Renewable-Aware AI Infrastructure Planning: An Optimization Framework for Matching Data Center Expansion with Renewable Generation and Energy Storage

Full text

Text extracted from the article PDF for reading and search. Tables, figures and equations are in the PDF, which remains the authoritative version.

1. Introduction

Artificial Intelligence (AI) has reshaped the electricity usage patterns of the digital infrastructure industry. In total, the IEA estimates the electricity demand of data centers at approximately 415 TWh in 2024, representing 1.5 per cent of global electricity demand, and predicts future demand growth to more than 945 TWh by 2030 at an annual rate exceeding 15 per cent (well above the annual rate expected for global electricity demand growth) [1]. In 2025, demand for electricity for data centers increased by another 17 per cent compared with 2024, while electricity consumption needed per single AI task remained lower [2]. Nearly half of the global demand growth for data centers to 2030 is expected to come in from the United States alone [1]. Data center electricity uses more than doubled in the USA in five years, rising from approximately 4.4 per cent of national electricity demand in 2017, to between 12 and 16 per cent in 2028 [4].

Renewable “energy” supplies are at the same time the largest and most limited one of these growth facilitators. According to the IEA, renewables (mainly wind, solar photovoltaics (PV) and hydropower) provide around 27 per cent of global data center electricity today and will contribute close to 50 per cent of incremental requirement growth to 2030, increasing by 22 per cent on average annually (2030 estimate) [1]. Natural gas and coal are still expected to serve over 40 per cent of incremental data center demand for the near term, despite the challenges of matching intermittent renewables, slow grid interconnection, and spiky growth in AI load demand [1,2]. The interconnection process has emerged as a first order bottleneck: In 2024-2025, the U.S. queuing system for active interconnections in the United States had approximately 2.3–2.6 terawatts (TW) of generation and storage projects in the pipeline, more than the entire national fleet; projects in these queues often take around five years to be built and tend to be focused on solar, with historical completion rates below 15 per cent [5,6]. That, coupled with growing interconnection cost (currently US$243/kW for solar, and US$265/kW for battery storage at completed projects), often hinders renewables from coming online when hyperscale data center operators are pushing for it [6]. Many such large-scale accounting exercises have revealed that estimates of energy usage in data centers require a significant, regular recalibration when efficiency improves, and when the growth trajectory of the data center changes [7], and new concepts of how to design data centers as active participants, more than consumers, in the electricity grid have emerged [8]. As response, industry and academia interest grew to join renewable-aware infrastructure planning: co-located or contractually coupled data center load growth with renewable generation and energy storage, to jointly optimize infrastructure siting, capacity expansion and operational dispatch selections, instead of making distinct, sequential and siloed decisions. In practice, matchable dispatchable solar and wind power with battery storage [9], corporate power-purchase agreements (PPAs) subsidizing new renewables purchase [36, 37] and work-load scheduling in the presence of renewable supply and demand price disparity [15, 17, 18] are examples of practical applications. These lines of research have, however, progressed largely independently, with distinct efforts on power-systems engineering around generation-storage sizing, on computer-systems research around workload scheduling, and on energy-siting research around land-use and multi-criteria siting evaluation, and far fewer frameworks that perform siting, sizing and scheduling under a specific uncertainty structure of AI driven demand growth.

This review is focused on that. It collates and integrates literature from a multitude of disciplines related to renewable-aware AI infrastructure planning, introduces a modular optimization framework that simultaneously optimizes data center siting, renewable power and storage power size and grid interconnection in solar wind uncertainty, and discusses the qualitative behavior of the optimization framework using a medium-sized set of scenarios as an example. The rest of the paper is structured as follows. Section 2 sets out the review methodology. In Section 3 an overview of the literature is compiled according to five thematic strands. Section 4 is a formality of the planning problem. Illustrative results obtained from the proposed framework are reported in Section 5. Section 6 discusses the findings in relation to practice and policy. Section 7 sets out limitations, and Section 8 concludes.

2. Methodology

The article not only provides a structured literature review but also specifies and illustrates an original optimization framework that is consistent with review articles in which a review is accompanied by a specific methodological contribution.

2.1. Literature Identification and Screening

The sources identified were based on combinations of the search terms “data center”, “data center”, “renewable energy”, “energy storage”, “battery sizing”, “carbon-aware”, “site selection”, “interconnection queue”, “power purchase agreement”, and “optimization”, “planning” across the Scopus, ScienceDirect and IEEE Xplore and arXiv databases, as well as grey-literature from IEAs, LBNL and major professional-services and industry bodies. The search window spanned January 2017 through September 2026, with preference for papers published within the last five years (2019–2025) that provided either new terms or procedures for the science introduced by materials and structures development work, as further studies compounded or advanced them [25, 34, 41, 42]. Relevant records were selected from five thematic strands namely, (i) electricity-demand characterization for AI/data center; (ii) renewable and storage co-sizing techniques; (iii) multi-criteria and geographic site selection for renewable generation; (iv) carbon- and cost-aware computing/workload scheduling; and (v) grid interconnection and renewable procurement economy. Records that were duplicated, in sources that were not English, or for which no venue or digital object identifier was traced were not included.

2.2. Framework Development Approach

The optimization framework proposed in Section 4 follows the previous line of work for data center energy-resource sizing (two-stage versus multi-stage stochastic and mixed-integer linear programming, MILP [9, 11, 12, 13, 14]), and for sizing microgrids and workload placement (carbon-aware workload placement [15, 19, 20]). It treats energy-resource locations, renewable energy capacities, and storage capacities as combined first-stage and second-stage decisions. The framework is also written at an abstraction level that matches those of the solution methods that were discussed in the literature (stochastic programming, two-stage decomposition, and hybrid optimization–learning approaches [14]).

2.3. Empirical Benchmarking

Unlike every aspect of the study in Section 5, modelled outcomes for the siting, sizing and dispatch of renewable-energy equipment, down to the level of the individual facility are not publicly released by any operator and this is not an end-to-end facility-level case-study of the full framework. However, this review does not accept the operating data provided by operators but uses the actual data wherever they have been provided. In its Section 5, Google released fleet-wide, quarterly and trailing twelve-month PUE data from 2008 onwards [45] which is used to compare the operational-overhead assumption that any renewable-and-storage sizing exercise should make. Epoch AI independently compiles a satellite and permit-based database of the world’s largest AI data centers, including their location, owner, IT power and compute capacity [46], which is used in Section 5 to characterize the real scale and geographic distribution the siting stage of the framework must plan around, and to cross-check that distribution against the interconnection-queue geography reported by LBNL [5,6]. No operator publishes a combined dataset to directly validate the joint renewable-fraction, cost and reliability outcome a model of full sizing, scheduling and procurement in Section 4 would yield at a particular facility, hence the latter is considered to be a directional expectation based on qualitative and quantitative findings reported in the papers that reviewed until now [9, 11, 21, 23, 27].

3. Literature Review

3.1. Ai-Driven Electricity Demand and The Changing

Load Profile The first strand of literature quantifies size and configuration of data center electrical consumption. Historically, data center load has increased at about 12 per cent per year, but, more recently, accelerated demand from accelerated (GPU-based) servers has been the leading cause of load growth with accelerating server electricity demand estimated to increase by about 30 per cent per year, vs. 9 per cent per year for conventional servers [1, 3]. Power, cooling and networking as co-designed resources is a framing that builds on the foundational treatment by Barroso and Hölzle of data centers as single computing centers, or warehouse-scale computers (WSC), of data and information [41]. Jones and others were worried about the early growth of data center electricity that structural changes are necessary to keep up with efficiency improvements [42]. In comparison, more recent analyses place greater emphasis on the temporal characteristics of AI workloads – which are large, fast spikes in power consumption during their training runs and inference bursts, increasing stresses on on-site power delivery systems and on power-grid interconnections in ways that traditional, comparatively flat data center workloads would not. [2] This evolving load profile is an important input to the demand-uncertainty treatment in the optimization framework identified in Section 4, being that siting and sizing decisions were taken based on a smoothed picture of demand would become severely under- or over-provisioned with materially different fluctuations in realized AI demand.

3.2. Renewable and Storage Co-Sizing Methods

A second more technical track focuses on what amount of renewable electricity and energy storage capacity (SA) a data center/microgrid should install and how it should be dispatched. Figini and Paolone develop a cost and carbon-efficient sizing problem for PV production and battery storage to enable PV production being more dispatchable than non-dispatchable loads, positioning the contribution in the context of previous analyses of hybrid grid storage solutions and stochastic sizing for integrated energy systems [9]. For AI-focused data center facilities, related work proposes risk-aware multi-stage planning models that explicitly propagate uncertainty related to interconnection-queue delays, and electricity price to capacity decisions at the facility level, sizing interconnection capacity, co-located renewable generation, and storage power and energy capacity [10]. Microgrid-composition studies have placed the co-location problem in the context of reducing the reliance on the grid and found that interactions of microgrid composition with data center load scheduling have not received adequate attention when compared with the amount of research done on data center load scheduling alone [11]. To optimize the day-ahead scheduling parameters, a two-stage optimization– learning approach has been proposed for multi-zonal data centers with renewable generation and batteries in particular and compared to the purely MILP baseline and purely RL baseline [14]. In the context of active distribution networks (ADNs) and for the capacity planning of distributed energy resources (DERs), when intermittency in renewables and storage appears as a stochastic phenomenon, approaches based on Bayesian optimization have also been proposed to account for the occurrence of scenario explosion [12]. It has been known since earlier operational-research treatments that onsite need for renewables and need for storage in data centers invariably have to be co-designed with workload scheduling, and not solved unilaterally; architectures that go by the name of “negotiation” style, which coordinate the placement of workloads with power-supply decisions, and models that explicitly schedule workloads for the periods of high availability of renewables onsite, were previously known and had existed prior to the AI-driven demand surge [13]. He and Shen propose a joint minimization of the operating cost of distributed green data centers with energy storage with the introduction of an explicit carbon cap, an early application of using storage sizing and carbon constraints in the same objective [24]. The work of Khosravi, Andrew and Buyya’s virtual-machine placement method to reduce energy and carbon cost in geographically distributed cloud data centers is a foundational precursor to the spatial load-shifting literature as discussed in Section 3.4 [25].

3.3. Energy-Storage Technology and Sizing

A parallel literature (PL) is dedicated to hardware technology selection and dimensioning, without relying on the data center application. Detailed reviews of hybrid energy-storage systems show, using cost, round-trip efficiency, technological maturity and environmental impact as the parameters, that battery energy-storage systems (BESS) are well suited for high-power, high-energy-density applications such as data center backup and dispatchability [30]. Techno-economic sizing methods can vary from simplified models of energy to power ratio, which is compared to commercial building tariff structures [27] to multi-objective co-design models and algorithms that optimize the battery chemistry, sizing and control of power electronics together optimally, such as by the algorithms NSGA-II [28]. Batteries at the transmission scale can be subject of a systematic review including transmission-scale grid applications such as frequency regulation, peak shaving and renewable firming [29]; the battery sizing for utility scale photovoltaic plant can be the subject of a review limited to the different kinds of technology-economic, reliability-oriented and hybrid mathematical-programming approaches [31]. The majority of these storage studies have shown the ranges of parameters (e.g. powertrain cost per kWh, round-trip efficiency, cycle life) that a data center-based sizing model like the one outlined in Section 4 should use but have not explicitly dealt with data center load profiles.

3.4. Carbon- and Cost-Aware Workload Scheduling

Another strand tightly tied with the computer-systems literature is that of renewable-awareness as a scheduling instead of a capacity-sizing problem. Focusing not just on technical optimization, Bashiret al. [16] are critical and caution in general to view sustainable computing as an efficiency problem because often structural, demand side, change is needed. Acunet al. [15] also propose a holistic framework for designing carbon-aware data centers that consider a combination of embodied and operational carbon, to approach 24/7 zero carbon operation. In both the previous work on delay-tolerant scheduling to clearly show emissions reductions by doing “waiting a while” before executing deferrable jobs [18] as well as an admission-control mechanism for cloud and edge workloads [17] formalizing temporal workload-shifting, we observed this phenomenon. In this vein, Hallet al. contribute carbon-aware versions of computing formulations for networks of data centers that provide probabilistic performance guarantees a realistic worry of pure carbon-following dispatch on degrading service-level reliability [19]. Lechowiczet al. frame the pause-and-resume decision as an online optimization problem which offers competitive-ratio guarantees for carbon-aware load shifting [20]; Lin and Chien investigate the green adaptation of data center computing capacity, not just only the timing of its load, [21]. The contribution by Sukprasertet al. is very important, as they empirically measure the extent of carbon-aware workload shifting and show that target emissions reduction achievable to real workload limitations is often less than what was assumed in idealized simulations [22]. Yan, Chow and Chen present a low-carbon data center operation scenario together with workload sharing and carbon-allowance trading for a distributed fleet over a network [23]; Souzaet al. propose a Carbon-Aware Scheduling approach to geographically distributed Web services with explicit provisioning decisions for carbon [44]. Caoet al. and Dodgeet al. offer broader surveys of, respectively, carbon-neutral data center design strategies [26] and measurement challenges when quantifying the carbon intensity of AI workloads specifically [43]. Combining all of this scheduling literature, it can be seen that there is realistic potential for operational, near-real-time renewable-following behavior to meaningfully reduce emissions and stress on the grid, and that this can only be achieved to the extent of the actual installed renewable and storage capacity, arguing for capacity planning and operational scheduling decision making to be treated as co-integrated optimization layers, not as separate ones.

