Navigating the Energy Crisis in AI Infrastructure

Reading Time: 7 minutes

Prefer watching instead of reading? Watch the video here. Prefer reading instead? Scroll down for the full text. Prefer listening instead? Scroll up for the audio player.

P.S. The video and audio are in sync, so you can switch between them or control playback as needed. Enjoy Greyhound Standpoint insights in the format that suits you best. Join the conversation on social media using #GreyhoundStandpoint.

---

A new KAIST roadmap reveals HBM8-powered GPUs could consume more than 15kW per module by 2035, pushing current infrastructure, cooling systems, and power grids to breaking point.

 “The power requirements outlined in the KAIST roadmap signal not just a thermal or architectural challenge, but an impending crisis of coordination between compute timelines and utility readiness,” said Sanchit Vir Gogia, CEO at Greyhound Research. “These electrical densities simply cannot be supported by existing grid infrastructure in most regions.”

He added that while hyperscalers are reserving gigawatt-class electricity allotments up to a decade in advance, regional utilities struggle to upgrade transmission, often requiring 7 to 15 years for execution. “Speed-to-power is now eclipsing speed-to-market as the defining metric of digital competitiveness.”

As AI modules push infrastructure to its limits, electricity is becoming a critical driver of return on investment. “Electricity has shifted from a line item in operational overhead to the defining factor in AI project feasibility,” Gogia noted. “Electricity costs now constitute between 40–60% of total Opex in modern AI infrastructure, both cloud and on-prem.”

As quoted in NetworkWorld.com, in an article authored by Gyana Swain published on June 17, 2025.

Beyond the Media Quote: Our View, In Full

Transmission Infrastructure: A 15,000-Watt AI Module Challenges the Grid

Greyhound Standpoint — According to Greyhound Research, the power requirements outlined in the KAIST TeraLab HBM roadmap—culminating in 15,360-watt GPU-HBM modules by 2035—signal not just a thermal or architectural challenge, but an impending crisis of coordination between compute timelines and utility readiness. These power demands are not hypothetical: modular AI accelerators already exceed 1,000W each today, and projections from both academia and industry point to >15kW per module within a decade as chiplet-based architectures scale and interposer-level integration intensifies. At the rack level, this pushes deployments well into megawatt-class territory for a single training cluster.

Such electrical densities simply cannot be supported by existing grid infrastructure in most regions. Transmission upgrades—whether high-voltage lines, substations, or distribution capacity—routinely require 7 to 15 years to execute, even under optimal conditions. Regulatory reviews, environmental clearances, community hearings, and capital constraints frequently push projects beyond their intended timelines. In Northern Virginia, the world’s largest hyperscale cluster, planned data centres are already outpacing available power, with the region’s utility scrambling to bring online additional capacity only by 2026. In California, PG&E warns that adding a substation for new AI facilities can take over five years, even without litigation or local pushback.

Critically, the notion of a 5–10 year grid upgrade cycle—often cited in policy briefs—is not just optimistic, it is dangerously misaligned with the current hardware roadmap. The reality is that hyperscalers are now reserving gigawatt-class electricity allotments up to a decade in advance and lobbying for faster permitting to avert compute bottlenecks. The U.S. Department of Energy and think tanks like CSIS have explicitly framed this as a strategic capacity race: “speed-to-power” is now eclipsing “speed-to-market” as the defining metric of digital competitiveness.

Absent urgent and coordinated intervention, we expect to see an increasing number of regional and national AI projects either delayed, restructured, or entirely relocated—not because of chip shortages or developer bandwidth, but due to power grid inadequacy. For enterprises and governments alike, this creates a new axis of planning: infrastructure foresight must now move in lockstep with compute ambition. In effect, electrical capacity has become the hard ceiling of AI progress. Without aligning policy, regulation, and utility planning to match the hardware roadmap, national AI deployments may be throttled at source.

Strategic Planning: When Power Costs Tip AI ROI Models

Greyhound Standpoint — According to Greyhound Research, electricity has shifted from a line item in operational overhead to the defining factor in AI project feasibility. For modern AI workloads—particularly those involving large language models, multi-agent simulations, and continuous fine-tuning—energy is not just a byproduct of scale; it is the economic substrate on which all model viability rests. Based on internal benchmarking and third-party estimates, electricity costs now constitute between 40–60% of total Opex in modern AI infrastructure, both cloud and on-prem. That figure, until recently abstracted into bundled platform pricing, is now being surfaced in board-level discussions.

A single high-performance AI server running at 1.5kW for fine-tuning tasks can incur over $0.87/hour in electricity costs at $0.15/kWh—and that does not account for redundancy, cooling, or backup infrastructure. With next-generation modules pushing towards 15kW, the economics quickly become exponential. Enterprises now face the challenge of modelling these costs not just per workload, but across daily usage variability, peak tariff zones, and future-proofing energy procurement contracts.

