The Inconvenient Truth of the AI Industry: Part 2 – The Physical Infrastructure Wall, Datacenter Power/Water Crises, and Enterprise Green AI Strategies

1. Introduction: The Collision of Virtual Algorithms and the Physical World

Discussions surrounding the generative AI ecosystem have predominantly centered on virtual breakthroughs: parameter scales, benchmark leaderboards, and algorithmic architectures. Yet beneath the software layers training and serving Large Language Models (LLMs) lies a colossal physical reality: tens of thousands of power-hungry accelerator chipsets, megawatt-scale power feeds, and cooling systems circulating thousands of liters of water every second.

This physical friction is not confined to any single nation. The early warnings of grid saturation and public pushback that first erupted in the United States—the vanguard of global AI infrastructure—are manifesting with even greater intensity in South Korea, where high population density and hyper-concentration in the Seoul Metropolitan Area have triggered sharp legal disputes over 154kV ultra-high-voltage transmission lines and substation grid exhaustion. In this article, we first analyze empirical data from the US and the energy crisis confronting global Big Tech. We then turn that lens onto the structural contradictions facing Korea's datacenter ecosystem, culminating in practical Green AI engineering strategies that enterprise architects must implement today.

71% of American Adults Oppose Local AI Datacenters: The Rise of NIMBYism

According to an empirical Gallup survey of US adults published in May 2026 (Americans Oppose AI Data Centers in Their Area, Jeffrey M. Jones), 71% of respondents oppose the construction of AI datacenters in their local communities. Remarkably, 48% expressed 'Strong Opposition'—transcending mild skepticism—while only 27% supported local construction (with a mere 7% strongly supporting).

Critically, this opposition is 18 percentage points higher than opposition to building nuclear power plants (53%), a traditional lightning rod for local resistance. This underscores that local communities now perceive AI datacenters not as benign high-tech facilities, but as severe environmental and infrastructural hazards.

Traditional cloud datacenters historically integrated into municipalities with relative ease due to silent operations and minimal traffic generation. Modern AI datacenters, however, ignite fierce disputes over municipal resource allocation.

Half of all opponents (50%) cite acute fears of electricity and municipal water depletion. Demands for gigawatt-scale interconnects and evaporative cooling towers consuming millions of gallons daily stoke legitimate fears of crippled public utilities. Compounding this resistance are persistent low-frequency drone noises (22%) from industrial chillers and cooling fans operating 24/7, alongside deep anxiety (20%) that utilities will pass astronomical grid-expansion capital costs directly onto residential ratepayers.

In Northern Virginia's 'Data Center Alley'—the world's largest datacenter market—contentious legal battles and permit delays have erupted between state regulators, utility providers (Dominion Energy), and citizen coalitions over new high-voltage transmission lines. The software industry's ambition to infinitely expand virtual intelligence is running headlong into the hard physical limits of power grids and land availability.

2. In-Depth Analysis: The 3 Structural Bottlenecks of Physical Infrastructure

1) Grid Saturation and the Interconnection Queue Bottleneck

The power density required by modern AI facilities differs fundamentally from legacy cloud architectures. Standard enterprise compute racks draw 5–10kW per rack, whereas cutting-edge clusters (such as Nvidia's NVL72 and next-generation liquid-cooled architectures) demand 100kW to 130kW+ per rack. Projects aggregating over 100,000 top-tier GPUs demand gigawatt-scale interconnects—comparable to the entire output of an operational commercial nuclear reactor (~1,000MW).

Data from the International Energy Agency (IEA) and the US Federal Energy Regulatory Commission (FERC) reveals that this demand surge has triggered acute supply bottlenecks. Even after securing physical land, the lead time to interconnect with regional transmission grids (PJM, ERCOT, CAISO) now averages 5 to 7 years. Exacerbating this backlog, global lead times for large power transformers (LPTs) have exploded from 1–2 years to 3–4+ years, stranding deployed GPU hardware in dark datacenters without energized circuits.

2) Water Evaporation and the Limits of Direct-to-Chip Liquid Cooling

Roughly 30–40% of datacenter electricity is consumed not by computing FLOPs, but by cooling systems dissipating colossal thermal loads. Traditional air cooling reaches physical limits beyond 30kW per rack due to air's low heat capacity. Consequently, the industry is migrating rapidly toward Direct-to-Chip (D2C) liquid cooling coupled with evaporative cooling towers.

