The AI Bottleneck: Why Power Infrastructure Is Becoming the Scarce Asset that Cannot be Ignored

Written by Adhum Carter Wolde-Lule

Director at Prism Power Group

Adhum Carter Wolde-Lule is a Director at Prism Power Group, a UK provider of critical power infrastructure and modular data centre solutions. With more than ten years’ experience across international trade, development and investment, his background spans infrastructure development, real estate and venture investment across the energy and technology sectors. He focuses on resilient power systems and intelligent energy technologies capable of supporting the evolving demands of data centres and modern industry

The bottleneck has moved

For the first three years of the AI boom, the scarce resource was Nvidia GPUs. Understandably, they became the focal point for investors, operators and governments, because that was the source of the gridlock.

Access to the latest processors determined who could deploy the most capable AI infrastructure and, by extension, who could compete at the frontier.

That is no longer the case, as the queue has moved, and today it is held not by semiconductor manufacturers but by grid operators.

The assets determining the pace of AI deployment are increasingly transformers, switchgear, substations and the copper that connects them all.

This shift fundamentally changes where value is likely to accrue over the coming decade.

According to the International Energy Agency, global data centre electricity consumption is expected to roughly double by 2030 after growing by around 17 per cent last year alone.

Electricity networks were never designed to accommodate demand arriving in 100MW, 300MW or 500MW increments, yet this has become the new reality of hyperscale AI infrastructure.

Grid interconnection has now become the principal rationing mechanism for AI expansion, with projects increasingly constrained not only by the availability of compute, but by the ability to secure power.

The consequences are already becoming visible.

A hyperscale campus joining the interconnection queue in Northern Virginia today could realistically wait until the early years of the next decade before receiving permanent utility power.

In many cases, the buildings themselves will be completed years before the electricity required to operate them becomes available.

Capital is therefore being deployed into assets that may remain commercially dormant while operators wait for connections that are entirely outside their control.

Power, equipment and materials

It is perhaps unsurprising, therefore, that the world's largest AI companies are beginning to behave more like integrated utilities than technology businesses.

One example that illustrates this changing landscape is Elon Musk's reported acquisition of APR Energy, a supplier of trailer-mounted gas turbines with more than a gigawatt of deployable generation capacity, in a transaction valued at approximately $1 billion.

The acquisition only came to light because of a regulatory filing rather than a formal announcement, and no detailed explanation has been provided as to its intended purpose.

Nevertheless, the strategic rationale is difficult to ignore. Owning deployable generation capacity removes dependence on utility connection queues and gives operators far greater certainty over deployment schedules.

I would expect many of the world's largest AI developers to pursue similar strategies.

Faced with a choice between waiting five to seven years for a permanent grid connection or deploying on-site generation within eighteen months, the commercial calculation becomes increasingly straightforward.

In today's AI infrastructure market, time has become every bit as valuable as power itself, because delays in energising facilities represent lost revenue, deferred returns and diminished competitive advantage.

Even where network capacity can ultimately be secured, the supporting electrical equipment presents another significant constraint.

Lead times for large power transformers in the United States have extended from roughly twelve months to more than double the time at well over 130 weeks, while switchgear manufacturers have order books stretching several years into the future.

This has created an unusual inversion of the traditional construction process.

Concrete, steel and structural works are rarely the limiting factors; projects can be permitted, financed and physically constructed before entering an extended period of inactivity while they await the electrical infrastructure required to bring them into operation.

It is no coincidence that a significant proportion of US data centres scheduled for completion over the coming years now face delays, with shortages of critical electrical equipment representing a major contributing factor.

Beneath these challenges sits an equally important structural constraint that receives considerably less attention than semiconductor supply: copper.

AI data centres require substantially greater electrical density than previous generations of facilities, consuming as much as three times more copper than conventional server farms.

Forecasts suggest global demand could more than double by 2040 yet bringing a new copper mine into production routinely requires a decade or more.

Unlike software or compute, there is no rapid scaling mechanism available for these underlying materials, nor is there a viable substitute capable of supporting deployment at comparable scale.

Policy will shape the market

Alongside physical infrastructure, policy remains an important variable.

Some jurisdictions have responded to rising electricity demand by imposing development moratoriums, while others are actively encouraging investment through planning reform and infrastructure support.

As a result, the geography of AI infrastructure over the remainder of this decade is likely to be shaped as much by utility commissions, planning authorities and energy regulators as it is by cloud adoption or model development.

Regulators are also asking a legitimate question about who should ultimately fund the substantial grid upgrades required to accommodate this new wave of demand, and the industry has yet to settle on a universally accepted answer.

What this means for investors

For investors, however, the implications are becoming increasingly clear.

Power infrastructure has evolved from a supporting asset into a strategic one, and in many respects, it is now scarcer than compute itself.

GPUs can be manufactured over the course of months as fabrication capacity expands.

By contrast, delivering substations, transmission upgrades and energised sites requires multi-year programmes constrained by planning, manufacturing capacity and skilled engineering resources.

A widening competitive gap is therefore emerging between operators that secured energy infrastructure early and those now entering increasingly congested queues.

The AI economy will undoubtedly continue to reward innovation in software and semiconductors, but the next phase of value creation is likely to be determined by the infrastructure that enables those technologies to operate.

Investors would therefore do well to look beyond the companies producing the chips and pay closer attention to those delivering resilient power systems, critical electrical equipment and deployable digital infrastructure.

The headlines may continue to focus on compute, but the investment opportunity increasingly lies in the assets that determine whether compute can be switched on at all.

Exploring opportunities in the private markets

ICP connects verified investors with selected early-stage companies through a secure, gated environment. Investors can review founder-supplied information and request introductions where they wish to continue the conversation, with ICP facilitating access and communication on a factual, non-advisory basis.
Learn more about joining the ICP investor network: