is operational could delay the time until a frontier model is trained or deployed. This could be costly for several reasons, including lost market share from consumers who switched to a higher-performing competitor model during the delay, less time for use of the frontier model for internal R&D, and lost reputation or clout associated with being perceived as the frontier company if a competitor releases a comparable model first. The first cost is much easier to robustly quantify, and is the only cost considered in the model. Because the time cost estimates included in this model do not account for time-to-market costs, they may underestimate the true cost of delays. This report considers the time requirements of permitting, construction, interconnection, and other pre-operational requirements like site selection. Only the costs associated with permitting and construction times are estimated; the time costs associated with interconnection and other pre-operational requirements are not included in the analysis due to a lack of data or relevance. Permitting Time US: 12 months UAE: 6 months Industry reports estimate that the average permitting time for data centers in the US is around 6-18 months.166,167 Permitting processes and timelines are relatively opaque in the UAE. As such, the relative speed of permitting for large energy projects is used as a proxy for permitting times for UAE data centers. 167 John Wright, “Navigating the Data Center Design and Approval Process,” Bohler, Oct. 2023. 166 Bray Dohrwardt, “How Long Does It Take to Develop a Data Center? A Step-by-Step Timeline,” Avisen Legal, Jan. 27, 2025. Table 5 | Permitting times for large energy projects in the US and the UAE. While data availability is limited, permitting for large energy projects appears to be significantly faster in the UAE. Table 5 compares the permitting times for large energy projects. The estimates for energy project permitting times – especially in the UAE – are based on limited data and should be treated as speculative (estimates of permitting times for solar energy in the UAE, for example, are based on only two data points). Nonetheless, the UAE permitting process appears to be faster than the US’s. Averaging across the different energy types, the UAE permitting process for large energy projects appears to take about half as long as the US permitting process. Assuming the relative permitting speeds for data centers are comparable to the relative permitting speeds of large energy projects, this report estimates the permitting for data centers in the UAE also takes about half as long as data center permitting in the US – around 6 months. These estimates may be inaccurate if data centers are given special priority in permitting. For example, US states have recently started offering incentive packages to data centers that include expedited permitting.168 Additionally, the US federal government issued an executive order in July 2025 to expedite permitting for data centers over 100 MW.169 These recent policy changes may not show up in the historical permitting time estimates for the US. In the UAE, data center permitting processes probably benefit from the fact that most data center projects are state-backed, and the state likely has a large amount of latitude to prioritize certain projects.170 Given the extent to which data centers are viewed as a national priority in the UAE,171 permitting timelines may be expedited more in the UAE than in the US, though scarce data limits the extent to which this can be quantified. 171 Gregory C. Allen et al., “The United Arab Emirates’ AI Ambitions: Key Implications for Maintaining U.S. AI Leadership,” Center for Strategic & International Studies, Jan. 4, 2025. 170 Khazna, the largest data center developer in the UAE, is a subsidiary of the state-backed firm G42 (Masri, 2025). 169 The White House, “Fact Sheet: President Donald J. Trump Accelerates Federal Permitting of Data Center Infrastructure,” The White House, July 23, 2025. 168 Melissa Reali, “Incentivizing the Digital Future: Inside America’s Race to Attract Data Centers,” Data Center Frontier, Aug. 5, 2025. Additionally, the model implicitly assumes that permitting and construction do not overlap (i.e., that construction cannot begin until permitting is completed). However, it’s possible that some parts of the permitting process may happen after the construction process begins. For example, some permitting processes in the US may allow for modular approvals,172 where construction is allowed to take place on approved parts of the projects while the builder waits for permitting approval on other parts. The degree to which overlap occurs is not well documented. Construction Time US: 24 months UAE: 29 months While it is possible to look at historical data of how quickly 100-MW data centers are built in the US, that is difficult to do for the UAE, where no data centers larger than 50 MW have been fully built (for which there is publicly available information).173 In the US, historical data derived from Epoch AI’s Frontier Data Centers174 database as well as news reports and satellite imagery suggests that data centers over 50 MW take around 24 months to construct.175 The full list of data centers, capacity, and construction time is available in the supplementary spreadsheet. In the UAE, in lieu of historical case studies for completed 100-MW data centers, a combination of three alternate sources are used: estimates of construction times for not-yet-completed 100-MW+ data centers; historical cases of construction time for completed data centers under 50 MW; and the relative construction speeds of large power plants. These are discussed in turn below. First, we can consider evidence from not-yet-completed 100-MW+ data centers. The UAE is in the process of constructing three such data centers, which overall do not appear to be progressing more quickly than US counterparts (See Appendix D for further discussion of the timelines of these data centers):

