The Power Paradox Why Energy Efficiency Will be Key to the Data Centre Build-Out

We are only at the beginning of the largest power infrastructure cycle since the 1950s, and it is colliding with hard limits in power supply, equipment availability and skilled labour.

Global data centre capex was around $500bn in 2025, and the four largest US Hyperscalers alone have guided to roughly $725bn for 2026, up 77% year-on-year and the largest single-year concentrated infrastructure cycle in the history of technology. Adding Oracle, Neoclouds, and European sovereign AI initiatives (the EU’s AI Gigafactories programme plus national efforts in France, Germany and the Nordics), 2026 investments clear the $1 trillion mark. From there, the trajectory points to $3-4 trillion a year by 2030, a 40-45% CAGR. Additionally, AI model sizes have grown ~1,200× in eight years while chip efficiency has improved only ~8×, constituting a 150× power gap that compounds with every release. In this power-constrained environment, energy efficiency becomes not only the central design parameter, but the central investment metric, and the central source of value across our portfolio.

Sizing the Build-Out

Of the $3–4 trillion in annual DC capex by 2030, approximately $1.2–$1.5 trillion per year flows into the “energizers” and “builders” that ENETIA invests in, which we view as a structurally protected segment defined by supply constraints, long lead times, and binding reliability requirements.

The Power Gap: Why Efficiency Becomes the Name of the Game

Why Chip Efficiency Cannot Catch Up

Semiconductor efficiency has historically tracked Moore's Law: doubling roughly every 18-24 months and delivering proportional reductions in energy per floating-point operation (FLOP). These efficiency improvements are now diminishing. With leading-edge production already at low single-digit nanometer node geometries, the underlying physics, such as quantum tunnelling, heat dissipation, and lithography resolution, impose binding limits that no incremental node shrink is likely to overcome.

Per Babak Falsafi (SDEA), in the same window that delivered roughly 8× improvement in compute efficiency per kWh across NVIDIA generations (V100 → A100 → H100 → B200), AI model sizes increased approximately 1,200×. The resulting ~150× structural power gap compounds with every release, and we are still at the foothill.

Three forces extend the trajectory further.

• (1) Inference is overtaking training, with billions of queries per day across consumer and enterprise applications; a prime example of Jevons' Paradox, where efficiency gains drive higher absolute consumption.
• (2) Reasoning models require ~100× more compute per query than standard inference, and emerging agentic AI requires an estimated additional 30-100×.
• (3) Physical AI (robotics, autonomous vehicles) and the race toward superintelligence compound demand further.

Consequently, the supply side is responding aggressively. TSMC has guided 2026 revenue +3% driven by AI, with capex +~37% to the high end of a record $52-56bn range. Yet we believe this expansion cannot close the structural gap, because the gap is not about more chips; it is about more efficient delivery of power into chips.

“In a world where every model release scales faster than chip performance can improve, energy efficiency becomes key – not just for sustainability, but as the binding constraint on whether the AI race can be powered at all.”

The Three Constraints: Power · Equipment · Labour

The scale of the build-out is colliding with three simultaneously binding constraints.
First, power: in the US, NREL maps confirm that data centre demand has outpaced an ageing grid, with 31% of transmission and 46% of distribution infrastructure at or beyond their service life. Grid connection queues in Northern Virginia, Texas and Phoenix routinely exceed five years. PJM, the largest US grid operator, has seen hyperscalers commit over $15bn in new power plants. In Europe, the EU Commission has announced a ~€1.2 trillion grid package, with grid congestion costs projected to rise from €5.2bn in 2022 to up to €26bn by 2030.

Second, equipment: lead times for large power transformers have extended from under twelve months to three to four years. Switchgear, UPS systems, bus ducts, transformers, generators and cooling equipment all run on multi-year backlogs. The supply chain, which historically was calibrated to slow-growing regulated utility demand, currently cannot ramp fast enough. This is the central reason that incumbent specialists with manufacturing scale and proven track records command durable pricing power.

Third, skilled labour: licensed electricians and high-voltage engineers are acutely scarce. In the US alone, NECA estimates a structural deficit of more than 80,000 electricians. Engineering and construction firms with DC certification command premiums, as price concerns rank low in customer priorities (~5% of total project spend), and reliability and time-to-market dominate. In a power-constrained AI infrastructure environment, the key metric is increasingly becoming tokens per watt: how much AI inference output a data centre can generate for every unit of electricity consumed.