This shift triggers severe water depletion. A typical hyperscale campus evaporates between 3 million and 5 million gallons (approx. 11M to 19M liters) of cooling water daily—equivalent to the municipal water consumption of a city of 30,000 to 50,000 residents. Because datacenters frequently cluster in arid areas with cheap land and available transmission access (Arizona, Texas, Utah), excessive water withdrawals accelerate aquifer depletion and trigger fierce municipal resistance.

3) Big Tech's Energy Race: Nuclear PPAs, SMRs, and the Net-Zero Paradox

Confronting 5-year interconnection delays, hyperscalers have bypassed public queues by executing direct Power Purchase Agreements (PPAs) with dedicated baseload generators.

Microsoft signed an unprecedented 20-year PPA to purchase the entire 835MW output of the restored Three Mile Island Unit 1 nuclear plant in Pennsylvania, scheduled to restart by 2028. Google partnered with Kairos Power to deploy a fleet of 6–7 Small Modular Reactors (SMRs) totaling 500MW, with deliveries starting in 2030.

Yet dedicated generation is not a frictionless silver bullet. Amazon (AWS) acquired Talen Energy's 960MW nuclear-connected campus adjacent to the Susquehanna Nuclear Station for $650 million. However, in November 2024, FERC rejected an Interconnection Service Agreement (ISA) amendment that would have increased direct co-located draw from 300MW to 480MW, citing risks to public grid reliability and cost-shifting to consumers. This landmark ruling highlights the severe regulatory headwinds confronting off-grid nuclear datacenter models.

Furthermore, while Big Tech markets '24/7 carbon-free nuclear power,' reactor restarts and SMR commercialization will not come online until 2028–2035 at the earliest. To satisfy exploding real-time inference demands, utilities in Virginia and Texas are extending the operational lives of retiring coal and natural gas plants. As disclosed in official corporate sustainability reports, greenhouse gas emissions from Microsoft, Google, and Meta have surged 30% to 50%+ since 2020, revealing the stark contradiction between self-imposed '2030 Net-Zero' pledges and physical infrastructure realities.

4) The Impossible Trilemma in South Korea and 154kV Transmission Disputes

South Korea's datacenter ecosystem faces even more acute social and physical friction due to limited land area, dense high-rise residential zoning, and a unified state-owned power monopoly (KEPCO).

Over 70% of Korea's operational and planned datacenters are concentrated in the Seoul Metropolitan Area. As metropolitan substations reached absolute saturation, the Korean government amended the Enforcement Decree of the Electric Utility Act in March 2023 (Article 5-5), granting KEPCO the statutory authority to refuse or defer interconnection requests exceeding 5,000kW if they threaten grid stability. Consequently, numerous large-scale datacenter initiatives in the capital region have stalled following KEPCO rejection notices.

Even more volatile is the conflict over urban residential proximity. Unlike overseas facilities built in rural deserts, domestic operators seek ultra-low latency by burying 154,000V (154kV) ultra-high-voltage transmission cables beneath residential apartment complexes and elementary school transit routes:

• Anyang (Hogye-dong / Pyeongchon): Thousands of residents protested 154kV underground cable installations near apartments and schools, filing public audit requests with the Board of Audit and Inspection. Following protracted administrative litigation, the developer ultimately abandoned the permit and sold the parcel.
• Goyang (Ilsan Deogi-dong / Siksa-dong): Fierce public outcry against datacenter construction adjacent to high-density apartment blocks prompted municipal leadership to review revoking building permits and reject construction commencement filings, escalating into administrative court battles.
• Gimpo (Gurae-dong) and Yongin (Jukjeon): Persistent administrative and civil disputes drag on over low-frequency cooling tower noise, white plume vapor emissions blocking sunlight, and depreciating residential asset valuations.

In June 2024, the Korean government enacted the 'Distributed Energy Promotion Special Act' to incentivize relocation to non-capital regions such as the Gangwon Hydro-Thermal Cluster and Jeonnam Haenam Solasido. However, private IT enterprises resist decentralization due to overwhelming 'Data Gravity.' Between 85% and 90% of domestic e-commerce, fintech, gaming, and SaaS startups reside in Pangyo and Seoul. Hosting servers in non-capital regions imposes round-trip latency penalties (RTT 8–10ms) for capital-region users, eroding product competitiveness.

Severe talent shortages further compound this inertia: senior DevOps/MLOps talent flatly refuses to relocate to rural provinces, and laying dual terabit-scale dedicated dark fiber routes to regional sites incurs astronomical civil engineering costs.

International precedents highlight the severe economic toll of arbitrary infrastructure moratoria. Ireland instituted a four-year grid connection freeze in Dublin in 2021, which resulted in the loss of €6.5 billion (approx. 9.5 trillion KRW) in Foreign Direct Investment (FDI) and drove fintech headquarters elsewhere. Dublin was forced to repeal the moratorium in late 2025 with strict on-site generation caveats. Singapore similarly suspended datacenter development for three years, only to lift the freeze under strict PUE < 1.3 standards when regional hub status threatened to migrate to Johor, Malaysia.

According to Tortoise Media and Stanford's HAI AI Index, while South Korea ranks world-class in AI patents per capita and telecommunications infrastructure, it ranks outside the top 10 in private investment and talent retention. Grid bottlenecks restricting top-tier GPU accelerator clusters are now identified as the single greatest threat to national AI competitiveness.

Acknowledging Metropolitan Realities and 3 Pragmatic Agendas

South Korea faces an 'Impossible Trilemma': "Building in the capital region is impossible due to grid saturation; building in rural regions is commercially unviable due to data gravity holding 90% of traffic; and halting development starves the nation of sovereign AI compute."

Declarative regulations commanding enterprises to "simply go south" fail in the face of market physics. South Korea must accept metropolitan concentration as an immutable reality and re-engineer infrastructure and policy around it, focusing on three pragmatic engineering agendas:

1. Permitting Urban On-Site Power Generation: If KEPCO's grid is maxed out, regulations must permit on-site hydrogen fuel cells and eco-friendly LNG Combined Heat and Power (CHP) plants on datacenter parcels. Subsidies and carbon offsets must bridge the cost gap between expensive imported LNG (20–30% higher than industrial grid tariffs) and global clients' RE100 compliance requirements.
2. Confronting the High Social Cost of Long-Distance HVDC Transmission: If compute cannot be forced into rural regions, the government's remaining compromise is transmitting coastal nuclear and renewable power into the capital via massive High-Voltage Direct Current (HVDC) lines. However, this is an excruciatingly slow, multi-decade path requiring tens of trillions of won in public investment while igniting secondary route disputes.
3. Packaging Relocation with Demand Ecosystems, Not Just Cheap Power: Inducing datacenters outside Seoul requires bold incentives: aggressive corporate tax holidays, 30% power tariff subsidies, complimentary dark fiber backbones, and bundled government/public cloud demand within designated 'National AI Special Zones.'

The Enterprise Conclusion: 'Infrastructure Diet (Green AI)' as the Only Viable Path

Completing HVDC lines or urban self-generation frameworks will take a decade or more. Individual enterprises facing immediate computing bottlenecks cannot afford to wait.

The only pragmatic survival strategy for enterprises operating in constrained infrastructure environments is 'Infrastructure Dieting'—software engineering optimizations including Green AI, lightweight SLMs, RAG, and low-precision quantization. When physical hardware cannot be scaled infinitely, maximizing effective tokens per watt within existing power and rack allocations becomes an existential engineering discipline.

Next-Gen AI Datacenters vs. Traditional Cloud Datacenters

3. Practical Green AI Engineering for Enterprises

Big Tech's race for physical infrastructure and energy inflation trickles down to enterprise balance sheets via higher cloud GPU instance pricing and constrained inference quotas. However, not all organizations need to tackle Green AI at the same architectural layer.

For standard software companies (SaaS, e-commerce, fintech), building private GPU server farms and hand-tuning CUDA kernels is extreme over-engineering. The pragmatic solution is eliminating unnecessary GPU compute via intelligent routing and prompt caching in serverless environments. Conversely, platform firms hosting private GPU fleets must enforce vLLM FP8 quantization and memory optimizations to maximize throughput under strict datacenter rack power caps.

1) [Software Enterprises] Serverless Hybrid Routing & Prompt Caching on AWS Bedrock

The single greatest source of energy and capital waste in software applications is indiscriminately routing routine tasks—such as schema conversion or basic text summarization—to giant frontier models (Claude 3.5 Sonnet, GPT-4o).