Moro Data Hub: The facility, built by a subsidiary of the Dubai Electricity and Water Authority, was first announced in 2021. On completion, the data center, which is being built in 10 phases, will be 100 MW and fully solar-powered. Data on square footage suggests that the facility will take about 108 months (9 years) to complete. While unusually long, this timeline should probably not be compared apples-to-apples with US timelines, for two reasons. First, the facility appears to be being built serially in phases, rather than all at once, and it’s possible that it could be constructed more quickly in parallel. Second, it’s possible that some of the delay is due to the facility’s complete reliance on solar power, which is 175 Corroborating this, US industry analyses report average construction times of 12-36 months (Dohrwardt, 2025). 174 Epoch AI, “Frontier Data Centers,” Epoch AI, last updated Dec. 12, 2025. 173 The largest fully operational data center in the UAE with publicly available capacity is Khazna’s Ibn Battuta complex, at about 34 MW (Khazna, 2025). For more discussion of the UAE’s largest data centers, see Appendix C. 172 Corin Dolan, “Building a Data Center: 10 Powerful Steps for Success 2025,” Accutech Communications, May 16, 2025. unusual in the industry; as such, the construction speed may be bottlenecked by energy production.

Khazna’s QAJ1: Khazna, a large regional data center company, announced the construction of their 100-MW AI-optimized data center, located in Ajman, in 2024. Current plans indicate that an “initial phase” of construction will be done in late 2026. At this rate, the initial phase would have been completed in 31 months. It’s not clear how large the initial phase is, or how on-track Khazna is to complete that phase.

