A generated megawatt is not necessarily a usable megawatt. Power has to be available in the right location and connected, transformed, distributed and controlled before it can support a data centre. The IEA estimates that new transmission lines can take four to eight years to build in advanced economies, while procurement now takes two to three years for cables and up to four years for large power transformers, with average lead times for both having almost doubled since 2021.

Our central investment view is that ‘speed to power’ is becoming one of the scarcest inputs in the AI infrastructure build-out. For investors this broadens the AI value pool to transmission capacity, grid interconnection, substations, transformers, switchgear, electrical engineering, contracting, energy-storage and distributed-generation businesses required to support AI compute. We expect a growing share of AI infrastructure spending to flow towards the companies that shorten the route from available electricity to powered data-centre capacity.

In this note is an analysis of how share prices, forward EPS expectations and P/E multiples have moved across our power bottleneck universe (Exhibit 6). Over the last 12 months, 19 companies out of 44 in our universe have de-rated, seven have re-rated and 11 have remained in a +/-10% P/E range. The median forward P/E for those that have de-rated is at a c 10% discount to the MSCI World forward P/E of 18.6x. Their 36% median increase in EPS has more than offset their 25% median share price decline. Such de-ratings may be creating more selective opportunities.

The AI investment proposition continues to evolve

Like AI itself, the AI investment proposition is constantly evolving. At Edison’s Growth Conference in May 2026, we heard the case for AI picks and shovels. We now develop this concept further through a series of thematic reports. While recent AI- related newsflow has been more balanced with the well-trodden positive case for AI pitched against the potential impact of slower model development and higher power prices, we continue to see potential across the value chain. Eaton pointed to $3.3tn of megaprojects in the US at Q126 with $54bn of project starts in the quarter, the third highest since its tracking began in 2021. At Q425 it noted only 16% of projects had started. This AI ‘picks and shovels’ thematic looks at the power bottleneck.

AI capex has become an energy-scale investment cycle

The scale of the AI build-out is now comparable with some of the world’s largest established capital-investment markets. The International Energy Agency (IEA) estimates that capex by the largest technology companies exceeded $400bn in 2025 and is expected to rise by a further 75% in 2026. It also notes that capex by just five technology companies is now larger than global investment in oil and natural-gas production.

This helps explain why power availability has moved from a supporting consideration to a potential constraint on AI deployment. The IEA describes a widening ‘scramble for electricity, grid connections, manufacturing capacity, chips and capital’, as the speed of the AI build-out increasingly runs ahead of the physical systems needed to support it.

For investors, the implication is that AI capex is increasingly becoming an infrastructure and energy-capex cycle as well as a technology-spending cycle.

Edison insight

The AI power trade is more specific than the simple idea that ‘AI needs more energy’.

In reality, the scarce resource is increasingly becoming timely access to sufficient power capacity, a much more complex constraint than headline energy generation capacity alone might suggest.

That requires not only energy generation, but also sufficient grid capacity, substations, transformers, switchgear, engineering capability, storage systems and, increasingly, more distributed and flexible power solutions. In addition, the scale of orders being reported by the largest electrical-equipment manufacturers suggests that AI-infrastructure investment is already flowing through this broader and more complex supply chain rather than remaining confined to data-centre developers alone.

This interpretation is also consistent with the IEA’s analysis, which finds limited evidence of a broad-based valuation uplift across the energy sector but identifies stronger links between AI and the valuations of electrical-equipment manufacturers, gas-turbine suppliers, selected nuclear companies and certain energy start-ups.

Our view is that this broadening of the value chain creates the potential for a second wave of AI ‘picks-and-shovels’ beneficiaries, particularly among companies that would traditionally have been analysed as industrial, utility, engineering or energy-transition businesses rather than as direct AI exposures.

The analytical challenge for investors is therefore to distinguish carefully between three different categories of exposure: direct beneficiaries, where data-centre demand is already a visible and material driver of projects, orders or electricity load; enablers, where AI is acting as an additional tailwind to an already attractive structural-growth market in areas such as grid investment, storage or distributed energy; and optionality exposures, where AI may improve the long-term strategic case but has not yet translated into a meaningful contribution to current earnings or cash flow.