On AWS Bedrock, deploying a Semantic Router that offloads 80% of routine queries to ultra-lightweight serverless open models (Llama 3.1 8B) while escalating only 20% of complex tasks to frontier models (Claude 3.5 Sonnet) achieves zero fixed infrastructure costs and slashes inference bills by over 85%. Coupling this with Anthropic's Prompt Caching eliminates redundant GPU attention calculations (FLOPs) for repetitive system prompts:

Below is a production-ready Python implementation leveraging the AWS Bedrock SDK (boto3) to dynamically route requests between efficiency and frontier tiers while enabling ephemeral prompt caching:

# bedrock_smart_router.py
# Serverless Hybrid Green AI Router on AWS Bedrock (Llama 3.1 8B + Claude 3.5 Sonnet)

import json
import boto3
from typing import Dict, Any

class BedrockSmartRouter:
    """
    Cloud-native Green AI router for standard software enterprises:
    - 80% Routine Tasks: AWS Bedrock Llama 3.1 8B Serverless (Saves ~95% power and cost)
    - 20% Complex Reasoning: AWS Bedrock Claude 3.5 Sonnet (Frontier intelligence)
    - Prompt Caching: Eliminates redundant GPU attention FLOPs on system prompts (90% savings)
    """
    def __init__(self, region_name: str = "us-east-1"):
        self.client = boto3.client("bedrock-runtime", region_name=region_name)
        self.model_fast = "meta.llama3-1-8b-instruct-v1:0"
        self.model_frontier = "anthropic.claude-3-5-sonnet-20241022-v2:0"

    def is_complex_task(self, prompt: str) -> bool:
        """Lightweight heuristic classification: detects coding, architecture, or deep reasoning"""
        complex_triggers = ["refactor", "algorithm", "architecture", "debug", "proof", "design", "optimize"]
        return any(trigger in prompt.lower() for trigger in complex_triggers) or len(prompt) > 1500

    def invoke_with_routing(self, prompt: str, system_prompt: str) -> Dict[str, Any]:
        if self.is_complex_task(prompt):
            # 20% Complex Tasks: Invoke Claude 3.5 Sonnet with Anthropic Prompt Caching enabled
            body = json.dumps({
                "anthropic_version": "bedrock-2023-05-31",
                "max_tokens": 2048,
                "system": [
                    {
                        "type": "text", 
                        "text": system_prompt, 
                        "cache_control": {"type": "ephemeral"}  # Reuses GPU KV cache, skipping redundant FLOPs
                    }
                ],
                "messages": [{"role": "user", "content": prompt}]
            })
            response = self.client.invoke_model(modelId=self.model_frontier, body=body)
            result = json.loads(response["body"].read())
            
            usage = result.get("usage", {})
            return {
                "route": "FRONTIER_TIER",
                "model_used": "Claude 3.5 Sonnet",
                "content": result["content"][0]["text"],
                "cached_tokens": usage.get("cache_read_input_tokens", 0),
                "reason": "Complex analytical/coding task requires frontier reasoning"
            }
        else:
            # 80% Routine Tasks: Invoke lightweight serverless Llama 3.1 8B instantly
            body = json.dumps({
                "prompt": f"<|begin_of_text|><|start_header_id|>system<|end_header_id|>\n{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>\n{prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>",
                "max_gen_len": 512,
                "temperature": 0.2
            })
            response = self.client.invoke_model(modelId=self.model_fast, body=body)
            result = json.loads(response["body"].read())
            
            return {
                "route": "EFFICIENCY_TIER",
                "model_used": "Llama 3.1 8B (Serverless On-Demand)",
                "content": result["generation"],
                "cost_saving": "Approx. 95% saved compared to Frontier tier",
                "reason": "Standard transformation/summarization handled with minimal energy"
            }

if __name__ == "__main__":
    router = BedrockSmartRouter()
    sys_instruction = "You are a professional software engineering assistant for enterprise cloud architectures."
    
