Why Energy Infrastructure Is the Smarter AI Investment Play

#

The obvious beneficiaries of the AI boom are the companies building the models. The less obvious beneficiaries are the ones keeping the lights on.

That gap between where investor attention concentrates and where durable returns may actually live is the tension worth examining. Electricity consumption by data centres grew 17% in 2025, roughly six times the pace of total global electricity demand growth, and the International Energy Agency (IEA) projects that figure to roughly double again by 2030. The AI layer driving this growth is also the most competitively volatile layer: technology platform markets tend to consolidate into two or three dominant players within a few years, leaving the losers with little to show investors. The physical infrastructure those competitors all depended on, however, keeps earning.

Here is what the structural demand signal actually tells a commercially minded investor: where the durable position sits, what threatens it, and what the current macroeconomic backdrop means for entry timing.

The electricity numbers that make the investment thesis concrete

Start with the baseline. The IEA estimates data centres consumed roughly 415 TWh of electricity in 2024, about 1.5% of global power demand. That figure alone is unremarkable. What matters is the trajectory it sits on.

In 2025, data-centre electricity demand grew 17%, lifting consumption to approximately 485 TWh, according to updated IEA figures. Compare that to the roughly 3% growth in total global electricity demand over the same period, and the divergence becomes the point: this is a demand source pulling sharply away from the broader power system.

Layer in the AI-specific driver. A April 2025 analysis published in Nature reports that AI workloads accounted for about 15% of total data-centre energy use and 24% of server electricity demand in 2024. The IEA expects AI-focused facilities to triple their consumption by 2030, making AI the principal engine behind the entire category’s growth.

Where the projections land by 2030

Even the conservative floor, the IEA’s 945 TWh, implies a near-doubling from current levels. That is the crux of the read for an investor: the disagreement between institutions is about how large the increase is, not whether it happens.

The spread between institutional projections, from the IEA-4E’s 200-400 TWh conservative floor to the IMF’s 1,500 TWh high scenario, reflects genuine methodological disagreements that any investor should understand before anchoring to a single figure; the AI energy impact numbers are more contested than the consensus framing suggests.

The generation requirement makes the physical stakes explicit. The IEA projects electricity generation specifically to supply data centres rising from 460 TWh in 2024 to over 1,000 TWh by 2030, and approximately 1,300 TWh by 2035. Every scenario on this range requires substantial new generation and transmission capacity to be built. That is precisely where regulated returns are earned, and it is the reason the range matters more to your positioning than any single headline number.

Why utilities earn the returns AI software companies cannot guarantee

Two investors can hold the same view on AI electricity demand and end up with radically different outcomes, depending on which layer of the stack they own. The demand dynamic is shared. The financial profile is not.

Consider the regulated utility model first. When a utility spends capital on lines, substations, and generation, that spending rolls into a rate base that earns an allowed return over multi-decade horizons. The earnings are structurally predictable in a way AI software margins are not, because the return is set by regulators against deployed assets rather than won in a competitive market.

Now apply the AI oligopoly thesis to sharpen the contrast. If AI platform markets consolidate into two or three dominant players within three to four years, consistent with how operating systems, search, cloud, and mobile all resolved, then most of today’s AI application-layer investors are backing companies that will not make the final set. The electricity those companies all consumed keeps earning regardless of who wins.

Sean Roosen of the Osisko Gold Group frames the logic directly: energy companies stand to benefit from AI adoption regardless of which AI firms ultimately prevail. That is the structural distinction between the two positions.

The capital scale confirms the build is already underway. According to Deloitte’s June 2025 analysis, capex across US electric and gas utilities is set to jump 22% year-on-year to approximately US$212 billion in 2025. Over the same period, eight major hyperscalers expect a 44% increase in AI-related capex to roughly US$371 billion. Both flows ultimately feed electricity infrastructure.

The two layers carry entirely different investment characteristics:

Scarcity adds a further edge. Allianz Research emphasises the long lead times and permitting constraints on new generation and transmission, which means existing, strategically located infrastructure carries scarcity value as AI demand scales faster than new capacity can be permitted and built.

Transmission infrastructure bottlenecks are the specific constraint that turns a decade-long capex commitment into a decade-long earnings lag; interconnection queues in the US have grown to exceed five years in some regions, meaning utilities that have committed capital to AI-serving generation projects may not begin earning regulated returns until well into the 2030-2032 window.

Understanding this mechanism is what separates a thesis-driven energy position from a momentum trade. The rate-base model converts AI’s non-discretionary electricity demand into a multi-decade earnings stream. Rather than betting on which AI firm wins the model race, you are taking a structural position on the energy floor that every AI firm requires.

What “picks and shovels” means in practice for resource and materials exposure

The electricity thesis does not stop at the meter. It extends further down the physical chain than many investors initially assume, into the metals and materials that grid expansion consumes.