Stargate UAE: The Stargate UAE data center, a planned 1 GW data center being built by Khazna in Abu Dhabi, was announced in May 2025 with plans for the first 200 MW to be completed in 2026. Based on analysis from Epoch AI, it appears that the construction process began in June 2025. Epoch estimates that the project is on track to be completed in December 2026, for a total of 18 months, though the project is young and delays are still possible. (While fast, this would still not exceed US timeline records, even if completed on time: xAI’s Colossus 1 in Memphis had 174 MW operational in a little over 4 months.176) Second, we can consider evidence from construction speeds for data centers under 50 MW. Here, there appears to be no substantial difference in construction speeds. Based on a limited sample of four under-50 MW data centers in the UAE, the average construction time is around 20 months; the average construction time for under-50 MW data centers in the US also appears to be about 20 months. Further detail on this analysis is available in the supplementary spreadsheet. Lastly, we can again consider evidence from comparative construction speeds for three types of large energy projects: nuclear plants, combined-cycle natural gas plants, and photovoltaic solar facilities. Based on limited data (especially for the UAE), construction speeds in the UAE appear to be faster for nuclear, slower for combined-cycle natural gas, and slower for solar, for a crude average of about 20% slower in the UAE. Multiplying the average construction time for 100-MW data centers in the US (24 months) by a factor of 1.2 produces an estimated construction time of around 29 months in the UAE. Further detail on this analysis is in the supplementary spreadsheet. To estimate construction timelines in the UAE, this report combines these three sources of evidence. Weights are assigned to the three 50 MW+ data centers, the average construction time for UAE data centers under 50 MW, and the relative speed of large energy projects in the UAE compared to the US, producing an estimated construction time of around 29 months, about 25% slower than in the US. See Table 6. The assignment of these weights is subjective; readers can adjust weights according to their judgement using the supplementary spreadsheet. 176 Epoch AI, “Data Centers/Satellite Explorer,” Epoch AI, accessed Dec. 18, 2025. Table 6 | Proxies used to estimate construction times of large data centers in the UAE. This report estimates that construction of large data centers takes longer in the UAE than the US. Commentary on Construction Time Estimates Estimates of UAE timelines are highly speculative, given the lack of direct data on data center construction in the UAE, the limited case studies for comparable large energy projects, and uncertainty about the comparability of energy projects to data center buildout. As such, these estimates of timelines should be approached with a high degree of uncertainty. However, there’s reason to expect the UAE to have relatively slower construction speeds for large data centers. The UAE data center market is relatively small: the UAE has not built any fully operational data centers greater than 35 MW (for which there is publicly available information177), and overall has a total data center capacity that is about 100 times smaller than the US. The UAE is also much newer to building data centers, with its first hyperscale data centers constructed in 2019. For further information on the UAE’s data center market size and history, see Appendix C. 177 It’s possible there are larger data centers without publicly available capacity information. Further, the UAE may struggle to access sufficient high-skilled talent needed for data center construction: while the UAE typically relies on imported foreign talent, especially in new industries, this approach may not work as well in the data center industry. The UAE is heavily reliant on migrant labor, including for skilled labor, with migrants178 making up 74% of the UAE’s population and an estimated 79% of the tertiary degree population.179 The UAE has historically relied on foreign talent to jumpstart new industries domestically. For example, the UAE’s first nuclear plant was jointly developed180 by Emirati and Korean companies, and involved 200 UAE engineers gaining training experience at Korean nuclear plants and 2,300 Koreans working in the UAE on the construction of the facility.181 There are many similar stories, including in the railway industry,182 the aerospace industry,183,184 and even the museum industry.185 However, the global data center industry is struggling to find qualified staff.186 Given that shortages are global, it may be more difficult for the UAE to import sufficient data center construction talent. Additionally, skilled labor shortages are likely a greater issue for larger data centers than smaller ones: large AI data centers involve more complex networking and infrastructure than traditional data centers, requiring more skilled workers like specialized electricians.187 Evidence from the Middle East data center market suggests these global shortages are already affecting the region. The Middle East data center industry already struggles with shortages of skilled labor,188 especially experienced workers.189 In the UAE in particular, the domestic construction industry is reportedly struggling with skilled labor shortages.190 Given the industry-wide skilled talent shortage in AI data centers, it’s possible that the UAE has struggled to import enough talent for the construction of its largest data centers. Interconnection Time The process of connecting to the grid in order to draw power from it is known as interconnection. For large electricity consumers such as hyperscale data centers, this process involves submitting an application detailing the expected size and characteristics of the electricity demand, which is then reviewed by the grid authority. 190 Adam Ralph and Graeme Baxter, “Construction Market Outlook: Robust Growth, Expanding Pipeline and an Innovative Future,” Turner & Townsend, 2025. 189 Douglas Donnellan et al., “Uptime Institute Global Data Center Survey 2024,” Uptime Institute, July 2024. 188 Ajay Mangara, “Middle East Data Centre Market Analysis: Demand and Opportunities for a New Digital Age,” Turner & Townsend, Feb. 2024. 187 Douglas Donnellan et al., “Uptime Institute Global Data Center Survey 2024,” Uptime Institute, July 2024. 186 Douglas Donnellan et al., “Uptime Institute Global Data Center Survey 2024,” Uptime Institute, July 2024. 185 Abu Dhabi Department of Culture and Tourism, “Developing a 360° Professional Capacity Building Plan,” France Museums, accessed Dec. 18, 2025. 184 H. E. S. Amiri et al., “The Emirates Mars Mission,” Space Science Reviews 218, no. 4, (2022). 183 eoPortal, “DubaiSat-1,” eoPortal, last updated May 25, 2012. 182 ER Communications, “Etihad Rail Moves to a Self-Operating Model After Concluding Knowledge Transfer Programme for the UAE National Railway Network with Deutsche Bahn,” Etihad Rail, accessed Dec. 18, 2025. 181 World Nuclear Association, “Country Profiles: Nuclear Power in South Korea,” World Nuclear Association, last updated Dec. 8, 2025. 180 Power Technology, “Barakah Nuclear Power Plant, UAE,” Power Technology, Sept. 25, 2024. 179 About 53% of immigrants had a tertiary degree in the UAE in 2020, compared to 41% of natives. (0.53*0.74)/(0.53*0.74 + 0.41*0.26)=0.79. Data comes from figure 6.6 (World Bank, 2023). 178 World Bank Group, “International Migrant Stock (% of Population) - United Arab Emirates,” World Bank Group, accessed Dec. 1, 2025. This report does not model the costs of interconnection delays, although they may be highly significant in some cases. This is primarily due to a lack of data. Data on large-load interconnection times – both in the US and the UAE – is scarce, making accurate estimates difficult. Interconnection processes are not standardized and don’t always involve a formal queue,191 and even in the US, there’s little data on the number, size, or wait time of projects attempting to draw energy from the grid.192 This makes it difficult to estimate likely interconnection timelines for a project of a given size. Additionally, even if there was reliable data on the typical lengths of interconnection review, it’s not clear how that translates into a delay for the hyperscaler. In particular, the average length of an interconnection review may be longer than the average length of the delay created by the review process. Hyperscalers can speculatively submit multiple interconnection applications before they have decided to begin a specific project, effectively allowing them to “start the clock” on the review process before they would otherwise be embarking on a project.193 Submitting multiple applications may also allow them to avoid worst-case delays, since they can choose whichever grid authority approves them for interconnection first. In principle, hyperscalers may also choose to embark on projects in limited ways (e.g. pursuing permitting or early-stage work) before interconnection is approved. The extent to which any of these specific dynamics apply in practice for hyperscalers is unclear given the lack of available data on the interconnection process. Furthermore, hyperscalers likely do not incur a large financial cost during the interconnection process, making it out of scope of the modeling in the report. Because applying for interconnection is not costly, hyperscalers likely would not need to secure financing for a specific project before submitting an interconnection request.194 As such, the time cost considered in this model – the financing cost during delay – likely doesn't apply to interconnection processes. This does not mean that interconnection delays are costless to hyperscalers, however: non-financial costs associated with time delays, like slower time-to-market for frontier models, would still be affected by interconnection processes that genuinely delay projects. 194 As mentioned, some utilities impose fees for the cost of reviewing interconnection requests. These range from $0 to over $1,000,000 for sites over 200 MW. These are insignificant on the scale of the multi-billion dollar construction costs associated with a 100-MW data center, and would be unlikely to require additional financing. (In the model estimated in this report, ten $1,000,000 requests would cost 0.2% of the total upfront capital costs associated with a 100-MW data center in the US). 193 Interconnection applications are usually not site-specific – a hyperscaler can typically apply for interconnection to a grid without identifying a specific plot of land. (This is distinct from permitting or construction processes, where a hyperscaler must know and have attained rights to a specific site). Because they do not need to have a site in mind to submit an interconnection application, this makes it easier to submit speculative applications in advance that they can later take advantage of when their project plans have become more specific. Interconnection applications also usually involve minimal financial commitment (Elevate Energy Consulting, 2025), making it cheap for hyperscalers to submit applications to multiple utilities. The ease of submitting speculative applications has in fact created a speculation problem in large-load interconnection queues (Elevate Energy Consulting, 2025): because the process is long and opaque and it’s not costly to submit interconnection applications, customers attempting to connect to the grid are incentivized to submit more applications that they could feasibly act on in the hopes that one will work out. This increases the reviewing burden for utilities, which further degrades the review process, and prompts further speculative requests. As evidence of the scale of the problem, data centers now make up 80% of the large-load interconnection queue in Western North America (Elevate Energy Consulting, 2025b). While certain utilities are beginning to implement interconnection application requirements like a security deposit or a demonstration that the applicant has exclusive rights to a specific plot of land (known as site control), interconnection application continues to mostly be a low-cost application process. 192 There is detailed data on generation interconnection times collected by the Lawrence Berkeley National Lab (Lawrence Berkeley National Laboratory, 2025), but this does not include data on large-load interconnection queues. 191 Elevate Energy Consulting, “Practical Guidance and Considerations for Large Load Interconnections (draft),” Elevate Energy Consulting, March 2025. The interconnection process is potentially still a significant barrier to the construction of large data centers. Even if hyperscaler projects aren’t delayed by the full duration of the interconnection review process, interconnection reviews could still be sufficiently lengthy that projects still face significant delays, especially if many grid authorities are experiencing similarly long delays. In Northern Virginia, for example, the local utility has warned that data centers over 100 MW attempting to connect to the grid may have to wait up to seven years195 for approval. (However, this likely represents the upper end of current US interconnection delays, as Northern Virginia is one of the most congested196 areas of the US electric grid; additionally, interconnection times are probably most serious for the largest data centers attempting to connect to the grid.197) As mentioned above, delays can be costly even if hyperscalers don’t incur any direct financial costs. While not included in the model, this report nonetheless attempts to consider whether hyperscalers building in the UAE may be less likely to encounter potential sources of interconnection delays than in the US. This analysis is discussed in Appendix F. On the one hand, the US’s larger grid may enable it to accommodate more data center capacity. On the other hand, the US likely has a more inefficient or backlogged interconnection review process. On net, it's unclear how interconnection times compare between the US and the UAE. The long-run picture may also differ from the immediate picture, since energy capacity growth may ultimately matter more than the efficiency of the interconnection review process. However, it is possible that interconnection delays are sufficiently shorter in the UAE as to make the UAE on net more attractive. Additional Sources of Delay There are likely other sources of pre-operational delays that are not accounted for in this model. These include the following:

Site selection: Before a data center is constructed, the builder must first pick a site on which to build. The process of choosing a piece of land involves evaluating the environment and terrain (e.g. determining whether the data center might experience flooding) and considering proximity to desirable features like population centers, fiber connectivity, and power. Power availability is often reported as one of the most significant factors considered during site selection, and so the site selection process likely overlaps significantly with the interconnection process.198 The process of evaluating environmental feasibility of a given plot of land may vary somewhat between the US and the UAE, but is unlikely to take a significant amount of time.199 199 Michael O’Shaughnessy et al., “Data Center Development Playbook: Streamlining Site Selection, Due Diligence, Entitlements, and Construction,” Bohler Engineering, May 2023. 198 Bloom Energy, “Onsite Generation Expected to Fully Power 27% of Data Center Facilities by 2030,” Bloom Energy, June 17, 2025. 197 Because interconnection delays are probably longer for larger projects, the hypothetical 100-MW data center considered in this report may face shorter interconnection delays than the largest hyperscaler-scale projects. However, because hyperscalers are actively considering facilities well above 100 MW, interconnection timelines are still likely to be an important factor in decisions about whether to build in the US or the UAE. 196 Dylan Meyer, “U.S. Grids Target Higher-Voltage Transmission to Alleviate Congestion,” FactSet, March 28, 2025. 195 Josh Saul, “Data Centers Face Seven-Year Wait for Dominion Power Hookups,” Bloomberg, Aug. 29, 2024.

