Executive summary

AI-based institutions—universities, hospitals, research labs, financial platforms, and public-sector agencies whose core operations depend on high-intensity AI compute—are rapidly turning into power institutions as well.

Global data-centre electricity demand is expected to roughly double by 2030, reaching ~900–1,400 TWh (around 3–4% of global electricity use). In the United States alone, data-centre load could reach ~100 GW by 2035, up sharply from previous forecasts. New AI-oriented facilities are already averaging ~47 MW of utility power per site, with projections of ~110 MW for typical new builds by 2030.

At the same time, interconnection queues are long, grids are strained, and environmental and social expectations are tightening. Big technology firms have responded by adopting an “all-of-the-above” power strategy: combining grid power, renewables, nuclear contracts, on-site gas generation, batteries, and microgrids to secure firm, low-carbon power for AI growth.

This white paper argues that no single power plant type is sufficient. The “best” infrastructure for AI-based institutions is a stack:

Firm, low-carbon baseload (nuclear, large hydro, high-capacity-factor renewables with long-duration storage or backing gas) to guarantee 24/7 availability. Local microgrid architectures combining grid connection, on-site generation (gas turbines/engines, fuel cells, or eventually SMRs), and battery storage for resilience and load flexibility. Aggressive efficiency and cooling design, as cooling alone can approach ~40% of data-centre energy use—higher for dense AI workloads. Flexible, software-defined energy management, using AI to orchestrate grid imports, storage, and on-site generation in response to price and reliability signals.

We outline design principles and reference architectures for three scales of AI-based institutions:

Tier 1 (1–5 MW): smaller universities, hospitals, regional government centres. Tier 2 (5–50 MW): major campuses, national labs, large financial or health systems. Tier 3 (50–500+ MW): hyperscale AI campuses and national AI infrastructure.

1. Context: Why AI institutions are different as power customers

1.1 Load characteristics of AI compute

AI-heavy institutions exhibit several distinct traits:

High power density: Dense racks with accelerators (GPUs/TPUs/custom ASICs) can draw 30–100 kW per rack, much higher than traditional enterprise loads. Near-continuous operation: Training and large-scale inference run 24/7 with limited downtime windows, making baseload and reliability critical. Rapid step-changes in demand: A new model, a new dataset, or a new institutional AI mandate can add tens of megawatts within a few years—or less. Cooling-intensive: AI workloads push cooling loads up; modern assessments suggest cooling can account for roughly 40% of facility energy, and AI can drive that higher without efficient system design.

The IEA estimates that data-centre electricity consumption could roughly double by 2030, growing around 15% annually—over four times faster than overall global electricity demand.

1.2 Strain on conventional grid-first strategies

Traditional “grid-first plus standby diesel” models are breaking down:

Grid connection delays: It can take many years to obtain interconnection rights in congested regions; data-centre projects are being delayed or cancelled because the grid cannot deliver the requested capacity. Local capacity constraints: Hyperscale AI facilities of hundreds of MW can exceed the capacity of local distribution networks. Carbon and air-quality concerns: Diesel backup is increasingly unacceptable to regulators and local communities when used beyond rare emergencies.

This is pushing AI-oriented institutions to behave less like passive ratepayers and more like energy developers in their own right.

2. Design principles for AI-supportive power infrastructure

Any “best-in-class” power plant strategy for AI-based institutions should be evaluated against six core criteria:

Reliability: Very high uptime (often targeting 99.999% or better) for both IT and cooling. Ability to ride through grid disturbances and regional outages. Scalability & modularity: Capacity to expand in 5–50 MW increments with minimal re-permitting. Modular blocks of generation and storage that can be added as AI demand grows. Low-carbon, long-term viability: Alignment with institutional sustainability commitments and regulatory trajectories toward decarbonization. Ability to maintain low lifecycle emissions even as load grows. Cost predictability: Long-term hedging against power price volatility via fixed-price PPAs or owned generation assets. Grid symbiosis: Ability to support and not destabilize the host grid—offering demand response, frequency support, and peak-shaving rather than only extracting capacity. Siting and social license: Minimal local air pollution, low visible impact where possible, and constructive roles in regional economic development (including repurposing brownfield power sites).

3. Core building blocks: Power plant options and their roles

3.1 Firm low-carbon power: Nuclear, hydro, and high-capacity-factor renewables

Nuclear is emerging as a central pillar in powering AI:

Major operators are signing long-term PPAs tied to existing and revived nuclear plants (e.g., Three Mile Island for Microsoft, Susquehanna for Amazon). Vendors and operators are exploring small modular reactors (SMRs) as campus-scale nuclear units in the 50–300 MW range, explicitly targeted at data-centre and industrial loads.

Advantages:

Very high capacity factor (often 90%+). Zero direct CO₂ emissions; strong alignment with 24/7 clean-energy goals. Long asset life (40–60+ years) matching multi-decade institutional horizons.