3.5. Site Selection and Spatial Planning for Renewable

Generation The fourth strand, found in the literature of renewable energy siting and geographic information systems (GIS), concerns the location of renewable capacity, and, by extension, the location of renewable energy paired capacity. Shaoet al.’s systematic review of multi-criteria decision making (MCDM) applications for renewable-energy site selection brings together 85 studies from the year 2001 to 18 2018, documenting the key steps of criterion selection, weighting and validation used in GIS-based Analytic Hierarchy Process (AHP) siting techniques [32]. Applying land-suitability, resource-intensities, cost of proximity and environment-constraint layers in the GIS-AHP approach is essentially consistent across countries and regions, with a few case studies built in the case of applications presented here: solar PV siting in Saudi Arabia [34], Turkey [33] and Peru [35], and validated for a small subset of sites. This literature was not created with data centers as its focus, but its taxonomy of criteria (resource availability, land-use conflict, proximity of transmission lines, water availability, and topographic constraint) is very similar to the siting stage of the proposed framework in Section 4, as increasingly siting a data center will require consideration of water and land constraints in addition to renewables-resource quality.

3.6. Grid Interconnection, Procurement and Forecasting

The third strand is infrastructural and market mechanisms for connecting renewable capacity with load on the data center. At the end of 2024, U.S. interconnection queues contained about 2.3 TW of generation and storage capacity, with typical queue durations being about five years and interconnection processing efficiencies consistently under 15 per cent, even though no increases due to processing efficiency improvements were recorded and there was also a significant wave of project withdrawals [5,6]. The most common non-utility approach to underwriting a renewable energy generation project is through PPA structures: in this regard, Mohseni Taheri, Nadarajah and Trivella [36] show, formally, that PPA structures can be physically- or virtual-based under demand and price uncertainty, while Stanitsas and Kirytopoulos [37] discuss the dynamics of engaging stakeholders within hybrid renewable-energy PPA structures. Another prerequisite for successful co-sizing and dispatch is the accuracy of the weather forecasts for the renewable resource, which was found to be comparable between several machine-learning techniques for solar and wind power forecasting, with artificial neural networks and ensemble techniques showing a better performance than simple statistical methods in the short time range that is most important in operational forecasting [38,39,40]. These studies show how relevant interconnection-queue reform and procurement-contract design are and forecast to the viability of renewable-aware data center planning as much as the developed capacity-sizing mathematics described in Section 3.2.

3.7. Synthesis

Table 1 summarizes the optimization approaches described above regarding the thematic strand, decision variables treated and mainly used method. Two patterns emerge. Firstly, the sub-problems in the literature are relatively well advanced. The traditions of sizing and scheduling are extensive and, to a lesser degree, so are those of siting and procurement. Secondly, most of the literature still provides frameworks for siting and renewable and storage size optimization separately, meaning that the closest analogues to the framework presented in Section 4 are the risk-aware source–load–storage–grid framework [10] and the dispatchability-oriented sizing model [9].

Table 1: Summary of reviewed optimization approaches for renewable-aware data center infrastructure planning, by thematic strand. Thematic strand Representative method(s) Primary decision variables Key source(s) Demand characterization Statistical/scenario-based load modelling; national energy-usage accounting Forecast electricity demand, load volatility [1], [2], [4] Renewable–storage co-sizing Two-stage stochastic MILP; hybrid MILP– reinforcement learning; Bayesian optimization Renewable capacity, storage power/energy, dispatch schedule [9], [10], [11], [12], [14] Storage technology & sizing Techno-economic models; multi-objective co-design (NSGA-II)

Storage technology choice, energy-to-power ratio, cost [27], [28], [30], [31] Carbon/cost-aware scheduling Online/competitive-ratio algorithms; admission control; carbon-allowance trading Workload timing and location, service-level constraints [15], [17], [19], [20], [22], [23] Site selection GIS-based multi-criteria decision analysis (AHP)

Candidate site ranking, land/resource suitability [32], [33], [34], [35] Interconnection & procurement Queue-attrition statistics; Markov decision process contract analysis Interconnection capacity, PPA structure and volume [5], [6], [36], [37]

