Introduction
The current data center boom appears to be an early-stage, brute-force response to the first wave of artificial intelligence demand. The market has encountered a rapid increase in demand for training, inference, and AI-enabled cloud services, and the immediate response has understandably been physical scale in the form of larger campuses, more megawatts, denser racks, more cooling capacity, and expanded grid interconnection. The International Energy Agency projects that global data center electricity consumption could more than double to roughly 945 terawatt-hours by 2030, with AI as a central driver of growth (International Energy Agency [IEA], 2025). The United States is expected to account for a large share of that increase, making the issue not only a technology investment question, but also a power, land use, and infrastructure planning question.
There is little risk that demand for AI will disappear. It is more likely that demand will persist while the architecture of supply changes. Markets tend to attack cost, but efficiency does not automatically reduce aggregate demand. Hardware improvements, software efficiency, workload routing, and the separation of high-value from commodity workloads will reshape supply as the market develops a clearer understanding of AI demand. Simultaneously, reasoning models, agents, evaluation, and test-time compute may increase the amount of inference required for high-value tasks. AI will almost certainly require more compute. In question is whether today’s first-wave AI infrastructure is in the right places, with the right power economics, cooling systems, network position, and workload flexibility for the compute demand that actually emerges.
If future demand does not match first-wave infrastructure assumptions, part of the current buildout may age poorly. Some data center assets will remain valuable because they control power, fiber, land, and interconnection rights. Others may become lower-margin commodity compute sites. A smaller but still meaningful subset may become stranded: physically present, but unable to earn an economic return before the end of its expected life. This paper defines that risk in spatial finance terms, compares the current boom to earlier technology infrastructure cycles, identifies countervailing demand forces, and describes the data center assets most exposed to stranding over the next 15 years. It is a framework for asset-level reasoning, not an estimate of aggregate exposure.
Leading Indicators That Could Affect a Stranding Thesis
The composition of AI demand is the key variable to monitor. The issue is not simply whether inference grows. Inference can point in different infrastructure directions depending on its form. Routine inference favors smaller models, caching, routing, local execution, and proximity to users or data. Reasoning-heavy inference favors dense accelerator clusters, high utilization, and centralized infrastructure. The stranded-asset risk increases if monetized inference growth is primarily routine and distributive. It decreases if reasoning-heavy inference absorbs large amounts of centralized capacity.
A practical monitoring framework would track five indicators:
| Indicator | If it points toward centralization | If it points toward distribution |
| Inference composition | Reasoning-heavy inference, agents, coding systems, simulation, evaluation, and synthetic data consume a rising share of accelerator time. | Routine inference such as classification, summarization, extraction, retrieval, and workflow automation dominates production usage. |
| Training and reasoning capex | Large accelerator clusters, frontier training, post-training, and test-time scaling remain the primary drivers of infrastructure investment. | New investment shifts toward regional, enterprise, colocation, and inference-oriented facilities. |
| Small-model and open-weight adoption | Frontier models retain most high-value workloads and routine tasks continue to call centralized models. | Small, specialist, private, and open-weight models capture routine enterprise use. |
| Edge and device-local inference adoption | On-device AI remains mostly assistive, intermittent, or dependent on cloud escalation. | NPUs, edge accelerators, and local inference become standard for privacy, latency, bandwidth, and cost control. |
| Data center siting and acquisition patterns | Capital continues to favor very large centralized training and reasoning campuses with dense accelerator clusters. | Capital shifts toward metro, regional, and connectivity-rich facilities closer to users and enterprise data. |
Current indicators point in both directions, but not as a simple tie. They suggest a market in which aggregate centralized spending remains enormous while marginal workload composition becomes more inference-heavy. Dell’Oro Group reported in March 2026 that global data center capital expenditure rose 57% in 2025 and forecast that full-year 2026 data center capex would surpass $1 trillion, supported by more than 10 million high-end accelerators as a primary capex driver (Dell’Oro Group, 2026). Goldman Sachs Global Institute likewise modeled a baseline of roughly $7.6 trillion in cumulative AI infrastructure capital expenditure between 2026 and 2031 across compute, data centers, and power, while emphasizing that the estimate is assumption-sensitive rather than a demand forecast (Goldman Sachs Global Institute, 2026). These indicators support the view that centralized infrastructure remains strongly supported by current capital flows.
At the same time, the workload mix is changing. Deloitte projects that inference workloads will account for roughly two-thirds of AI compute in 2026, up from one-third in 2023 and half in 2025, while also arguing that most computation will still run on high-end chips in data centers or enterprise on-premises AI factories rather than consumer edge devices (Deloitte Center for Technology, Media & Telecommunications, 2025). Reuters described I Squared Capital’s May 2026 acquisition of 10 data center facilities across nine U.S. markets as a bet on AI inference and as evidence of a shift from large centralized training data centers toward facilities closer to end users (Reuters, 2026). Stanford’s 2025 AI Index adds the efficiency side of the picture: increasingly capable small models helped reduce the inference cost of GPT-3.5-level performance by more than 280-fold between November 2022 and October 2024, while hardware costs declined and energy efficiency improved (Stanford Institute for Human-Centered AI, 2025).
These indicators collectively point to a more specific version of the stranded-asset thesis. Total AI infrastructure demand may remain very large, and centralized reasoning and training facilities may remain highly utilized. The greater risk is that some first-wave assets are optimized for the wrong mix of workloads. Remote training-oriented mega-sites may be repriced if the marginal commercial workload shifts toward routine inference that values proximity, connectivity, enterprise integration, and reuse optionality. Conversely, stranding risk may be limited if reasoning-heavy inference, agents, simulation, evaluation, and synthetic data absorb centralized capacity faster than routine inference distributes.
Stranded Assets in Spatial Finance Terms
A stranded asset is generally understood as an asset that, before the end of its expected economic life, can no longer generate the economic return assumed when the investment was made. Carbon Tracker defines stranded assets as assets that are no longer able to earn an economic return prior to the end of their economic life because of market or regulatory changes (Carbon Tracker, n.d.). Lloyd’s uses a related definition, describing stranded assets as assets that suffer unanticipated or premature write-downs, devaluation, or conversion to liabilities (Lloyd’s, 2017). Although the concept is often associated with fossil fuels and climate transition risk, it is broader than that. Technology shifts, market structure changes, regulatory constraints, social opposition, input cost changes, and operational obsolescence can all strand assets.
The term should not be used too broadly. A facility that remains occupied but earns less than expected is not necessarily stranded. For this paper, four outcomes are distinct:
| Outcome | Definition |
| Performing asset | Meets or exceeds the original investment case. |
| Repriced asset | Remains useful and occupied, but earns a lower return than expected. |
| Impaired asset | Requires a write-down, major retrofit, renegotiated power terms, recapitalization, or change of use. |
| Stranded asset | Cannot earn an economic return before the end of its expected life, or becomes a liability through contractual, regulatory, operating, or market changes. |
This distinction addresses a central tension in infrastructure analysis. Efficiency gains can increase total use through a rebound effect while still damaging assets built around older cost, utilization, and margin assumptions. A full data center operating at lower margin is better understood as repriced. A facility facing uneconomic power costs, unusable cooling design, weak reuse options, or stranded supporting infrastructure is closer to true stranding.
Spatial finance treats asset location as financially material. The Oxford Sustainable Finance Group defines spatial finance as the integration of geospatial data and analysis into financial theory and practice (Caldecott, 2022). The underlying premise is that asset-level location data, combined with geospatial datasets, can reveal risks that are obscured by corporate-level disclosures or aggregated financial reporting (Oxford Sustainable Finance Group, n.d.). For data centers, that means the risk is not only whether the AI sector grows. It is whether a specific facility, in a specific location, with specific power, water, fiber, labor, tax, and regulatory conditions, remains competitive as AI infrastructure becomes more efficient and specialized.
In spatial finance terms, AI data center stranding should be assessed at the asset level. Relevant variables include grid interconnection status, power price, power source, transmission constraints, water availability, cooling design, fiber connectivity, distance to demand, hazard exposure, local political tolerance, labor availability, and reuse potential. A data center campus is not a generic building. It is a bundle of locational rights, energy dependencies, environmental exposures, technical design choices, and workload assumptions. Stranding occurs when those assumptions no longer support an economic return.
A simple spatial-finance screening framework would evaluate the following variables:
| Variable | Why it matters | Stranding risk indicator |
| Power price and volatility | Determines operating cost and competitiveness. | High fixed-cost power above regional alternatives. |
| Interconnection certainty | Determines whether projected capacity can be delivered. | Long queue position, delayed transmission upgrades, or uncertain energization. |
| Power source and emissions exposure | Affects regulatory, reputational, and customer constraints. | Dependence on high-emission generation without credible mitigation. |
| Water availability | Affects cooling feasibility and social license. | Water-stressed basin, drought exposure, or rising municipal opposition. |
| Cooling adaptability | Determines ability to support future rack densities. | Air-cooled or inflexible design with limited retrofit path. |
| Fiber and network proximity | Affects workload suitability and latency. | Weak connectivity for latency-sensitive inference or enterprise workloads. |
| Distance to demand | Matters for inference, regulated data, and edge-adjacent uses. | Remote site suited mainly to batch or training workloads. |
| Hazard exposure | Affects uptime, insurance, and resilience. | Flood, wildfire, heat, drought, storm, or seismic exposure. |
| Regulatory and social tolerance | Affects permitting and operating continuity. | Ratepayer exposure, tax backlash, local opposition, or permitting constraints. |
| Reuse optionality | Determines downside value. | Narrow design tied to one workload, one customer, or one hardware cycle. |
This distinction illustrates that the building itself may not be the primary stranded asset. The impaired asset may be the power contract, the substation investment, the cooling system, the gas generation asset, the local tax abatement, the land position, or the premium valuation attached to the site as scarce AI infrastructure. A facility can remain physically operational while the financial asset is repriced or impaired.
The screening framework can be illustrated with two stylized facilities. The first is a remote AI training campus with a large land position, a projected multi-hundred-megawatt load, dependence on new transmission or behind-the-meter generation, limited metro proximity, constrained water availability, and a design optimized for dense accelerator clusters. Such a facility may still be valuable if frontier training and reasoning demand remain concentrated and if power is delivered at a durable cost advantage. Under the framework above, however, its risk profile would be driven by interconnection certainty, water exposure, reuse optionality, customer concentration, and distance from latency-sensitive inference demand. Depending on those factors, it would likely remain useful but exposed to repricing or impairment if the expected workload mix shifts.
The second is a metro-adjacent inference facility with a smaller footprint, strong fiber connectivity, proximity to enterprise and regulated data environments, modular cooling, and multiple plausible uses across inference, private AI, retrieval, storage, and conventional cloud workloads. Such a facility may face higher land and power costs, but it has more reuse optionality and is better positioned for latency-sensitive and data-proximate workloads. Using spatial finance screening variables, it would be better positioned for durable value, though still sensitive to power price, cooling adaptability, and local regulatory tolerance. This comparison is not meant to classify any specific real-world asset. It shows how the spatial variables change the financial interpretation of two facilities that might both be described generically as AI data centers.
Historical Analogs from the Technology Sector
The data center boom has several analogs from this century. None is exact, and none predicts an outcome for AI. They show how demand shocks, infrastructure buildouts, efficiency gains, and workload reorientation interact.
The closest analog is the fiber optic overbuild around the dot-com boom. Telecommunications companies built capacity in anticipation of rapid internet demand growth. Demand was real, but the timing, routing, financing, and competitive assumptions were often wrong. Much of the resulting capacity sat unused as “dark fiber,” and many companies that financed the buildout failed. Over time, however, bandwidth demand grew into some of that infrastructure, while technologies such as dense wavelength division multiplexing increased the capacity of already-installed fiber (Hecht, 2016). The lesson for AI data centers is that infrastructure can be directionally right and financially mistimed. The asset may eventually become useful, but not necessarily at the valuation or return assumed by the original investor.
The analogy has limits. Dark fiber could sit idle at relatively low marginal cost and later be lit as demand caught up. AI accelerators face a much shorter economic life, with rapid hardware depreciation, capital cost, and operating expense even when utilization falls. The durable asset may be the land, power, fiber, and interconnection position while the vulnerable asset is often the silicon and the facility design built around it.
A second analog is enterprise server sprawl in the early 2000s, followed by virtualization, cloud migration, and more efficient data center operations. Organizations responded to new applications by buying more physical servers, often with low utilization. Efficiency improvements later changed the economics. The Uptime Institute reported that average power usage effectiveness improved substantially between 2007 and 2014 before gains slowed as easier improvements were absorbed (Uptime Institute, 2025). The American Council for an Energy-Efficient Economy similarly notes that data center energy use roughly doubled from 2000 to 2010, but rose much more modestly from 2010 to 2020 because of more efficient IT devices, virtualization, and movement of workloads into efficient large facilities (Nadel, 2025). The lesson is that initial demand is often met with capacity, while later demand is met through utilization.
A third analog is video streaming infrastructure. The rise of online video created enormous demand for bandwidth, storage, and delivery capacity. Cisco forecast that IP video would account for 82% of global IP traffic by 2022, reflecting how large the delivery problem had become (Cisco, 2018). The mature response was not simply more bandwidth. It included content delivery networks, adaptive bitrate streaming, caching, compression, and better codecs. The AI equivalent would be model routing, semantic caching, context reuse, smaller models, and workload placement. If AI inference becomes a distributed service rather than a centralized frontier-model interaction, some of today’s centralized capacity assumptions may weaken.
There are also counter-analogs. Cloud hyperscale infrastructure expanded aggressively and was repeatedly described as capital intensive, but enterprise software, mobile applications, analytics, streaming, and later AI absorbed large amounts of capacity. Mobile broadband followed a similar path. Network efficiency improved, but better networks enabled more usage, richer applications, and higher consumer expectations. In those cases, efficiency did not reduce the need for infrastructure. It expanded the market that infrastructure could serve.
The analogs therefore do not prove that AI data centers will be stranded. They show the range of possible outcomes. Premature capacity can be impaired, absorbed, repriced, or transformed into durable infrastructure. The differentiating variable is whether the asset remains aligned with the next architecture of demand.
Risk Assessment Over the Next 15 Years
Over a 15-year horizon, the risk of stranded AI data center assets is material but uneven. It is highest where facilities are built around permanent scarcity of inefficient compute. It is lower where sites control durable power access, strong fiber connectivity, flexible cooling design, and the ability to support multiple workload types.
The most exposed assets are generic powered shells in weak locations. A remote site with cheap land but expensive power, limited water, poor fiber, weak labor access, or constrained transmission may be valuable during a capacity panic but less attractive in a more mature market. The same is true for rural mega-campuses built primarily around training workloads. Training can tolerate remote locations better than latency-sensitive inference, but training demand may concentrate among fewer firms, fewer model classes, and more specialized facilities. If monetizable AI volume shifts toward inference, workflow automation, enterprise-private AI, or edge-adjacent services, remote mega-campuses may fall into lower-margin uses.