From compute scarcity to power scarcity

The starting point remains exceptional spending on compute. The research underpinning this series identifies power, computing power and data as interdependent inputs into the AI infrastructure build-out, with the physical infrastructure surrounding the processor becoming increasingly important as deployments scale.

What has changed is the intensity and scale of the electricity requirement. The IEA estimates that the power density of AI servers increased 11-fold between 2020 and 2025 and could increase a further fourfold by 2027. Electricity consumption from AI-focused data centres consequently grew substantially faster than the wider data-centre market in 2025.

Efficiency improvements do not necessarily remove the constraint. More efficient chips and models can lower compute requirements per task, but inference and reasoning workloads remain power intensive, while lower costs can stimulate greater usage. Deloitte expects compute demand to continue rising as inference scales, while Goldman Sachs Research forecasts AI-token consumption increasing substantially through 2030 as agentic AI adoption develops. The relevant question for infrastructure providers is whether total deployed compute and utilisation continue to rise, rather than whether individual calculations become more efficient.

The result is an industry facing two related capacity questions: can enough compute be manufactured and can that compute be energised quickly enough?

The latter increasingly depends on infrastructure that was not designed for the speed, scale or concentration of today’s data-centre load growth.

A megawatt is more than generation

Generation is only one part of the constraint. Incremental capacity must still pass through transmission, interconnection, transformers and switchgear before it can become usable data-centre power. The US Department of Energy (DOE) identifies several of these equipment categories as already experiencing supply-chain pressure, including transformers, circuit breakers, substation equipment and power electronics. Distribution-transformer lead times increased from roughly three to six months in 2019 to one to two years or longer by 2024, while demand for distribution transformers increased 41% from 2019.

Grid expansion presents an additional timing mismatch. The DOE’s 2026 Draft National Transmission Needs Study identifies accelerating electricity demand from data centres, manufacturing and other large loads as one of the drivers of increased transmission requirements.

Nuclear: Longer-duration optionality on the AI power theme

Nuclear (see Edison explains: The nuclear resurgence) has become more strategically relevant as data-centre operators look for large volumes of reliable, low-carbon electricity. The IEA expects nuclear to make a meaningful contribution to meeting incremental data-centre power demand over the next decade, with the first small modular reactors (SMRs) expected to come online around 2030.

The opportunity spans existing nuclear generation, life extensions and restarts, advanced reactors and SMRs, and the nuclear fuel cycle. However, the timing differs materially from transformers, switchgear and grid engineering. Existing plants can potentially participate sooner through power contracts or uprates, whereas advanced reactors remain dependent on licensing, financing, supply-chain build-out and fuel availability.

For investors, we therefore view advanced nuclear primarily as long-duration optionality on the AI-power theme rather than a major source of earnings from the current data-centre capex cycle. If AI-related electricity demand remains structurally high and grid constraints persist, nuclear could become a more important part of the solution set over time; however, the near-term beneficiaries are more likely to be the companies supplying the equipment and infrastructure needed to connect and distribute power today.

From forecast demand to equipment orders

The power bottleneck is no longer just a forward-looking story about rising electricity demand; it is increasingly visible in the order books and capacity plans of the electrical-infrastructure industry itself.

One of the clearest signals comes from the large, diversified equipment manufacturers, whose scale provides a more reliable read-through on underlying demand than any single niche supplier. GE Vernova, for example, reported that its Electrification segment booked $2.4bn of data-centre-related orders in Q126 alone, already exceeding the total for the whole of 2025. By the end of Q226, year-to-date orders linked to data centres had risen to over $5bn, more than double the prior-year total.

A similar pattern is emerging at Siemens Energy. In Q326, its Grid Technologies division saw strong order growth, with a notable contribution from transformer demand tied to data-centre projects. The division’s order backlog now stands at €51bn, with Europe and North America driving much of the momentum.