4. Problem Statement

This planning problem is formally stated in an informal way as follows: Given a set of potential development sites, multi-period forecast of the electric load served by the facility, forecast of renewable energy supply capacity and cost at the developed site(s), and forecast of interconnection load capacity and cost, determine (i) which site(s) to develop, and in what sequence, (ii) how much renewable energy supply capacity to install or contract at the developed site(s), (iii) how much energy storage power and capacity to install and (iv) how much load to serve from the interconnection, so as to minimize the expected levelized cost to serve the facility’s load over the planning horizon, while maximizing the share of demand met by renewable energy sources. Formally, $t \in T$ denotes time periods within a representative operating year, $\omega \in \Omega$ uncertainty scenarios (covering AI workload growth scenarios and renewable resource realizations, as used in [10]), and $s \in S$ denotes candidate locations for the new site assessed by the multi-criteria siting layers introduced in the GIS-AHP literature [32–35]. First-stage (here-and-now) decisions include the site-selection binary variables, installed renewable capacity by technology, and the installed storage power and energy capacity; second-stage (wait-and-see) decisions include the period-by-scenario dispatch of renewable generation, storage charge/discharge, grid imports, and the reallocation of delay-tolerant computational load across time (following the literature on workload-shifting [17, 18, 20]). The goal is to minimize the total annualized capital cost of renewable and storage capacity, annualized interconnection (as documented cost escalation for queued projects [6]), and expected operating cost over all scenarios while including the cost of curtailing generation and the cost of reliability shortfall, with the latter defined as the cost of being less reliable than the probabilistic performance guarantees proposed by Hallet al. [19]. Constraints enforced are energy balance (site/site/period); energy storage state-of-charge dynamics bounded by capacity of installed energy storage (site/all/period); power limit (production/input/storage capacity) bounded by installed energy storage power capacity (site/period/input/storage capacity); ceiling for energy import from the grid, reflecting interconnection capacity, which may itself be a first-stage decisions (site/period/import/first degree decisions) [5,6] ; and land, water, and policy suitability constraints drawn from the energy storage site selection literature [32–35]. This formulation contains several existing models as special cases. If storage capacity is set to zero, then the problem of site selection becomes an exogenous problem, recovering a pure carbon-aware scheduling problem, similar to those studied in [15,17,18,20]. Fixing the site and demand scenario set to a single realization recovers a deterministic sizing problem of the kind solved by simplified techno-economic models [27]. Leaving the stochastic, multi-site, multi-technology structure intact results in a problem that is structurally similar but more general than the risk-aware co-optimization model of [10], and the dispatchability-oriented sizing model of [9] that the present approach generalizes in this respect by permitting an explicit nesting of the present optimization problem, which includes both the site-selection decision and the workload-shifting decision, into a stochastic program.

The challenge of modelling the uncertainty of AI workloads, renewable resource and multi-site interconnection costs simultaneously is tractability, which is explicitly identified in the distributed-energy-resource capacity planning literature as an issue that leads to the use of noise-aware Bayesian optimization alternatives to a combinatorial growth in the scenario tree [12]. The framework proposed here builds upon the two-stages decomposition/scenario reduction approach that has been followed by other recent works in that area, and the hybrid optimization–learning structure used by [14] (where the day-ahead optimization is a deterministic MILP core with a learned real-time adjustment layer) is recommended as the operational-scale structure of solution adopted here, where a full stochastic MILP is still retained to compute the coarser-grained siting and capacity-sizing decisions with long term-horizons.

5. Results

The three exhibits in this section support the essential assumptions of the framework with evidence. Figure 1 places the exercise in the context of a demand trajectory established in section 3.1. Figure 2 provides a capacity-sizing check against Google’s own published fleet-wide operational efficiency data to compare the assumption for operational efficiency applied in any capacity-sizing exercise. Table 3 uses Epoch AI’s independently compiled database of the world’s largest AI data centers to show the real scale and geographic distribution the siting stage of the framework must contend with. In combination, they offer an empirical foundation for the stochastic sizing exercise begun in Section 4, which is conducted without the use of assumed parameters.

Fig 1: Global data center electricity consumption, 2017–2035, and the renewable share of data center electricity supply. Consumption figures for 2024 and 2030 and the renewable-share estimate for 2024 follow IEA Base Case projections [1]; 2035 consumption and post-2024 renewable-share trajectories are illustrative extrapolations consistent with the reported growth rates in [1,2].

Figure 1 places the scenario comparison in the context of the overall demand trajectory defined in Section 3.1 for global demand of electricity for data centers more than doubling from 2004 to 2030, with the renewables component of supply projected to increase from about a quarter in 2004 to nearly two-fifths in 2030 in the IEA’s Base Case [1]. The interaction between these three elements of siting, sizing and scheduling is most pronounced with such a high rate of demand growth and increasing, but still partial, interconnection of renewable energy sources, and the cost of being undersized or over scheduled is greatest, whether for renewables and storage or for relying too heavily on congested grid interconnection.