Stranding risk is also shaped by capital-cycle behavior. The most durable assets are likely to be controlled by operators with deep demand visibility, flexible balance sheets, and the ability to adjust siting and design as workloads evolve. Greater risk sits with speculative developers, miners converting facilities to AI, merchant data center projects, neoclouds dependent on narrow customer demand, and utility or power infrastructure built around aggressive load forecasts. AI infrastructure is not only a technical response to demand, it is also a capital cycle, and capital cycles can produce assets whose financial assumptions move faster than their physical usefulness.
Power infrastructure presents another major risk category. IEEFA has warned that utility planning may be assuming substantially more data center demand than the technology industry itself is projecting, creating risk of overbuilt fossil fuel and grid infrastructure (Kunkel & Wamsted, 2025). The Belfer Center has also described AI data centers as a major challenge for U.S. grid planning because of the scale, speed, and uncertainty of load growth (Belfer Center, 2026). In this context, stranded assets may sit outside the data center balance sheet. Generation assets, substations, transmission upgrades, gas infrastructure, and long-term power commitments can all be stranded if load forecasts prove too high, facilities are delayed, workloads migrate, or efficiency gains reduce demand per unit of AI output.
Cooling and thermal design are also central. AI rack densities are rising, but facilities vary widely in their ability to support high-density liquid-cooled workloads. Uptime Institute’s 2025 survey found that average PUE levels had shown little change for six consecutive years, with improvement constrained by legacy infrastructure and regional barriers to efficient cooling (Uptime Institute, 2025). A facility designed for a prior thermal envelope may remain usable, but it may not be competitive for dense AI workloads without expensive retrofit. Conversely, a facility designed too narrowly for today’s highest-density GPU clusters may face reuse challenges if future systems change rack, power, cooling, or networking requirements.
Water-dependent cooling assets deserve particular attention. Data centers located in water-stressed regions or areas with rising political sensitivity around water use may face operating restrictions, higher costs, or retrofit requirements. Spatial finance analysis is applicable here because hydrology, drought exposure, competing municipal demand, and local regulatory tolerance are all location-specific.
The lowest-risk assets are those that combine durable power rights, strong transmission access, reliable fiber, flexible cooling, modular design, and credible reuse pathways. In those cases, the primary asset is not the building. It is the site’s position within the energy and digital infrastructure landscape. Such assets may survive changes in workload composition because they can shift among training, inference, batch processing, cloud services, simulation, rendering, storage, or other compute-intensive uses.
A practical 15-year risk typology would divide assets into four tiers:
| Tier | Conditions | Likely outcome |
| Tier 1: Adaptive platform | Durable low-cost power, strong fiber, modular cooling, high retrofit capacity, multiple workload options, and credible long-term demand access. | Retains premium value and remains financeable. |
| Tier 2: Repriced commodity site | Operable and occupied, but with weaker location, less efficient design, higher power cost, or narrower workload suitability. | Lower margins, lower valuation, but continued use. |
| Tier 3: Impaired facility | Requires major retrofit, power renegotiation, recapitalization, customer replacement, or change of use. | Write-down or restructuring likely. |
| Tier 4: Stranded commitment | Uneconomic power, unusable or obsolete design, blocked expansion, weak reuse path, unsupported public infrastructure, or loss of social license. | Liability, abandonment, forced conversion, or cost transfer. |
This typology is qualitative rather than predictive. It does not estimate the fraction of the buildout at risk. Its purpose is to distinguish physical occupancy from financial performance and to identify the asset-level conditions that separate durable infrastructure from vulnerable infrastructure.
This paper also does not attempt to assign stranded-asset exposure across capital structures. That question requires separate treatment because the economic loss may sit with different parties depending on the financing and contractual structure of the project. A data center developer may bear exposure through land, shell, fit-out, and customer concentration. A hyperscaler or tenant may bear exposure through long-term lease or reserved-capacity commitments. A utility may bear exposure through generation, transmission, and substation investments, while local governments and ratepayers may bear exposure through incentives, infrastructure subsidies, or regulated cost recovery. The present analysis focuses on spatial and workload conditions that create stranding risk. A subsequent analysis should examine who bears that risk once it is transmitted through leases, power contracts, debt, equity, utility regulation, and public incentives.