Importantly, this is not just a story of demand signals but of supply beginning to respond. Eaton, for instance, announced a new US facility in April 2026 dedicated to medium-voltage switchgear, explicitly aimed at meeting rising demand from AI data centres alongside utilities and industrial customers. Hitachi Energy is also committing more than $2bn across its North American supply chain, expanding large power transformer and high-voltage equipment capacity, with part of this investment directly linked to AI-driven computing growth and broader electrification trends.

The same shift is increasingly evident in Europe. Schneider Electric is supporting SoftBank’s proposed French AI infrastructure programme by supplying electrical systems and prefabricated power modules, while also planning additional local manufacturing capacity to accelerate deployment timelines.

Taken together, these developments point to three important conclusions. First, AI-driven power demand is already translating into tangible equipment orders rather than remaining a purely theoretical forecast. Second, the response is capital intensive, with manufacturers actively expanding factories and supply chains in an effort to reduce lead times. Third, the bottleneck is increasingly shifting from electricity generation itself to the infrastructure required to connect, transform and distribute that power to data-centre sites.

In this context, traditional indicators such as order intake, backlog growth, book-to-bill ratios, lead times and announced capacity expansions become particularly valuable for assessing whether the AI-power theme is continuing to strengthen or beginning to normalise.

Storage: Reliability as well as generation

The power requirement is not simply a question of annual electricity consumption. AI training and model use can create large and rapid changes in load, increasing the importance of reliability and power quality.

The IEA estimates that 20–25GW of battery storage could be installed at data centres globally by 2030, potentially enabling facilities to support grid operations where incentives are appropriate.

This creates two potential investment channels. Utility-scale storage can help electricity systems absorb additional large loads and energy generation, while onsite storage can support data-centre reliability and load management.

Metlen Energy & Metals is a good example of a broader power-system beneficiary rather than a direct data-centre exposure. Through M Renewables, it develops and delivers utility-scale battery-storage projects, including standalone battery energy storage systems (BESS) and hybrid renewable-storage systems, with storage positioned as a tool for grid flexibility and stability. Rising data-centre electricity demand could therefore increase the need for the type of storage and balancing infrastructure in which Metlen participates, without necessarily translating into material direct data-centre revenue. We therefore distinguish between direct AI infrastructure beneficiaries and broader power-system beneficiaries in the company table.

US first, but not US only

The US currently offers the clearest evidence of exposure to this theme because the scale of new data-centre development has collided with grid infrastructure and equipment constraints simultaneously.

The constraint is increasingly relevant in Europe. In the UK, the government’s AI Growth Zones programme identifies timely grid connections as the biggest blocker to new AI infrastructure and includes measures intended to accelerate grid access and planning. Data centres have also been designated critical national infrastructure.

The European build-out is beginning to translate into electrical-infrastructure investment. Schneider Electric’s involvement in SoftBank’s proposed French AI infrastructure programme includes electrical infrastructure, prefabricated power modules and additional local manufacturing capacity.

The investment routes may nevertheless differ by geography.

In the US, the response has included utility investment, hyperscaler-power procurement, private-financing structures and increasing interest in dedicated onsite generation.

In the UK and Europe, the opportunity may be more closely linked to regulated grid investment, electricity-market reform, renewable generation, interconnection and policy-led attempts to shorten connection timelines.

The common denominator is that AI infrastructure adds a new source of large, concentrated and relatively time-sensitive electricity demand to power systems already undergoing substantial investment.

What could weaken the thesis?

AI infrastructure demand disappoints

One of the most important risks to the investment case is that AI development and deployment slows from current expectations, reducing the amount of data-centre capacity ultimately required. This could reflect slower model progress, more efficient compute usage, weaker-than-expected enterprise adoption or a normalisation in hyperscaler spending. In those scenarios, some of the capacity currently being announced may be delayed, cancelled or prove unnecessary. We are already seeing signs that not all planned projects will be delivered on time, and a moderation in AI-infrastructure spending would reduce pressure on electricity systems and weaken the incremental contribution of AI to demand for grid connections, transformers, switchgear and other power equipment.