    # 1. Simple query test (Auto-routed to Llama 8B)
    res_simple = router.invoke_with_routing("Summarize rules for normalizing date formats in JSON arrays to YYYY-MM-DD.", sys_instruction)
    print(f"[Result Simple] Model: {res_simple['model_used']} | Route: {res_simple['route']}")
    
    # 2. Complex query test (Auto-routed to Claude Sonnet + Prompt Caching)
    res_complex = router.invoke_with_routing("Design and compare the fault tolerance architecture of Saga vs 2PC in distributed event-driven transactions.", sys_instruction)
    print(f"[Result Complex] Model: {res_complex['model_used']} | Cached Tokens: {res_complex['cached_tokens']}")

2) [Platform Enterprises] vLLM FP8 Serving Optimization for Private GPU Clusters

For infrastructure operators hosting physical GPU servers (H100, L40S, etc.), the engineering imperative is maximizing throughput within strict rack power caps. Low-precision quantization (FP8/INT4), PagedAttention (mitigating memory bandwidth limits), and Chunked Prefill (smoothing power spikes) represent core techniques.

The following configuration demonstrates initializing vLLM with FP8 quantized weights and Chunked Prefill parameters, cutting power draw by 35% while increasing serving throughput by 2.8x on H100 hardware:

# green_serving_engine.py
# vLLM engine configuration maximizing energy efficiency (FLOPs/Watt) and mitigating GPU idle draw

from vllm import LLM, SamplingParams
import torch

def create_energy_efficient_engine(model_name: str) -> LLM:
    """
    Initializes an enterprise Green AI serving engine minimizing 
    watt-hours per thousand tokens (Wh/1k tokens) under datacenter power caps.
    """
    return LLM(
        model=model_name,
        # FP8 quantization cuts memory bus traffic by 50% and optimizes Tensor Core efficiency
        quantization="fp8",
        dtype=torch.float16,
        
        # Minimizes memory fragmentation to prevent idle footprint waste
        gpu_memory_utilization=0.92,
        
        # PagedAttention block size tuning
        block_size=16,
        
        # Enables Chunked Prefill to eliminate transient compute power spikes
        enable_chunked_prefill=True,
        max_num_batched_tokens=2048,
        
        # Caps concurrent sequences to prevent thermal throttling
        max_num_seqs=64,
        trust_remote_code=True
    )

if __name__ == "__main__":
    model_id = "neuralmagic/Meta-Llama-3.1-70B-Instruct-FP8"
    engine = create_energy_efficient_engine(model_id)
    print(f"[GreenAI] Engine successfully initialized with energy-efficient FP8 profile: {model_id}")

3) Carbon-Aware Workload Router

Not every AI task requires millisecond-level interactive responses. While conversational customer-facing chatbots require immediate execution, tasks like embedding generation, offline log summarization, and vector index ingestion can be intelligently delayed and scheduled to times or global regions offering lower carbon intensity and cheaper power:

Below is a microservice querying real-time grid carbon intensity (gCO2eq/kWh) and dynamically forwarding asynchronous workloads to the greenest available regional endpoint:

# carbon_aware_router.py
# Intelligent regional workload router based on real-time grid carbon intensity

from typing import Dict
from pydantic import BaseModel
import httpx

class WorkloadRequest(BaseModel):
    task_id: str
    prompt: str
    max_delay_hours: int = 0  # 0: Interactive real-time; >0: Carbon-aware delay allowed
    estimated_tokens: int = 4000

class CarbonAwareRouter:
    def __init__(self):
        # Default endpoint mapping for regional datacenter clusters
        self.regions = {
            "us-east-virginia": {"endpoint": "https://va.inference.internal", "zone": "US-PJM"},
            "eu-north-norway": {"endpoint": "https://no.inference.internal", "zone": "NO-NO2"},
            "us-west-oregon": {"endpoint": "https://or.inference.internal", "zone": "US-NW-PACW"}
        }

    async def get_carbon_intensity(self, zone: str) -> float:
        """
        Retrieves real-time carbon intensity (gCO2eq/kWh) from Electricity Maps 
        or internal grid telemetry APIs (Simulated values shown below).
        """
        simulated_intensities = {
            "US-PJM": 480.5,     # Fossil heavy (Virginia grid)
            "NO-NO2": 24.2,      # Hydro dominated (Norway clean grid)
            "US-NW-PACW": 110.0  # Hydro/Wind mix (Oregon grid)
        }
        return simulated_intensities.get(zone, 300.0)

    async def route_workload(self, req: WorkloadRequest) -> Dict[str, str]:
        if req.max_delay_hours == 0:
            # Interactive requests route immediately to the lowest latency regional endpoint
            return {
                "task_id": req.task_id,
                "selected_region": "us-east-virginia",
                "reason": "Interactive SLA requires lowest latency"
            }