Countervailing Demand in Reasoning, Agents, and Test-Time Compute
The efficiency thesis has an important counter-force. Some AI workloads are becoming less expensive per task, but others are becoming more compute-intensive. Reasoning models explicitly use additional inference-time computation to improve output quality. OpenAI described o1 as improving with both train-time compute and more time spent thinking at test time (OpenAI, 2024). RAND similarly describes reasoning models such as o1, o3, and DeepSeek-R1 as using test-time compute to produce intermediate reasoning and improve performance on complex tasks (Duke et al., 2025). OpenAI’s o3 and o4-mini release also positioned reasoning models for complex math, coding, visual, and tool-using tasks (OpenAI, 2025).
This trend cuts against a simple efficiency narrative. If the market increasingly values multi-step reasoning, coding agents, simulation, evaluation, tool use, verification, and agentic workflows, then some inference demand may become more computationally intensive, not less. A coding agent may generate code, run tests, inspect errors, revise its approach, call tools, and repeat that loop. A scientific or financial reasoning model may spend far more compute per answer than a conventional chatbot. Evaluation and synthetic data pipelines can also create additional demand that sits between training and inference.
The result is likely to be bifurcation rather than uniform efficiency. Routine inference may become cheaper, smaller, more local, and more distributed. High-value reasoning may become more compute-intensive, centralized, and dependent on dense accelerator clusters. This distinction changes the stranded-asset thesis. The argument is not that aggregate AI compute demand will fall. The argument is that future AI demand may not map cleanly onto the first wave of centralized data center buildouts.
A facility optimized for massive training may remain valuable if frontier training and reasoning demand continue to scale. A facility optimized for generic AI capacity may be more exposed if routine inference migrates to smaller models, enterprise environments, regional facilities, or edge devices. The relevant risk is mismatch: between facility design and future workload mix, between site geography and workload geography, and between power commitments and realized compute revenue.
The relative magnitude of these forces remains unresolved. If high-value reasoning, agents, simulation, evaluation, and synthetic data grow faster than routine inference becomes cheaper, centralized AI infrastructure may remain highly utilized and stranding may be limited. If routine inference dominates monetized usage, and if smaller models, caching, routing, and local execution absorb a larger share of demand, first-wave centralized facilities may be repriced or impaired. The stranded-asset risk turns on the composition of demand, not aggregate AI adoption.
Efficiency Drivers Already on the Horizon
The innovations most likely to reprice or strand first-wave AI infrastructure are not speculative. They are visible in current practice and early market direction. They should be understood alongside the countervailing rise of reasoning and test-time compute, not as a uniform reduction in AI demand.
The first is the rise of smaller and more specialized models. Small language models are increasingly attractive for enterprise use because many tasks do not require frontier-model reasoning. Bounded workflows such as classification, extraction, routing, summarization, document question answering, geospatial enrichment, and policy compliance can often be handled by smaller models or specialist systems. A survey of small language models notes that they are favored for low inference latency, cost effectiveness, customization, and adaptability (Wang et al., 2024). If smaller models absorb a large share of routine inference, the market will still use AI, but the required infrastructure per transaction may fall.
The second is model compression. Quantization, pruning, distillation, sparsity, and related techniques reduce compute, memory, and bandwidth requirements. These techniques are already central to inference optimization. Their effect is to reduce the amount of hardware and power required for a given level of service. The result may not be lower total AI usage. It may be much higher usage at lower unit cost. That distinction matters. Efficiency can increase total demand while still impairing assets that were underwritten on older cost-per-token assumptions.
The third is inference routing and model cascades. Mature AI systems are unlikely to send every request to the largest available model. They will use small models first, escalate only when needed, retrieve before generating, cache repeated context, batch nonurgent jobs, and reserve frontier models for tasks that justify their cost. This architecture resembles the maturation of video delivery, where caching and routing became as important as raw capacity. It also changes geography. Some inference will be centralized, but some will move closer to users, regulated data, enterprise systems, or edge devices.