Site planning: Once a site has been selected, engineers plan the specific layout and construction process of the data center build. As with site selection, environmental factors may affect the difficulty and length of this process,200 but this is unlikely to constitute a significant difference in total time between the US and the UAE. Total Time Cost The estimates of total time until a data center is operational can be used to estimate the direct financial cost of the pre-operational delay. As noted, these only account for the financing cost during delays, and do not account for other costs, including lost competition from delayed time-to-market. Table 7 summarizes the results of the construction and permitting estimates. Because construction is a larger share of the overall timeline, the UAE’s relative slowness on the construction process is more meaningful than their relative speed on the permitting process. Table 7 | Estimated permitting and construction times in the US and the UAE. This report estimates that permitting is significantly faster in the UAE, while construction appears to be faster in the US. The cost of the delay is estimated using a) the total length of the delay; b) the value of capital expenditures; and c) an industry-specific discount rate. The details of these calculations are discussed in Appendix H. The unannualized cost of delay across permitting and construction is estimated at $537 million in the US and $518 million in the UAE. Combined Annual Cost To calculate a total annual cost, annual operating expenses and annualized capital expenses and time costs are added together. Table 8 shows each annual expense sorted by greatest discrepancy. 200 Schneider Electric, “A Practical Guide To Data Center Planning and Design,” Schneider Electric, 2016. This report estimates that the US’s greatest cost advantages are its lower energy costs (due to both cheaper electricity and greater energy efficiency due to being in a more temperate environment), potentially faster construction times (due to a more advanced data center industry), and lower maintenance costs (due to the generally more temperate environment in the US). The US’s greatest cost disadvantages are facility construction costs (possibly due to construction labor costs) and permitting time. Costs net out such that the US appears to be very slightly cheaper, though this should be interpreted as being within the margin of error. Table 8 | US cost advantage compared to the UAE, by data center cost component. The US’s greatest cost advantages are cheaper power, faster construction time, and lower maintenance costs. The UAE’s greatest advantages are faster permitting time and lower facility construction costs.