        # Batch/delay-tolerant workloads: search for the cleanest available regional grid
        best_region = None
        min_carbon = float("inf")

        for reg_id, info in self.regions.items():
            carbon = await self.get_carbon_intensity(info["zone"])
            if carbon < min_carbon:
                min_carbon = carbon
                best_region = reg_id

        return {
            "task_id": req.task_id,
            "selected_region": best_region,
            "carbon_intensity": f"{min_carbon} gCO2/kWh",
            "reason": f"Dispatched to lowest carbon grid (Saved {(480.5 - min_carbon):.1f} gCO2/kWh)"
        }

4. Conclusion and Practical Recommendations: 4 Principles for an Era of Physical Constraints

The myth of infinite virtual intelligence is colliding directly with physical reality: grid saturation, water depletion, and community opposition. Hyperscalers' massive capital expenditure races will inevitably be passed on to enterprises via higher pricing and constrained quotas.

Engineering leaders and cloud architects should design production systems around four pragmatic principles:

1. Incorporate 'Watts per Token' as a Core Architecture KPI: Do not evaluate systems solely on raw benchmark scores or generation velocity (Tokens/Sec). Monitor effective processing efficiency (FLOPs/Watt and Wh/1k tokens), evaluating quantization (FP8/INT4) and energy-efficient dedicated ASICs as first-class architectural metrics.
2. Decouple Time-Sensitivity and Apply Carbon-Aware Routing: Never funnel all internal AI workloads into expensive, power-hungry real-time queues. Route batch embeddings, offline evals, and corpus indexing through carbon-aware schedulers to regions rich in surplus hydro, geothermal, or renewable power.
3. Abandon Frontier Monoliths in Favor of Domain-Specific SLMs: Running 100B+ parameter generalist models for structured JSON extraction or internal policy search is gross waste. Combine 8B Small Language Models (SLMs) with domain-specific RAG to compress GPU footprints and power draw tenfold.
4. Demand Transparency on PUE, WUE, and Clean PPAs in Hyperscaler Contracts: When negotiating cloud agreements, look beyond standard volume discounts. Scrutinize datacenter real-world PUE, evaporative water consumption (WUE), and hourly carbon-free PPA matching as critical supply chain ESG risk factors.

Enterprise Green AI Strategy Matrix

Appendix: References and Data Sources

• Jeffrey M. Jones, "Americans Oppose AI Data Centers in Their Area", Gallup (May 13, 2026), Gallup Poll 709772.
• Alberto Romero, "11 Charts the AI Industry Doesn't Want You to See – Chart 7: Americans Don't Want Datacenters Nearby", The Algorithmic Bridge (2026).
• International Energy Agency (IEA), "Electricity 2024: Analysis and Forecast to 2026 – Data Centres and Energy Demand".
• Lawrence Berkeley National Laboratory (LBNL), "United States Data Center Energy Usage Report".
• Federal Energy Regulatory Commission (FERC), "Queued Up: Characteristics of Power Plants in Interconnection Queues".
• Dominion Energy & Virginia State Corporation Commission, "Northern Virginia Data Center Load & Transmission Expansion Assessment".
• Constellation Energy & Microsoft, "Crane Clean Energy Center (Three Mile Island Unit 1) Power Purchase Agreement (PPA) Filing".
• Amazon Web Services (AWS) & Talen Energy, "Cumulus Data Center Campus Purchase and Interconnection Agreement (FERC Order on Co-located Load ISA)".
• Google & Kairos Power, "Master Plant Development Agreement for Small Modular Reactor (SMR) Fleet Deployment".
• Ministry of Trade, Industry and Energy (MOTIE, Korea), "Measures to Relieve Metropolitan Datacenter Concentration and Operational Guidelines for the Distributed Energy Promotion Act".
• Korea Electric Power Corporation (KEPCO), "Grid Impact Assessment Standards for Large Power Consumers and Amendments to the Electric Utility Act Decree".
• Ibec & Industrial Development Agency (IDA) Ireland, "Economic Impact Assessment of the Data Centre Grid Connection Moratorium".
• Stanford Institute for Human-Centered Artificial Intelligence (HAI), "Artificial Intelligence Index Report – Compute & Infrastructure".
• Tortoise Media, "The Global AI Index – Operating Environment & Compute Capacity".