Fig 2: Google data center fleet trailing-twelve-month (TTM) power usage effectiveness (PUE), Q1 2009–Q1 2026, benchmarked against the 2025 industry-Figure 2 shows actual revealed operating data, and not the result of a simulation. Google’s fleet-wide TTM PUE was reported at 1.19 in Q1 2009, but dropped to 1.09 by 2024– 2026, 83 per cent lower than the overhead-energy implied by the Uptime Institute’s Global Data Center Survey 2025 [45] industry-average PUE for 2025 of 1.54. This is relevant to the sizing exercise detailed in Section 4 as total facility power, which renewable capacity and storage needs to be sized for, is total IT power multiplied by PUE, rather than total IT power. As a result, and because PUE is a scenario-dependent factor, it does not imply that an operator who is already operating near the efficiency frontier in Figure 2 will be under-provisioning from the efficiency benchmark of average 1.5 (see also Section 4), while a lower value indicates that he will be over-provisioning.

The second part of the framework in Section 4, the sizing of the storage, comes from the literature on storage but not from a single named case-study because there is no publicly available resource that provides detailed data on the storage used in each data center deploying AI. Table 2 shows our synthesis of that literature for a comparison of storage technologies [29,30,31]. Lithium-ion battery systems are lingering as the major short-term technology option for applications involving dispatchability in data centers, as delivered by their high round-trip efficiencies, response-time and commercial maturity, although flow batteries, flow batteries combined with drying and heat retention, and hydrogen-based storage are outlined in the reviewed literature as other options for multi-day or longer duration firming, rather than diurnal balancing [30].

Table 2: Comparison of energy storage technologies relevant to data center renewable-aware planning, synthesized from [29], [30] and [31]. Technology Typical duration Round-trip efficiency Response time Commercial maturity Primary data center role Lithium-ion battery Minutes–hours 85–95% Milliseconds High Diurnal balancing, backup, frequency support Flow battery (e.g., vanadium redox) Hours 65–80% Seconds Medium Extended-duration renewable firming Compressed air / pumped storage Hours–days 40–70% Minutes Medium (site-dependent) Bulk, long-duration balancing Thermal storage Hours–days 50–90% (application-dependent) Minutes Medium Cooling-load shifting, waste-heat reuse Hydrogen / fuel cell Hours–seasonal 30–45% Seconds– minutes Low–medium Seasonal firming, long-duration backup Table 3 maps the true size and location of the world’s biggest AI data centers from an independently compiled database of satellite detection and permits from Epoch AI [46] to demonstrate the scale and whereabouts of data centers the siting stage in the framework needs to plan for. By September 2026, Epoch AI has identified 93 of them, accounting for a combined 13.5 gigawatts (GW) of IT power capacity and a total facility power capacity of approximately 20–40 per cent, as documented in preceding sites [46]. Individual facility size ranges from less than 50 megawatts (MW) up to more than 900 MW at the largest individual sites and is summarized in Table 3, which lists the eight largest by compute capacity.

Table 3: The eight largest AI data centers tracked worldwide by AI compute capacity, as of September 2026. IT power is reported only where Epoch AI states a specific figure; “-” indicates a site for which Epoch AI does not report a discrete power value. Source: Epoch AI, AI data centers [46]. Facility Owner Location IT power (MW) Compute (kH100-eq) Colossus 2 SpaceX AI Memphis, TN, USA 946 1,112 Microsoft Fairwater Atlanta Microsoft Fayetteville, GA, USA 636 769 Anthropic-Amazon New Carlisle Amazon New Carlisle, IN, USA 910 686 Google Pryor (North) Google Pryor, OK, USA - 637 Meta Prometheus Meta New Albany, OH, USA 496 600 OpenAI Stargate Abilene Oracle Abilene, TX, USA 421 509 Microsoft Fairwater Wisconsin Microsoft Mount Pleasant, WI, USA - 446 DayOne Kempas Unknown Kempas, Johor, Malaysia - 368 Two characteristics of Table 3 directly relate to the problem set forth in Section 4. Next, US facilities are over-represented in this largest-scale tier, which is among the smaller number of states in Epoch AI’s broader siting database (14 tracked facilities in Texas, 7 in Ohio, 7 in Virginia, 4 in Nebraska, and 4 in Iowa) [46] that also aligns heavily with the states with the longest active interconnection lists at LBNL’s data source [5,6], providing direct empirical backing for the siting– interconnection tension possible subject of Section 6.2. In addition to that, the reported utilization of the IT components (60-80 per cent of nameplate capacity on average, with supporting infrastructure overhead of 20-50 per cent [46]) fits into the range of the total-power multiplier of this framework, as well as the achievable size of a facility as shown in Figure 2, and these two now give a range for the facility size and total-power-multiplier of this framework that is empirically bounded rather than assumed.