Limited grid capacity and slow capacity growth: Customers cannot draw power from the grid if there is not surplus power. On this source of delay, the US probably benefits significantly from having a larger grid. As of 2024, the US grid had a capacity of about 1,294 GW, while the capacity of the UAE’s grid was about 53 GW.264 As a function of this, the US’s grid is growing faster than the UAE’s: over the last five years, the US added an average of 27 GW a year while the UAE added an annual average of 2.7 GW. This does not necessarily indicate that there’s more available capacity in the US grid, since the larger population and industrial capacity of the US means that there is also greater demand for electricity. However, it may make the US able to accommodate greater additional demand, for a few reasons: ○ The US likely has greater grid-associated resources, including skilled labor,265 infrastructure manufacturers,266 and specialized construction equipment, which may make the US better-equipped to build capacity quickly. ○ Because the US’s grid is larger, grid improvements that are proportional to grid capacity unlock greater capacity in the US. For example, curtailment policies,267 where large electricity consumers agree to limit their electricity usage for a small share of the year (e.g. 0.5%) and which could be applied across the grid, could theoretically unlock around 98 GW of extra capacity in the US – including several regional systems that could each accommodate at least 10 GW of extra capacity.268 Assuming comparable grid utilization in the UAE, these practices could only unlock an additional 4 GW in the UAE.269 ○ The US likely has more unused interconnection locations. For example, the DOE has considered whether US data centers could use the interconnection points from retiring coal plants.270 Additionally, grid capacity growth may be bottlenecked by global supply chain delays.271 While significant, these presumably affect the US and the UAE equally. 271 Jesse Cohen et al., “Gas Turbine Supply Constraints Threaten Grid Reliability; More Affordable Near-Term Solutions Can Help,” RMI, June 18, 2025. 270 Pacific Northwest National Laboratory, “Redeveloping Coal Power Plants: Data Centers,” Pacific Northwest National Laboratory, Sept., 2024. 269 Assuming the load factor (i.e., the ratio of average demand to peak demand) (Norris et al., 2025) is similar in the US and the UAE, the UAE could unlock an additional 53/1294*98 ≈ 4 GW. 268 According to estimates by Norris et al. 2025, PJM, MISO, and ERCOT could respectively accommodate an additional 17.8, 14.8, and 10.0 GW at 0.5% curtailment. 267 Tyler H. Norris et al., “Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems,” Nicholas Institute for Energy, Environment & Sustainability, 2025. 266 For example, a majority of US distribution transformer consumption is sourced domestically (Industry and Security Bureau, 2021). 265 Goldman Sachs, “The Power Industry May Need More Than 750,000 New Workers by 2030,” Goldman Sachs, July 23, 2025. 264 Ember, “Electricity Data Explorer,” Ember, accessed Dec. 18, 2025