Rack Density – Towards 1 MW in a Few m²

The current pace of power density increases inside the data centre is staggering. NVIDIA's reference architectures, from V100 (2017) to Feynman Ultra (2029+), trace a ~70× increase in power per rack: 14 kW (V100, 2017) → 26 kW (A100, 2020) → 40 kW (H100, 2023) → 132 kW (GB200, 4Q 2024) → ~180 kW (Rubin, 2026) → ~600 kW (Feynman, 2027+) → approximately 1 MW per rack (Feynman Ultra, 2029+).

To put this into perspective, by the end of the decade, a single rack, being just a few square metres of floor space, will consume approximately one megawatt, the equivalent of roughly 1,000 European households.

Air cooling, the standard for the past three decades, is physically incapable of handling such densities. Direct liquid cooling, single-phase immersion, and 800V DC architectures become not optional but mandatory. The entire thermal and electrical chain from grid connection through medium-voltage switchgear, transformers, UPS, busways, PDUs, and right down to the chip-level coolant needs to be structurally redesigned. This redesign is the source of the multi-decade investment opportunity.

From Bottleneck to Opportunity: Where Value Accrues

Key Efficiency Technologies

Solving the power gap requires a complete redesign of the data centre value chain. No single technology dominates, but instead, a portfolio of innovations across cooling, electrical infrastructure, networking, and software each contribute meaningful efficiency gains. Crucially, each layer is supplied by specialist incumbents with deep engineering know-how, multi-year order backlogs, and high barriers to entry. These are precisely the kind of companies ENETIA invests in.

Why the Build-Out Is Still Early

Sam Altman has described the AI buildout as “a long unfolding exponential curve of technological progress.” We are at the beginning, not the end. The shift from training to inference is just starting; reasoning and agentic AI are still in early commercial deployment; physical AI and superintelligence remain potential future value drivers. JLL Research projects global data centre power capacity more than doubling from current levels to over 200 GW by 2030; in Europe, the first powered land transactions for hyperscale DCs are only now closing. The cycle ahead should be measured in decades, not quarters.

Portfolio Implications – The Infrastructure Premium

Of the $3–4 trillion in annual DC capex projected by 2030, approximately ~40%, approx. $1.2–$1.5 trillion p.a., flows into the physical infrastructure layers that ENETIA invests in. While smaller than the chip layer in absolute dollar terms, this segment is structurally more attractive: less commoditised, more geographically distributed, longer replacement cycles, and protected by binding supply constraints that sustain elevated pricing and margins.

The infrastructure layer is also mission-critical and non-deferrable: a GPU cluster cannot operate without reliable power delivery and thermal management. Hyperscalers' primary decision criteria are reliability, manufacturability at scale, long-term serviceability, and time to market, all factors that decisively reward established specialists over new entrants. Each of the nine technology categories above is supplied by a small set of incumbents with deep moats; this concentration is itself a source of pricing power.

For the ENETIA Energy Transition strategies, this creates a durable, multi-year tailwind across Green Buildings (cooling, HVAC, building automation), Grid & Electrical Equipment (switchgear, transformers, cables, engineering), Energy Management & Storage (power semiconductors, BESS, active cables), and Efficient Industries. 

ENETIA Portfolio Exposures

▸  Cooling & thermal management

▸  Electrical equipment & switchgear

▸  Power cables & grid connection

▸  Active networking cables (AEC/ACC)

▸  Power semiconductors (SiC/GaN)

▸  Engineering & construction

▸  Renewable energy production

▸  Efficient industrial equipment

Our View – Why Now

The $3-4 trillion AI infrastructure cycle is not a one-year capex peak. It is the start of a structural multi-decade build-out of the physical infrastructure that powers digital intelligence. Power demand is growing at 4-7% p.a. globally for the first time in twenty years; grid infrastructure is ageing in the US and EU simultaneously; equipment lead times are at multi-year highs; and labour is scarce. In this environment, the companies that solve the energy efficiency problem – reliably, at scale, with manufacturing depth – capture disproportionate value. That is the core thesis of the ENETIA Energy Transition strategies.