PaperlayerView exact revision on arXiv

Source-linked AI summary

Moving bits instead of watts: Accelerating global AI data-centre deployment at the energy-digital-water-land nexus

Jiawei Qu, Xiaoying Shi, Kai Hou, Hongjie Jia, Jinyue Yan

arXiv:2610.35088v1eess.SYcs.SY

TL;DR

Grid constraints and long connection queues threaten data-centre growth, while the global feasibility of transferring workloads to other regions remains uncertain. The paper models siting and workload allocation across resource and policy scenarios. It finds that computation transfer can bring most connection-delayed capacity online while reducing costs, emissions and transmission needs.

  • Problem

    How interdependent resource constraints shape the global feasibility and grid-connection benefits of computation transfer remains poorly understood, amid data-centre growth and grid-connection delays.

    Source: §1 · p. 4

  • Method

    A global framework estimates regional construction costs and optimizes data-centre siting and workload allocation across energy, water, land, connectivity and financing constraints.

    Source: §1 · pp. 4–5

  • Results

    More than 95% of delayed capacity can be brought online by 2030; construction costs fall 43.7–46.7%, energy costs 17.3–20.4% and CO₂ emissions 37.9%.

    Source: §1 · p. 5, §4 · p. 23

  • Takeaways & Limitations

    Computation-transfer-based siting provides an alternative to conventional grid-expansion-dependent planning by shifting computing demand toward regions with favourable energy and digital infrastructure.

    Source: §1 · p. 5

    The model imposes no explicit fibre-capacity constraints, and its deployment benefits depend on workload latency tolerance and service thresholds.

    Source: §4 · p. 23

Abstract

from arXiv · show

Artificial intelligence is driving an unprecedented expansion of data centres. However, this growth is increasingly stalled by grid transmission constraints that trap new data centres in years-long connection queues. Unlike electricity, computation can be moved across the globe through fibre-optic networks, but whether this spatial transfer can bypass constrained grids and accelerate global AI data-centre deployment remains unknown. Here we develop a global framework that places future data-centre capacity in energy-abundant regions and routes computational workloads through fibre-optic networks rather than transmitting electricity. We evaluate the framework through more than 11 million hourly operation simulations at subnational resolution, integrating constraints on energy systems, water, digital connectivity, land and financing risk. Results show that across multiple data-centre development and climate-policy scenarios, leveraging global data transfer allows new data centres to be sited where these conditions are favourable, lowering construction costs by 44-47% and carbon emissions by 38%. By treating the global fibre-optic network as an energy-system asset, it avoids 34-67 GW of high-voltage transmission capacity, comparable in magnitude to Italy's national peak electricity demand. Crucially, this methodology could bring more than 95% of connection-delayed data centres online in 2030, with Canada, the Nordic countries and the Arabian Peninsula emerging as preferred locations for data-centre development. Our findings establish computational mobility as an alternative to energy-system expansion, informing the coordinated global development of energy and artificial-intelligence infrastructure.

1. Introduction

Grid-connection delays threaten data-centre expansion, while workload transfer over fibre offers a way to site capacity where resources are favourable. The paper develops and evaluates this approach across global scenarios.

1. Introduction

  • The framework estimates costs across more than 4,000 subnational regions and jointly optimizes data-centre locations and workload allocation.It provides an economic basis for siting and an alternative to conventional grid-expansion-dependent planning.Source: §1 · pp. 4–5
  • The analysis draws on more than 11 million dispatch simulations across subnational regions worldwide.The framework is evaluated under multiple data-centre development and climate-policy scenarios.Source: §1 · pp. 4–5
  • More than 95% of capacity projected to remain connection-delayed in 2030 could be brought online through computation-transfer-based siting.Under continued local energy-system expansion, 10–15% of global data-centre capacity is projected to remain delayed.Source: §1 · p. 5
  • Construction costs fall by 44–47% and carbon emissions by 38% when data centres are built in regions with favourable energy, water, fibre and land conditions.The comparison is with energy-system expansion.Source: §1 · p. 5
  • 34–67 GW of additional transmission capacity is avoided by treating the global fibre-optic network as an energy-system asset.The avoided capacity is comparable in magnitude to Italy’s national peak electricity demand.Source: §1 · p. 5

2. Methodology

The methodology combines regional data and cost estimates with an optimization model that reallocates future data-centre capacity through computation transfer. It evaluates deployment across resource constraints, scenarios and operating and financing modes.