Slow interconnection review processes: Interconnection requests probably take longer to be reviewed in the US. The US requires large loads attempting to connect to the grid to submit interconnection applications, which enter an informal queue of projects waiting to be reviewed by the regional utility or transmission authority. These review processes can take years, and the growing number and complexity of interconnection applications has reportedly increased the reviewing burden. By contrast, there is little publicly available evidence of comparable interconnection queues or long review processes in the UAE, though minimal data on the Emirate interconnection process makes this difficult to assess confidently.272 Emirati governments, which are strongly centralized,273 may also have more latitude to prioritize projects of national importance, like data centers. It's unclear on net how interconnection delays compare between the US and the UAE. The lack of publicly available data on large-load interconnection processes, particularly in the UAE, makes this an area where future research could improve the model's accuracy. 273 Bertelsmann Stiftung, “United Arab Emirates Country Report 2024,” Bertelsmann Stiftung, 2024. 272 Individual cases suggest quick interconnection times, though it’s not known how early interconnection planning began. For example, if Stargate UAE is brought online on the planned schedule (i.e., by the end of 2026), it will have gone from planning to connection in about 18 months. Since the US had only announced that it would allow the export of advanced Nvidia chips to countries like the UAE a month prior (Nellis, 2025), it’s unlikely that there was significant pre-construction interconnection planning. (It’s not yet confirmed that the UAE will have sufficient power availability by the completion of the Stargate data center, so there is still the possibility of interconnection delays).

The discount rate for these calculations reflects the rate of return that investors financing the data center investment expect. This is calculated by weighting a company’s debt and equity to produce a weighted average discount rate, called the weighted average cost of capital (WACC). The WACC varies by industry; using the computer services industry as a proxy, the WACC for hyperscalers is probably around 9%274 (Equinix, for example, has a WACC of 9.5%275).

The lifespan of capital investment depends on the item: different products last for different amounts of time. In particular, servers and networking equipment depreciate more quickly than the surrounding facility. Based on an average of lifespan estimates from hyperscalers’ own accounting and third-party analysis, this report assumes that the average lifespan of the facility and servers are around 14 and 5 years, respectively. See the supplementary spreadsheet for further detail. Applying these values for the discount rate and the lifespan produces the following CRF values:

Facilities (14-year lifespan): 0.13

Servers and networking (5-year lifespan): 0.24 The values can be used to annualize the upfront costs. 275 Finbox, “WACC for Equinix Inc,” Finbox, accessed Dec. 18, 2025 274 Aswath Damodaran, “Cost of Equity and Capital (US),” Stern, NYU, last updated Jan. 2025.

t = expected length of delay

i = discount rate

P = upfront capital cost The discount rate, i, is determined by the industry WACC. The effective duration of the delay, t, is shorter than the full calendar length of the delay because construction financing is typically advanced in stages rather than fully disbursed at the outset. This report assumes a linear drawdown of construction financing over the delay period, implying that on average, half the capital is outstanding for the full delay duration. As such: 𝑡 = 0. 5 · (𝑑𝑒𝑙𝑎𝑦 𝑑𝑢𝑟𝑎𝑡𝑖𝑜𝑛) To enable comparison with annual operating costs and annualized capital expenditures, delay costs are expressed as equivalent annual costs. Delay-related financing costs are first allocated proportionally across capital components and then annualized using a capex-weighted average capital recovery factor (CRF). This weighted CRF reflects the differing economic lifetimes of underlying assets (i.e., servers and facilities, which have different lifespans) and is used solely to express delay costs on an annual basis. Annualizing delay costs and reporting them separately allows them to be normalized and explicitly compared with other annual cost components. To the extent that alternative annualization assumptions would change estimated cost levels, these changes would primarily affect absolute magnitudes rather than relative comparisons across countries, since the same methodology is applied consistently in all cases (e.g. the relative cost of permitting delays between the US and the UAE would remain accurate even if the estimated magnitude of permitting costs for a given country is imprecise).