Supply catches up

At the same time, it is important to recognise that high prices and extended order backlogs tend to attract capital and capacity expansion, and we are already seeing this response from major industrial players such as Eaton and Hitachi Energy, which are actively investing to increase production. If manufacturing capacity begins to grow faster than underlying demand, lead times and pricing power could start to normalise even in a world where total electricity consumption continues to rise, which means investors need to be careful not to confuse a period of temporary scarcity economics with something more durable and structural in nature.

AI is only one driver

AI is an incremental accelerator rather than the sole driver of grid investment. Electrification, renewable integration, ageing infrastructure, manufacturing and resilience requirements were already increasing spending before the current data-centre build-out. This provides a broader demand underpinning for the supply chain but also makes attribution important: growth in an electrical-equipment or grid company’s backlog should not automatically be classified as AI-related. Companies such as Siemens Energy explicitly highlight data-centre demand as just one component of a much wider investment cycle in electrification and grid modernisation. While this diversification of demand sources is clearly supportive from a resilience perspective, it also makes attribution more complex, and investors should therefore be cautious about assuming that all backlog growth in power-infrastructure companies is directly linked to AI.

Financing and AI returns disappoint

Another increasingly important consideration is that the scale of data-centre and AI infrastructure build-out is now so large that it is requiring meaningful external financing alongside hyperscaler balance sheets, with structures such as special-purpose vehicles, private credit and asset-backed financing becoming more common as a way to fund these projects. While this does broaden the pool of available capital, it also introduces a new sensitivity: if AI monetisation were to fall short of expectations or if financing conditions were to tighten, the cost of capital could rise and projects could be delayed or even cancelled, even in a scenario where long-term electricity demand remains structurally supportive. The IEA highlights that financing conditions and return expectations are becoming increasingly important determinants of the pace of data-centre deployment.

A strong theme can still be a poor investment

Finally, it is important to remember that even a very strong thematic backdrop does not automatically translate into strong investment returns, particularly if a significant portion of the growth story is already reflected in valuations. In some cases, companies with exposure to AI infrastructure may already be pricing in extended periods of high utilisation, strong order growth and attractive margins. For that reason, the durability and quality of earnings and order conversions become more important than simply identifying thematic exposure in isolation, and the key question for investors is not simply whether a company benefits from AI but rather how much of that AI-linked earnings growth is already embedded in the current valuation.

Listed company universe: The power bottleneck

The table below has been compiled to include both large-, medium- and small-cap companies that fit into the wider power-supply chain within the broader AI ‘picks and shovels’ theme. Inclusion does not imply that AI/data centres are currently material to earnings or that the shares are attractively valued. We have ordered the large-cap benchmarks from highest to lowest 12-month forward P/E ratio, and the remainder (which can be viewed as idea generation) are listed in the reverse order.

Valuation dispersion is creating a more selective opportunity

The valuation data show a clear divergence across the AI power-bottleneck universe over the last 12 months. Of the 37 companies with usable P/E data, 19 have de-rated (P/E move of less than 10%), seven have re-rated (P/E move of more than 10%) and 11 are broadly unchanged. The de-rating cohort has seen a 25% median share-price decline despite a 36% increase in 12-month forward EPS, driving a 42% median contraction in forward P/E. By contrast, the re-rating cohort has delivered a 53% median price gain despite a 2% decline in forward EPS, with P/E expanding by 48%. The broadly unchanged cohort has seen both price and EPS rise by 16%, with a median P/E change of just 2%.

Despite these very different trajectories, current median forward P/E multiples are relatively close at 16.6x, 19.7x and 18.4x, respectively. It is also worth noting that the median 12-month forward P/E multiples across all three groups are less than 20x, with the median 12-month forward P/E of the de-rating group at 16.6x and at a c 10% discount to the MSCI World Index forward P/E of 18.55x.

This suggests the opportunity is becoming more selective: a number of smaller companies have de-rated despite improving earnings expectations, potentially creating relative-value opportunities where earnings upgrades prove sustainable. However, de-rating alone does not imply undervaluation, with execution, balance-sheet and cyclicality risks remaining important.