Challenges:

Long permitting and construction timelines. High upfront capital costs and political/regulatory uncertainties. Need for robust safety, security, and waste-management regimes.

For AI-based institutions with suitable sites or access to nuclear-heavy grids, long-term nuclear PPAs or, in the long run, co-sited SMRs are among the strongest candidates for “best” baseload.

Hydro and geothermal can play similar roles where geography allows, providing firm or near-firm low-carbon baseload, but these are more location-specific.

3.2 Dispatchable thermal generation: Gas turbines, engines, and fuel cells

To complement nuclear and renewables, institutions are increasingly considering:

High-efficiency gas turbines/engines (combined-cycle or reciprocating engines) located on or near campus. Fuel cells (e.g., SOFC or PEM) fed by natural gas, renewable gas, or hydrogen.

Analysts expect data centres to require more on-site generation as their load growth outpaces grid capacity, with utilities contracting fuel-cell and gas capacity as “bridge solutions.”

Gas turbines/engines:

Mature technology, fast to deploy relative to nuclear. Can be configured for combined heat and power (CHP) to feed absorption chillers for cooling. Emissions must be managed; low-NOx turbines and eventual fuel switching (e-methane, hydrogen blending) are important for long-term viability.

Fuel cells:

Higher electrical efficiency (often 50–60%) and potentially lower local pollutants than diesel or simple-cycle gas. Modular (MW-scale blocks) and can be arrayed into campus-scale microgrids. Capable of providing continuous power plus grid services.

3.3 Renewables plus storage

Most major AI/data-centre operators have strong renewable portfolios:

Companies like Meta have matched 100% of their power use with renewables since 2020, while others pursue 24/7 carbon-free energy goals.

For AI institutions, renewables are essential but not sufficient on their own:

Utility-scale wind/solar PPAs reduce average emissions and power cost but are intermittent. On-site or co-located solar can cover daytime cooling and non-critical loads. Battery energy storage systems (BESS) can shift renewable generation to evening peaks and serve as primary backup instead of diesel.

New developments show data-centre owners directly funding large batteries—sometimes 30 MW or more—to accelerate grid interconnection and secure both reliability and flexibility.

3.4 Microgrids and “power islands”

Microgrids knit together these generation resources into coherent, resilient systems:

They allow the campus to operate independently of the main grid during outages while coordinating on-site generation, renewables, and batteries. For the grid, they can act as stabilizing nodes, offering frequency response and peak-shaving instead of being pure loads.

Emerging practice (and expectation) is for large AI-oriented facilities to be built as microgrid-ready from day one, with:

Islanding capability for at least several hours to days. Black-start capability for IT and cooling. Control systems that prioritize critical loads and modulate non-critical AI tasks.

3.5 Cooling as a first-class design problem

Because cooling is often 30–40% of overall energy use—and even higher for dense AI clusters—“best” power plant infrastructure must be designed jointly with cooling:

High-efficiency liquid cooling and rear-door heat exchangers to reduce fan power. District-cooling style loops, where multiple buildings share cooling plant, improving load factor and reducing stranded capacity. Thermal storage (chilled water or phase-change materials) to shift cooling loads away from grid peaks and onto cheap or on-site generation.

As a rule of thumb, the “best” infrastructure treats cooling as co-generation: pairing CHP, heat reuse, and thermal storage to reduce electrical demand.

4. Reference architectures for AI-based institutions

4.1 Tier 1 (1–5 MW): Smaller universities, hospitals, regional agencies

Typical profile:

A teaching hospital or university AI lab that runs 1–2 small data rooms plus GPU clusters, with strong uptime needs but limited capital.

Recommended power strategy:

Grid connection as primary supply, ideally with: Long-term renewable PPAs (wind/solar) for carbon reduction. Participation in local demand-response programs. On-site generation: Small gas CHP or fuel-cell plant (1–5 MW) to cover critical loads and provide heat for campus buildings. Battery storage (1–3 MWh): To ride through short disturbances and support grid services. Cooling integration: Central chilled-water loops for AI facilities, potentially with ice storage for peak-shaving.

Why this works:

It leverages existing grid capacity while building a mini-microgrid around the core AI facilities, affordable for institutions that cannot finance large bespoke plants.

4.2 Tier 2 (5–50 MW): Major campuses and national research institutions

Typical profile:

A national lab, large university, or multi-hospital system running AI for research, diagnostics, planning, and services.

Recommended power strategy: “Campus microgrid with firm anchor”

Firm anchor supply: Long-term nuclear or hydro PPA where available, or contracts with gas plants capable of low-carbon fuels over time. On-site microgrid: 10–30 MW of gas turbines/engines or fuel cells, potentially in CHP configuration. 10–50 MWh of BESS for ride-through, peak-shaving, and participation in grid services. Renewables & storage: Co-located or nearby solar and wind to supply a portion of energy and reduce marginal emissions. Advanced energy management: AI-based energy orchestration that: Schedules non-urgent training jobs in low-price, low-carbon hours. Switches between grid, on-site, and stored energy in response to real-time prices and system constraints.