2. Methodology

  • The framework combines data inputs, construction-cost quantification and a computation-transfer-based siting model.It compiles 12,445 data-centre clusters and integrates energy, water, digital-connectivity and land information.Source: §2 · p. 5, Figure 1 · p. 6
  • Construction costs are quantified across more than 4,000 subnational regions for grid-supplied and self-generation deployment under local and multinational financing.Locations with high capital-recovery risk are excluded from candidate locations.Source: §2 · p. 5
  • The siting model reallocates future capacity from energy-constrained regions to feasible receiving regions while satisfying latency, energy, water and land constraints.The framework is assessed across four IEA data-centre development scenarios and nine IIASA climate-policy scenarios.Source: §2 · pp. 5–6
  • The optimization spans grid-supplied and self-generation deployment and local and multinational financing modes while minimizing total costs and delayed capacity.Climate-policy and data-centre development scenarios drive the model.Source: Figure 1 · p. 6

2.1. Data inputs

The input layers define the framework’s spatial scope and characterize data centres, energy, water, digital connectivity and land. Together, they inform where capacity can be deployed and at what cost.

2.1.1 Spatial and temporal scope

  • Construction costs are estimated at subnational scale across more than 4,000 regions, with siting conducted for 2030 and detail in major data-centre countries.The analysis groups the world into nine regions and focuses on 24 major countries.Source: §2.1.1 · pp. 6–7

2.1.2 Data centres

  • The baseline includes 12,445 data-centre clusters, with future capacity drawn from four IEA scenarios and ten representative cooling technologies.Technology-specific PUE and WUE estimates use hourly local climate data.Source: §2.1.2 · p. 7

2.1.3 Energy layer

  • Energy inputs cover power-system and gas-network access, reliability, generation costs and hourly demand, using detailed transmission networks in major countries.The reconstructed network includes more than 240,000 high-voltage lines and identifies potential connection delays rather than modelling global power flows.Source: §2.1.3 · p. 7, 1) Energy availability · p. 7, 2) Power system reliability · p. 7

2.1.4 Water layer

  • The water layer combines industrial freshwater withdrawals with technology-specific consumption intensities to estimate cooling and on-site generation water demand.Source: §2.1.4 · p. 8

2.1.5 Digital connectivity layer

  • Digital-connectivity inputs combine terrestrial and submarine fibre networks, landing stations and internet exchange points to represent workload access and latency.Source: §2.1.5 · p. 8

2.1.6 Land layer

2.2. Construction cost quantification

Construction costs combine facility components with regional adjustments, and the model checks its estimates against market data. Financing assumptions also screen out locations where capital costs exceed expected returns.

2.2.1 Data-centre construction costs

  • Construction-cost estimates include building, mechanical and electrical systems, general requirements and contractor fees, but exclude servers and tenant-owned IT equipment.Regional adjustments account for construction prices, imported mechanical equipment, cooling technology, grid reliability, gas access and digital connectivity; self-generation also adds land and on-site generation costs.Source: §2.2.1 · p. 9
  • The model’s comparison with estimates for 51 major markets is an external consistency check, with regional deviations of 6.2–13.9% and a global normalized deviation of 11.5%.Because model results are subnational and market estimates are primarily city-level, this is not a strict one-to-one validation.Source: §2.2.1 · pp. 9–10

2.2.2 Financial feasibility screening

  • Candidate locations are screened under local or multinational financing, and excluded when capital costs exceed expected returns.Local financing uses country-specific WACC; multinational financing combines an international AI-investment cost of capital with a country risk premium.Source: §2.2.2 · p. 10

2.3. Computation-transfer-based siting model

The model jointly optimizes data-centre siting and capacity by minimizing system costs and delayed capacity while linking long-term planning to hourly operation. It applies coupled energy, water, digital-connectivity and land constraints.