6. Main Discussion

6.1. From Siloed Optimization to A Unified Source–

Load–Storage–Grid Framework The review in Section 3 shows that the constituent sub-problems of renewable-aware data center planning are individually well studied, but that the interfaces between them are comparatively under-theorized. Storage-sizing studies typically take renewable capacity and load as given [27,28,29]; carbon-aware scheduling studies typically take physical renewable and storage capacity as given and optimize only the timing dimension [15,17,18,20]; and siting studies typically evaluate land and resource suitability without reference to co-located storage or to the specific temporal load profile of AI training and inference [32–35]. The framework proposed in Section 4 addresses this fragmentation by treating siting, sizing and scheduling as first- and second-stage decisions within a single stochastic program, following the structural logic of the more integrated recent contributions [9,10,11]. Figure 3 depicts this architecture schematically: an input layer combining workload, renewable-resource, grid-tariff/carbon-intensity and site-suitability data feeds a stochastic optimization core, which produces a coordinated set of siting, capacity and dispatch outputs, with realized operating data feeding back into subsequent planning cycles.

Fig 3: Schematic architecture of the proposed renewable-aware AI infrastructure planning framework, showing the input data layer, the stochastic optimization core, the resulting planning outputs, and the feedback loop that updates subsequent planning cycles.

6.2. The Interconnection Bottleneck as A Binding, Not

Merely Economic, Constraint One of the recurring patterns in the literature surveyed here is that the ability to connect to the grid is becoming increasingly a physical and administrative constraint and not just a cost parameter. Historically, the majority of renewable and storage capacity that is proposed has not been commercialized, and queue duration has grown as projects have been submitted despite the recent start of queue reductions to limit speculative volume [5,6]. The above has two ramifications for the framework proposed here. First, the grid import ceiling in the model formulation (Section 4) should be assigned a cost and a delay distribution similar to the risk-aware approach to interconnection-capacity co-sizing of [10] instead of treating it as an exogenous and freely available backstop. Second, there is an emerging rationale for on-site or directly contracted renewable generation and storage on-site, as those facilities are less susceptible to delays in the transmission line queue, than those that are part of a mix of renewables, a point that has also been raised in the microgrid-content literature or in the context of the corporate-profile approach to procurement [11,36,37].

6.3. Procurement Structure as A Planning Variable

The literature on the corporate power-purchase-agreements (PPA) that were reviewed in Section 3.6 indicates that physical and virtual PPA structures are not only a financial hedge-taking choice but are also related to the physical planning problem discussed in this chapter. Unlike virtual PPAs, which mainly help contract facilities with price risk, physical PPAs provide renewable electricity certificates along with power and are more closely aligned to objective of co-location and dispatchability of current framework [36]. The optimal PPA structure, as shown in the solution of the problem in [36] depends on demand and price uncertainty in a similar way to the way a physical siting and sizing depends on load and cost, and so a natural extension of the formulation in Section 4, identified as a direction for future work in Section 7, is to combine the decision on PPA structure with the siting and sizing of the physical plant into a single stochastic program, rather than making them separate decisions that could be solved sequentially with their own idealizations of uncertainty.

6.4. Forecasting Quality as A Precondition for

Framework Performance The forecasts for renewable resources and demand are crucial and are key factors that will limit the performance of any stochastic sizing and dispatch framework. As discussed in Section 3.6, literature in machine-learning forecasting indicates that accuracy improves albeit with some forecast-horizon dependence with a neural or ensemble forecasting approach over simpler statistical methods for both solar and wind energy forecasts [38,39,40]. Forecasting a range should not exclusively be based on the historical climatological average and is therefore not a second-order implementation detail but a direct influence on the ability to deliver a dispatch which closely resembles the realized availability of renewables, which the second stage of the proposed framework, the dispatch layer, covers. This second stage dispatches hourly, and other lower, degrees of freedom, and the forecast in this stage should consider the range, not just the trimmed mean of past supply and demand. Forecasting a range should not be a second-order implementation detail; it has a first-order influence on the ability to deliver-a-dispatch which closely resembles the realized-availability-of-renewables that is the dispatch layer’s second order concern; the degrees-of-freedom for which it is dispatched are at the hourly, and other lower, degree of freedom [14].