This dynamic is dramatically reshaping the calculus of cloud versus on-premises deployments. While in-house data centres may offer control and compliance benefits, they are increasingly exposed to price volatility, inefficient power usage effectiveness (PUE), and regulatory carbon scrutiny. In contrast, hyperscalers like AWS, Google Cloud, and Azure operate at sub-1.2 PUE levels, leverage preferential bulk electricity rates, and co-locate near renewable energy zones. They also engage in long-term power purchase agreements (PPAs) to cap cost exposure—an advantage that is not easily replicable by enterprise IT departments.

This delta in energy efficiency and pricing is now large enough to tip entire ROI models, particularly for inferencing-heavy or bursty AI use cases. Importantly, this isn’t just a CFO concern. Sustainability mandates, particularly in Europe and parts of Asia, are beginning to tie ESG scoring to the carbon intensity of compute, not just its financial cost. That means boards must now ask not only, “What is the cost per model run?” but also, “Where is the energy coming from, and how does it reflect on our emissions disclosures?”

Moreover, modern AI workloads are highly spiky and unpredictable. Transformer-based models in production can show huge variability in power draw across training epochs and inference batches. This volatility creates challenges for capacity provisioning, but also for forecasting energy exposure—something hyperscalers mitigate through load balancing across regional power markets. Enterprises with localised infrastructure are often exposed to worst-case pricing during peak grid load.

In effect, electricity is now the unseen negotiator at every AI strategy table. ROI cannot be divorced from wattage—and more importantly, from the provenance of those watts. Enterprises must now develop power-aware AI investment strategies, treating kilowatt-hours as a first-class design variable alongside algorithmic efficiency and model accuracy. Those who fail to do so risk building unsustainable architectures—financially, operationally, and reputationally.

AI Deserts: Power Scarcity May Trigger Uneven AI Development

Greyhound Standpoint — According to Greyhound Research, the geography of AI is being redrawn not by talent, data, or innovation policy—but by the availability of electricity. What began as isolated anecdotes of delayed data centre builds has now hardened into a structural global pattern: regions with constrained grid capacity are losing out on the next wave of AI infrastructure, while those with foresight and surplus energy are rapidly consolidating dominance. This is creating a new global fault line—between “AI-rich” and “AI-poor” geographies.

Real-world data supports this emerging bifurcation. Singapore, one of Asia’s most advanced digital economies, has virtually no spare power capacity for data centres—less than 7 MW remains, and government-imposed moratoriums have slowed new permits for years. Amsterdam and the broader Netherlands have placed hard caps on data centre growth due to grid overload, including a ban on hyperscale builds over 70 MW. In Mexico’s Querétaro region, a rising tech cluster, AI projects have been delayed by up to 18 months because promised transmission upgrades failed to materialise—leaving only 0.6 MW available in one of the most active data centre zones.

In stark contrast, regions like Norway’s Oslo corridor and Brazil’s São Paulo have surged in hyperscaler interest, largely due to stable grid capacity, abundant renewable energy, and favourable regulation. Even West Virginia in the U.S.—historically overlooked in digital investment—has passed legislation to expedite AI-focused data centres with access to legacy coal power and faster permitting cycles.

These decisions have economic consequences that will last a decade or more. Hyperscalers are not merely siting compute where fibre exists—they are chasing power. Regions without sufficient electricity will not only miss out on data centre capex and job creation, but also the downstream ecosystem benefits: AI research hubs, edge inferencing zones, start-up clusters, and sovereign model development. What emerges is the concept of the “AI desert”: a region starved not of talent or funding, but of megawatts.

The risk is not just economic—it is strategic. Regions unable to host sovereign AI models due to electrical constraints will become increasingly dependent on external cloud regions, many of them controlled by foreign entities. This introduces data sovereignty concerns, latency limitations, and geopolitical exposure. Energy policy, once seen as separate from innovation, is now a direct lever of AI autonomy.

Furthermore, energy-constrained markets tend to exhibit higher lease rates—Singapore now commands 2–3x the cost per kilowatt-hour compared to power-abundant U.S. zones like Chicago. This creates a second-order barrier: even where physical capacity exists, smaller players and local innovators may be priced out entirely.

Ultimately, this divide will deepen unless national and municipal authorities treat power provisioning as part of their digital competitiveness strategy. That means not just adding capacity, but reconfiguring grid architecture, streamlining substation construction, and rebalancing power to edge zones. In parallel, AI architects must incorporate geographic energy mapping into workload siting and model deployment. The outcome will be uneven: some cities will become AI capitals; others, AI deserts. And unlike data, power cannot be cloned or compressed—it must be generated, routed, and consumed. This is the new constraint frontier.

Copyright Policy. All content contained on the Greyhound Research website is protected by copyright law and may not be reproduced, distributed, transmitted, displayed, published, or broadcast without the prior written permission of Greyhound Research or, in the case of third-party materials, the prior written consent of the copyright owner of that content. You may not alter, delete, obscure, or conceal any trademark, copyright, or other notice appearing in any Greyhound Research content. We request our readers not to copy Greyhound Research content and not republish or redistribute them (in whole or partially) via emails or republishing them in any media, including websites, newsletters, or intranets. We understand that you may want to share this content with others, so we’ve added tools under each content piece that allow you to share the content. If you have any questions, please get in touch with our Community Relations Team at connect@thofgr.com.