2.3. Computation-transfer-based siting model

  • The siting model optimizes data-centre locations and capacity through an hourly framework linking long-term planning with short-term operation and energy flows.Its objective combines total system costs with delayed data-centre capacity under coupled resource constraints.Source: §2.3 · p. 10, §2.3.1 · p. 10

2.3.1 Objective function

  • Annualized costs combine construction and operating costs for 2030, with construction annualized using financing-specific capital costs and the assumed lifetime.Operating costs include fixed and variable operation and maintenance, plus electricity costs; local and multinational financing use different discount-rate definitions.Source: §2.3.1 · p. 10

2.3.2 Energy layer constraints

  • Data-centre workloads can shift both across hours and locations, unlike conventional load, which shifts only across hours within each location.The model distinguishes fully flexible, latency-constrained spatially transferable, and ultra-low-latency workloads excluded from spatial transfer; accelerated-server workloads proxy AI-related demand.Source: §2.3.2 · p. 11
  • Hourly energy constraints bound generation by installed capacity and renewable availability, model network flows and transmission limits, and use storage for intertemporal balancing.Separate power balances represent the public grid, grid-supplied data centres, and self-generation data centres; water demand includes cooling and behind-the-meter generation.Source: §2.3.2 · p. 12, §2.3.3 · p. 12

2.3.4 Digital connectivity layer constraints

  • Latency-sensitive workloads must meet a 100 ms end-to-end threshold, with service interactions represented as a request-response round trip including network and processing delay.The model validates feasible transmission distances against independent benchmarks, with deviations below 0.6%.Source: §2.3.4 · p. 13

2.3.5 Land layer constraints

  • Generation capacity is constrained by suitable land area and resource quality, excluding protected and other unsuitable areas and bounding wind and solar potential.The land-suitability mask excludes forests, wetlands, water bodies, steep terrain, built-up areas, and—under conservative assumptions—cropland and high-value agricultural land.Source: §2.3.5 · p. 13

3. Results

Across 36 scenarios combining four data-centre development pathways and nine climate-policy pathways, the results compare local energy expansion with siting data centres in resource-favourable locations and transferring workloads. Computation-transfer-based siting lowers construction costs and emissions, while results also show that local expansion leaves connection delays and that outcomes depend on location and policy assumptions.

3. Results

  • The analysis combines four data-centre development pathways with nine climate-policy pathways, yielding 36 scenarios.The data-centre pathways span low-growth Headwinds to high-growth Lift-Off, with High Efficiency and Base representing efficiency-driven and central trajectories.Source: §3 · pp. 13–14

3.1. Quantification of data-centre construction costs

  • Energy factors dominate grid-supplied construction-cost differences in Europe and North America, while land becomes the key cost determinant under self-generation.In Europe, self-generation land-price factors add 1.81 USD W⁻¹; in water-stressed West Asia, water constraints add up to 1.50 USD W⁻¹.Source: Figure 2 · p. 15, Figure 3 · p. 16, §3.1 · p. 17

3.2. Cost increases and residual connection delays under local energy expansion

  • Local energy expansion leaves 10–15% of global data-centre capacity connection-delayed in 2030; construction costs are 40% and 54% higher in the top half of delay-share regions across scenarios.Global energy costs increase by 13–33%; D4 raises construction costs by around 5% versus D1 but lowers energy costs by approximately 4%.Source: §3.2 · pp. 17–18, Figure 4 · p. 18

3.3. Data-centre redistribution through computation-transfer siting

  • Nearly 90% of relocated capacity originates in established US and European hubs, while Canada, Nordic countries, the Arabian Peninsula and Latin America host most relocated capacity.Canada’s share falls from 46.7% in Headwinds to 29.3% in Lift-Off; Nordic capacity constraints shift much of Europe’s remaining relocated capacity to the Middle East in Lift-Off.Source: Figure 5 · p. 19, §3.3 · p. 20
  • Receiving regions use different supply modes: self-generation accounts for over 60% of new Nordic capacity, while grid supply accounts for approximately 90% in China.The passages attribute the Nordic pattern to renewable resources and land costs, and China’s to expanded transmission and generation infrastructure.Source: §3.3 · p. 20