7. Challenges and Limitations

A number of limitations for reviewed literature and proposed frameworks are found. First, computational tractability is still a major restriction accounting for multi-site, multi-technology, multi-period, uncertainty simultaneously generates ‘combinatorially large’ scenario trees, which have prompted alternative, approximate and learning-based approaches to stochastic programming [12, 14] that do not fully reflect such uncertainty, but are still applicable within reasonable runtime ladders. Second, empirical evidence supporting the full integrated framework is still limited because of the availability of data: Section 5 compares individual parameters operational overhead and facility size to empirical evidence of actual, publicly reported data from Google [45] and Epoch AI [46] but no operator reports the joint renewable fraction, cost and reliability outcome of the full framework as an experimental result, which is why such an outcome is provided as a “directional expectation” drawn from the literature, just like several of the reviewed studies rely on synthetic or partially-reported operational data [9,11]. Third, the empirical scheduling literature offers an important corollary to respect within any capacity-planning framework: The attainable emissions and cost-savings from workload shifting, both temporally and spatially, are often much less than idealized simulation results, when taking the real-world nature of service levels, data localities and job-inter-dependency requirements into account [22]. The capacity-planning framework that does not explicitly consider workloads to be flexible with respect to dispatch decisions in the second stage of the planning process may exaggerate the portion of the load that can be served by renewables; modelling these workload flexible constraints in practice would necessitate direct modelling of the job-level flexible data-locality and deadline constraints found in the literature of workflow scheduling which is at a more abstract level than those found in the planning literature.

Fourth, the siting literature reviewed in Section 3.5 is based on utility scale renewable energy generation and does not usually consider siting without regard to data center cooling requirements; adding water-availability constraints explicitly for use in formulating data center siting requirements is identified as a gap and not a solved problem. Fifth, policy and regulatory uncertainty, both in terms of the rate at which changes occur in the interconnection-queue process, as well as in terms of how behind-the-meter electricity generation will be treated going forward, and the evolving regulatory landscape for grid-interactive, flexible operation of data centers (modeling of which is beginning to materialize in a significant capacity, building up to less capacity buildout [8] altogether), are another source of uncertainty that is hard to capture in a purely techno-economic optimization framework and would benefit from closer engagement with the policy-analysis literature. Lastly, the evidence-base of this review does not just consist of peer reviewed academic research but also of grey literature from energy agencies and consultative organizations, with findings based on the latter marked as such in the text; findings based on grey literature must be read in terms of the current industry practice, not as findings from peer-reviewed empirical research.

8. Conclusion

In this review, we have compiled the multidisciplinary literature relevant to the field of renewable-aware AI infrastructure planning and proposed a structured optimization approach (across energy infrastructure siting, renewable- and storage-capacity sizing, and grid interconnection) under AI-driven uncertainty in demand. The constituent sub-problems are well-developed and widely used: these include demand forecasting, generation-storage co-sizing, carbon-aware scheduling, multi-criteria siting and renewable procurement. Integrated frameworks that cover the complete source–load–storage–grid chain, however, are relatively new and scarce, and grid interconnection delay and cost has now become an uncertainty that must be explicitly captured in all techno-economic planning frameworks, instead of treated as a free backstop. Benchmarking the framework’s key parameters against Google’s disclosed fleet-wide PUE data and Epoch AI’s independently compiled database of the world’s largest AI data centers shows that operational overhead varies by a factor of roughly 1.4 across the industry (1.09 to 1.54), that individual facility scale now spans two orders of magnitude, and that the states attracting the largest AI data center buildout overlap substantially with the states carrying the longest interconnection queues empirical evidence that reinforces, rather than merely illustrates, the case for treating siting, sizing and interconnection as jointly optimized decisions. Finally, the actual outcome of the framework (the renewable fraction, cost, and reliability affecting a particular facility) is only a directionary expectation based on the published literature and not a verified empirical result as no operator has yet released the combined data set to test this result on a specific (real) facility. Future research should include three directions: experimental validation of integrated planning frameworks should be undertaken using operating data on facilities collected from the field where they are used (such as the data center database by Epoch AI), but this approach is one that is starting to become possible; the capacity-planning literature needs to explicitly represent procurement-contract structure, interconnection-queue dynamics and water-availability constraints as first-class decision variables rather than exogenous factors; and the literature on near-real-time scheduling needs to be coupled more closely with the capacity-planning literature, so that the workload flexibility assumed at the planning level is matched by what is considered attainable in operational practice. As AI-powered electric loads keep increasing relative to this rate of renewable interconnection, it is likely planning tools that optimize the siting of electric loads coupled to generation, storage and grid access will gain greater value, with all planning achieved in parallel and not sequential.