Transmission and distribution (T&D) spending is materials-intensive by nature. S&P Global Market Intelligence reports combined US utility T&D capex of US$84.9 billion in 2025, up more than US$12 billion on the prior year. That capital buys conductors, transformers, and towers, which translates directly into demand for copper, aluminium, and steel.

The chain runs as follows:

1. AI compute demand requires firm, expanded electricity supply.
2. That supply requires new generation, lifting fuel and materials demand on the generation side.
3. Generation must reach data centres, requiring grid expansion.
4. Grid expansion consumes copper and aluminium for conductors, steel for towers, and specialised components for transformers.

AI infrastructure metals demand extends well beyond the copper conductors in transmission lines; specialised transformer components, aluminium busbars, and the steel lattice structures of high-voltage towers each represent distinct procurement flows that scale with every new generation and distribution circuit brought online to serve data centres.

Each link is a demand source with a multi-year duration. The IEA’s projection of generation for data centres crossing 1,000 TWh by 2030, up from 460 TWh in 2024, implies sustained materials demand across the entire period, not a single procurement spike.

Macro conditions shaping the entry context in late 2026

The AI demand story does not sit in isolation. It sits inside a broader resource cycle where macroeconomic conditions independently shape whether that demand floor translates into commodity price support in the near term.

US Treasury bond yields have risen, increasing the cost of issuing new debt. According to Sean Roosen of the Osisko Gold Group, this creates headwinds for sovereign borrowers while reinforcing the appeal of hard assets and resource-based investments in a rising-rate environment.

The dollar dynamic is more nuanced. Roosen notes that confidence in the US dollar is currently restraining gold’s upward move in dollar terms, yet gold has appreciated significantly when measured against weaker currencies. Some regions, particularly in the Middle East, are reportedly reducing reliance on the US dollar in trade.

That distinction matters for your read on gold: dollar strength is the near-term constraint on the USD price, not a fundamental argument against the hard-asset case. The IMF’s framing that AI-related electricity demand could outpace even electric vehicle demand reinforces a multi-year call on energy and resource supply, sitting alongside broader electrification as an independent tailwind.

The IMF electricity demand comparison finds AI-driven global power needs could reach around 1,500 TWh by 2030, a figure roughly 1.5 times higher than projected electric vehicle demand over the same period, reinforcing the multi-year resource and energy supply call that sits alongside broader electrification as an independent tailwind.

For an investor deciding where to position within the resource sector, the AI electricity thesis provides the demand floor. The macro layer, yields and dollar dynamics and currency weakness outside the US, determines whether that floor converts into commodity price support in the months ahead rather than the years.

Where the thesis can break down: risks a commercially minded investor must price in

A serious position deserves an honest stress test. The risks below do not invalidate the thesis. They define what to monitor, and pricing them in is what leaves you calibrated rather than exposed.

The three primary risks are:

Efficiency is the risk to take most seriously. A May 2025 IEA-4E review emphasises that improvements in server efficiency and workload management can materially reduce energy per unit of AI compute, and the Nature analysis notes significant room for gains in specialised accelerators and liquid cooling.

Here is the interpretive point. The efficiency scenario does not eliminate the electricity demand story; it compresses its magnitude. An investor who has priced in efficiency headwinds is positioned to hold through that compression rather than sell at the wrong moment.

Permitting delays and grid congestion compound the timing question. S&P Global notes grid congestion is already a key issue, and stricter interconnection rules or regional constraints could slow how quickly utilities deploy capital and begin earning returns. For a commercially oriented reader, these are the specific conditions under which the thesis underperforms, and knowing them is the informational edge over an investor who adopted the position uncalibrated.

Regulated transmission returns are not unconditional, and a coordinated challenge from state governments to FERC’s return-on-equity incentive adders is already compressing the margin assumptions embedded in some utility rate bases, adding a layer of regulatory risk that a structurally positioned investor must price in alongside permitting delays.

Where the signal is clear and where the judgment call begins

Separate what is structurally certain from what requires judgment, and the position becomes actionable.

What is certain: electricity and grid infrastructure are non-discretionary inputs for AI at any scale. Regulated utilities earn returns on that capex regardless of which AI firms win, with US utility capex at US$212 billion in 2025 confirming the build is already underway. The demand floor is material across every credible scenario, with even the IEA Base Case requiring over 1,000 TWh of new generation by 2030.

What requires judgment: which utilities and resource producers, at what entry point, with what tolerance for permitting delays and commodity cycles. That is where your own assessment does the work.

The distinction between the two AI positions holds at the close. Investors in AI application-layer companies are making a competitive selection bet on who survives consolidation. Investors in energy infrastructure are making a structural demand bet on what every survivor requires. Both can sit in a portfolio, but they are different risk types, and Roosen’s point stands as the anchor: energy companies benefit regardless of which AI firms prevail.

Three variables will determine whether the structural thesis converts into near-term performance:

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions. Past performance does not guarantee future results, and financial projections are subject to market conditions and various risk factors. These statements are speculative and subject to change based on market developments.