Why this works:

It balances institution-scale self-reliance with continued participation in the broader grid, avoiding over-build while achieving strong resilience and decarbonization.

4.3 Tier 3 (50–500+ MW): Hyperscale AI campuses and national AI infrastructure

Typical profile:

A multi-building AI mega-campus or national AI compute hub, with hundreds of MW of load and high public visibility.

Recommended power strategy: “Co-located power campus + microgrid”

Brownfield power-plant co-location where possible: Repurposing former coal or gas stations as AI campuses, leveraging existing transmission, cooling infrastructure, and land (e.g., Drax’s coal-era site conversion). This accelerates interconnection and can be politically attractive as a just-transition strategy. Dedicated firm generation: Direct access to one or more nuclear units (existing or SMRs), sized to cover much of the campus baseload. On-site or nearby gas-plus-CCS / fuel-cell capacity as flexible peaking and contingency resource. Large-scale microgrid & storage: 100+ MWh of BESS and potentially other storage (e.g., flow batteries, thermal storage) to shape load, provide black-start, and support the regional grid. Microgrid controls that can island a substantial portion of the campus for days. Renewables and transmission strategy: An “all-of-the-above” portfolio of wind, solar, and possibly offshore or geothermal, connected via new or existing high-voltage lines. Contract structures to ensure 24/7 clean-energy matching rather than annual offsetting.

Why this works:

At this scale, AI infrastructure is a power plant, and vice versa. The optimal solution is a unified power-and-compute campus, designed holistically to minimize emissions, maximize reliability, and actively support regional grid stability.

5. Implementation roadmap for AI-based institutions

5.1 Phase 1 (0–3 years): Risk triage and quick wins

Map criticality tiers of AI workloads and associated power requirements. Audit grid connection constraints and local capacity; identify realistic interconnection timelines. Deploy BESS and small-scale microgrids around the most critical facilities (e.g., hospitals, control centres) to reduce dependence on diesel. Sign initial renewable PPAs and investigate potential nuclear/hydro contracts where aligned with institutional mission. Optimize cooling and efficiency to reduce immediate load growth (better PUE, retrofits, liquid cooling).

5.2 Phase 2 (3–10 years): Campus-scale microgrids and firm supply

Develop campus microgrids tying together on-site generation, storage, and controllable loads. Co-locate with existing power infrastructure where possible—retired coal or gas plants, or industrial parks with strong transmission access. Scale up fuel-cell or gas-CHP capacity, with long-term plans for cleaner fuels and CCS where appropriate. Integrate AI-driven energy management, tightening the coupling between compute scheduling and energy availability.

5.3 Phase 3 (10–30 years): Deep decarbonization and nuclear/long-duration storage

Commission or contract SMRs or advanced reactors tailored to institutional campuses where regulatory and economic conditions permit. Deploy long-duration storage (e.g., flow batteries, hydrogen, pumped hydro) for multi-day coverage and seasonal balancing. Transition gas assets toward low-carbon fuels or phased retirement. Formalize “AI-ready grid” partnerships with utilities and regulators, treating AI campuses as grid resources rather than passive loads.

6. Key risks and trade-offs

Over-building vs. stranded assets: Over-estimating long-term AI power demand can strand capital in under-utilized plants; under-estimating can constrain institutional growth. Some analysts warn of speculative AI data-centre bubbles inflating load projections. Regulatory friction: Nuclear, gas, and major transmission lines face complex permitting processes—though AI is now being explored to accelerate nuclear permitting itself. Local community impacts: Siting large AI-power campuses can strain water resources (for cooling), change land use, and affect housing and labour markets. Brownfield strategies and transparent community benefit agreements can mitigate this. Cyber-physical security: As AI institutions become integrated power actors, their microgrids become critical infrastructure, demanding robust cyber and physical security architectures.

7. Conclusion: What “best” looks like

For AI-based institutions, the best power plant infrastructure is not a single technology but a layered architecture:

Foundation: long-term contracts or co-location with firm low-carbon sources (nuclear, hydro, high-CF renewables with storage). Middle layer: campus-scale microgrids integrating gas/fuel-cell generation, renewables, and substantial batteries. Top layer: AI-driven orchestration of compute and energy, plus highly efficient, integrated cooling systems.

This architecture allows AI-based institutions to:

Secure the reliability required for life-critical and mission-critical AI use. Scale in step with algorithmic and institutional ambitions. Align with decarbonization imperatives and community expectations. Evolve from grid burden to grid partner, stabilizing and decarbonizing the wider system.

For any given institution, the details will depend on geography, regulatory context, and mission. But the essential shift is universal: AI-based institutions must treat power plants, microgrids, and energy markets as core strategic infrastructure—not a background utility.