The fourth is context caching and reuse. Repeatedly processing the same documents, prompts, embeddings, and intermediate reasoning is wasteful. Systems that cache context, reuse embeddings, maintain durable state, and separate deterministic processing from model reasoning will reduce redundant compute. Over time, the AI stack will look less like a series of isolated model calls and more like an optimized distributed system.
The fifth is hardware and system-level efficiency. New accelerators, lower-precision inference, improved memory systems, liquid cooling, rack-scale architectures, and better workload schedulers all attack energy per useful output. The important point is not that every vendor claim will hold. It is that the direction of competition is clear. Power is now a binding constraint. Hardware and systems companies have strong incentives to deliver more AI output per watt, per dollar, per rack, and per square foot.
The sixth is workload separation. The current phrase “AI data center” hides meaningful differences among training, inference, fine-tuning, evaluation, synthetic data generation, simulation, retrieval, embedding, and batch analytics. These workloads do not all need the same site, hardware, network, cooling system, or latency profile. As the market matures, facilities optimized for one workload may not command a premium across the whole AI stack.
An informative analogy is the recurring movement between centralized and distributed computing. Mainframes and terminals concentrated compute. Client/server architectures split workloads between servers and local machines. Thick clients used growing local compute capacity for richer application logic. Cloud recentralized many workloads because scale, operations, deployment, and utilization improved. Edge computing moved some workloads outward again because latency, resilience, privacy, bandwidth, and physical-world interaction mattered. AI is likely to follow a similar pattern. Frontier training and high-value reasoning may remain centralized, while routine inference, sensitive inference, and latency-sensitive inference move toward regional, enterprise, device-local, or edge environments.
Conclusion
The current AI data center boom reflects real demand, but it also reflects the inefficiency and uncertainty of a young market responding quickly to a new infrastructure requirement. The first answer to AI demand has been more physical capacity. The next answer will be more complicated, including better utilization, more efficient models, smarter routing, improved cooling, stronger workload segmentation, and, at the high end, more compute-intensive reasoning and agentic workflows.
The stranded-asset risk over the next 15 years is therefore material but uneven. The highest risk lies in infrastructure built for a permanent shortage of one kind of inefficient AI compute, such as remote mega-sites with weak geographic advantages, power infrastructure justified by aggressive load forecasts, water-dependent cooling assets in constrained regions, and facilities whose design is tightly coupled to current GPU-era assumptions. The lowest risk resides in adaptable assets with durable power access, strong fiber, flexible thermal design, and credible reuse across multiple compute workloads.
The analogs from fiber, enterprise server sprawl, streaming video, cloud hyperscale, and mobile broadband point to a range of possible outcomes. Demand shocks can produce stranded assets, but they can also produce infrastructure that is eventually absorbed. Efficiency can reduce unit costs, but it can also expand demand. AI is unlikely to produce a simple overbuild story. More likely, it will produce a tiered geography of compute with premium adaptive sites, repriced commodity sites, impaired facilities, and stranded supporting infrastructure.
Spatial finance provides the right framework because the outcome will be decided asset by asset, location by location. The question is not whether AI will use more compute. It will. The question is whether a specific facility, in a specific place, with specific power, water, cooling, fiber, regulatory, and reuse characteristics, remains aligned with the workload architecture that emerges.