3.4. Costs and emissions benefits through computation-transfer siting

  • Compared with local self-generation, computation-transfer-based siting lowers global construction costs by 45.8% in Headwinds and 43.7% in Lift-Off, and avoids about 34–67 GW of additional transmission capacity versus transmission expansion.In Headwinds, global construction costs fall from 17.7 to 9.6 USD W⁻¹; the avoided capacity is comparable in magnitude to Italy’s national peak electricity demand.Source: §3.4 · p. 21
  • 37.9% lower annual CO₂ emissions from new capacity in 2030, from 184.9 to 114.8 (101.7+13.1) Mt CO₂ yr⁻¹, accompany computation-transfer-based siting.The global reduction includes higher emissions in some receiving regions, with the Middle East and North Africa increasing by 29.4 Mt CO₂ yr⁻¹.Source: Figure 7 · p. 21, §3.4 · p. 22

Uncertainty analysis

  • More stringent climate policy increases delayed capacity by accelerating VRE penetration and the need for spatial balancing, transmission capacity and system flexibility.Siting geography remains broadly stable across economic, policy-implementation and climate-policy assumptions, while delayed-capacity magnitude is sensitive to climate policy.Source: Uncertainty analysis · p. 22

4. Discussion and Policy Implications

The discussion finds that computation-transfer-based siting can bring delayed capacity online while reducing construction and energy costs and emissions, but emissions outcomes vary by corridor. Its deployment potential depends on workload latency tolerance, fibre capacity, and cross-border governance.

4. Discussion and Policy Implications

  • Regions with connection-delayed capacity have 40-54% higher construction costs and 13-33% higher electricity costs than regions without delays; 10-15% of global capacity remains delayed by 2030 under energy-system expansion.The paper frames computation transfer as matching capacity to stronger resource endowments while routing workloads accordingly.Source: §4 · p. 23
  • More than 95% of delayed capacity can be brought online by 2030, while construction costs fall 43.7-46.7%, energy costs 17.3-20.4%, and CO₂ emissions 37.9%.The emissions reduction is not universal: relocating capacity from Europe to North Africa increases emissions by 16.8 Mt CO₂ yr⁻¹.Source: §4 · p. 23
  • A 10% change in latency-tolerant workload share changes delayed capacity by less than 1.8%; tightening the latency threshold from 100 ms to 50 ms raises remaining delayed capacity by around 3%.The analysis does not explicitly account for energy security, data security, or environmental compensation issues.Source: §4 · p. 23
  • Aggregate submarine-cable capacity is sufficient in most cases, but some corridors, including ARCOS, could face constraints under bandwidth-intensive multimodal transfer.The analysis does not impose explicit fibre-network capacity constraints.Source: §4 · p. 23
  • Planners should distinguish latency-tolerant workloads from real-time inference and interactive services to identify capacity that can move and capacity that must remain near demand hubs.The discussion also calls for coordination of cross-border data, computation, and infrastructure regulations, alongside carbon accounting and environmental responsibility.Source: §4 · p. 23
  • Realizing the deployment potential depends on removing national barriers to data flows and cross-border data-centre deployment while ensuring coordinated data-security governance.The paper also recommends frameworks for carbon accounting, carbon trading, and environmental responsibility in receiving regions.Source: §4 · p. 23

Appendix Fig. 4 Deployment model of data centres across major countries.

Appendix Fig. 5 summarizes avoided transmission capacity under computation-transfer-based siting across scenarios, reliability assumptions, equivalent transmission distances, and VRE penetration.

Appendix Fig. 4 Deployment model of data centres across major countries.

  • The heatmap reports avoided transmission capacity at a 200-km equivalent transmission distance across scenarios and three grid-reliability assumptions: no N-1 constraint, N-1, and N-2.Compare avoided capacity across the reliability assumptions and scenarios.Source: Appendix Fig. 4 Deployment model of… · p. 30
  • Upper plots show sensitivity to equivalent transmission distances from 100 to 1,000 km.Compare how avoided transmission capacity varies with the assumed equivalent distance.Source: Appendix Fig. 4 Deployment model of… · p. 30
  • The right-hand panel relates global avoided transmission capacity to VRE penetration, with points for individual scenarios and lines for fitted trends.Compare the scenario points against the fitted trends.Source: Appendix Fig. 4 Deployment model of… · p. 30
Loading 2610.35088v1…