References
Dell’Oro Group. (2026, March 17). Data center capex surges 57 percent in 2025 as AI deployments accelerate, according to Dell’Oro Group. https://www.delloro.com/news/data-center-capex-surges-57-percent-in-2025-as-ai-deployments-accelerate/
Deloitte Center for Technology, Media & Telecommunications. (2025, November 18). Why AI’s next phase will likely demand more computational power, not less. Deloitte Insights. https://www.deloitte.com/us/en/insights/industry/technology/technology-media-and-telecom-predictions/2026/compute-power-ai.html
Belfer Center for Science and International Affairs. (2026). AI, data centers, and the U.S. electric grid: A watershed moment. Harvard Kennedy School. https://www.belfercenter.org/research-analysis/ai-data-centers-us-electric-grid
Caldecott, B., McCarten, M., Christiaen, C., & Hickey, C. (2022). Spatial finance: Practical and theoretical contributions to financial analysis. Journal of Sustainable Finance & Investment. https://doi.org/10.1080/20430795.2022.2153007
Carbon Tracker. (n.d.). Stranded assets. https://carbontracker.org/terms/stranded-assets/
Cisco. (2018). Cisco Visual Networking Index: Forecast and trends, 2017–2022. https://www.cisco.com/c/dam/m/en_us/solutions/service-provider/vni-forecast-highlights/pdf/Global_Device_Growth_Traffic_Profiles.pdf
Duke, B., Walent, A., & Mazarr, M. J. (2025). When AI takes time to think: Implications of test-time compute for AI policy. RAND Corporation. https://www.rand.org/pubs/commentary/2025/03/when-ai-takes-time-to-think-implications-of-test-time.html
Goldman Sachs Global Institute. (2026, May 1). Tracking trillions: The assumptions shaping the scale of the AI build-out. Goldman Sachs. https://www.goldmansachs.com/insights/articles/tracking-trillions-the-assumptions-shaping-scale-of-the-ai-build-out
Hecht, J. (2016). Boom, bubble, bust: The fiber optic mania. Optics & Photonics News. https://internethistory.org/wp-content/uploads/2020/01/OSA_Boom.Bubble.Bust_Fiber.Optic_.Mania_.pdf
International Energy Agency. (2025). Energy and AI. https://www.iea.org/reports/energy-and-ai
Kunkel, C., & Wamsted, D. (2025). Risk of AI-driven, overbuilt infrastructure is real. Institute for Energy Economics and Financial Analysis. https://ieefa.org/resources/risk-ai-driven-overbuilt-infrastructure-real
Lloyd’s. (2017). Stranded assets. https://www.lloyds.com/strandedassets
Nadel, S. (2025). Opportunities to use energy efficiency and demand flexibility to reduce data center energy use and peak demand. American Council for an Energy-Efficient Economy. https://www.aceee.org/white-paper/2025/10/opportunities-use-energy-efficiency-and-demand-flexibility-reduce-data-center
OpenAI. (2024). Learning to reason with LLMs. https://openai.com/index/learning-to-reason-with-llms/
OpenAI. (2025). Introducing OpenAI o3 and o4-mini. https://openai.com/index/introducing-o3-and-o4-mini/
Oxford Sustainable Finance Group. (n.d.). Spatial finance. University of Oxford. https://sustainablefinance.ox.ac.uk/research/spatial-finance/
Reuters. (2026, May 26). I Squared bets on AI inference with $225 million data center buy from Cogent. https://www.reuters.com/business/media-telecom/i-squared-bets-ai-inference-with-225-million-data-center-buy-cogent-2026-05-26/
Stanford Institute for Human-Centered AI. (2025). The 2025 AI Index report. Stanford University. https://hai.stanford.edu/ai-index/2025-ai-index-report
Uptime Institute. (2025). Uptime Institute global data center survey 2025. https://datacenter.uptimeinstitute.com/rs/711-RIA-145/images/2025.Annual.Survey.Report.pdf
Wang, F., Zhang, Z., Zhang, X., Wu, Z., Mo, T., Lu, Q., Wang, W., Li, R., Xu, J., Tang, X., He, Q., Ma, Y., Huang, M., & Wang, S. (2024). A comprehensive survey of small language models in the era of large language models: Techniques, enhancements, applications, collaboration with LLMs, and trustworthiness. arXiv. https://arxiv.org/abs/2411.03350
