Why AI Power Bottlenecks Are Slowing Data Centres and Lifting Copper

Chips get the headlines, but power equipment decides when an AI data centre switches on. Large power transformers now take roughly 128 weeks to arrive in the US, and some heavy-duty gas turbines are quoted at 3-7 years. Software moves in weeks; steel and copper move in years.

Hyperscaler capital spending (Alphabet, Amazon, Meta and Microsoft combined) is guided at roughly $720-745 billion for 2026, yet much of the equipment needed to power it is already sold out. For mining and energy investors, that means physical supply chains set the pace, not announcements.

Why grid hardware, not chips, sets the pace for data centres

Start with what sits on a developer’s critical path. A transformer steps grid voltage up or down so power can be used safely. Switchgear is the heavy equipment that routes and isolates that power, and breakers cut the current when a fault occurs. Without all three, a finished data centre sits dark.

The lead times tell the story. Wood Mackenzie’s survey for Q2 2025 found average US delivery times of about 128 weeks for large power transformers and 143-144 weeks for generator step-up (GSU) transformers, which connect power plants to the grid. Later readings show both exceeding 160 weeks by early 2026.

Demand has outrun factories that take years to expand, and grain-oriented electrical steel, the specialised steel inside large transformers, is a hard constraint. Wood Mackenzie’s modelling points to persistent deficits rather than a passing delay.

Power Magazine reports that large units “remain locked above two years.”

A queue of 128-plus weeks means a project not already in the order book is unlikely to be energised on a typical AI build timeline. Equipment makers with full order books hold pricing power, and you should factor that in.

Transformers: the longest queue

Wood Mackenzie modelling shows a 2025 US supply deficit of about 30% for power transformers and about 10% for distribution transformers. Since 2019, demand has risen 119% for power transformers, 274% for GSUs and 34% for distribution units.

Data centres compete with end-of-life replacement, grid modernisation and the clean-energy pipeline. Foley & Lardner notes that equipment once available within a year now sometimes carries 3-4 year lead times.

Switchgear and breakers: the quieter shortage

Wood Mackenzie estimates 2025 shortages of roughly 40% for high-voltage switchgear, 25% for breakers and 20% for medium-voltage switchgear. An Institute for Supply Management (ISM) survey indicated US companies cannot obtain enough switchgear.

Data centres made up under 2% of the electrical equipment market in 2020, and could approach 40% in accelerated scenarios. One caveat: the evidence base is mostly US data. Utilities and developers elsewhere draw on the same global manufacturers and materials, so the pressure is unlikely to stay local.

A 2024 industry survey cited in the IEA transmission grid report found that growing infrastructure needs are pushing up component prices and lead times for transformers and cables.

Gas turbines and on-site power: when waiting for the grid is not an option

If the grid connection is years away, why not build your own power? Many operators are trying exactly that, using on-site and front-of-the-meter gas generation (plants built beside the project rather than linked through the grid).

Battery storage is one of the tools developers weigh when grid connections are years away, though its role alongside generation is narrower than headlines suggest and does not remove the need for core grid hardware.

The workaround meets its own wall. Heavy-duty turbine lead times have stretched to 3-7 years, and GE Vernova’s chief executive expects the company to be nearly sold out through 2030.

Relief is slow for three reasons:

• Hot-section blades, the parts that withstand the hottest gas, are hard to make.
• Other specialised components are in short supply.
• New plants take years to build.

Clem Chambers, interviewed by host Jeremy Saffra, argued that fixing one bottleneck simply exposes the next. As an illustration, the host said Elon Musk’s AI company runs about 69 gas turbines, plans to buy $2.8 billion more and is reportedly making its own blades. Chambers said UK blade makers were acquired by American firms weeks before he looked to invest. These accounts are reported, not independently verified.

Chambers also said a planned UK AI supercomputer reportedly cannot secure enough electricity until 2035, which he attributes to a lack of political will for nuclear or coal.

With turbine capacity booked towards 2030, treat any AI project promising near-term self-generated power with scepticism. OEM backlogs are a visible gauge of how long the squeeze lasts.

How much electricity will AI data centres actually need?

A terawatt-hour (TWh) is one billion kilowatt-hours, and national consumption is counted in the thousands of them. Data centre load is different from ordinary growth because it is concentrated in a few places, runs constantly and arrives fast.

Because data centre load is concentrated in a few places and runs around the clock, you face a very different planning problem from ordinary demand growth, and that is why grid equipment queues bite so hard.

Globally, the International Energy Agency (IEA) estimates data centres used about 415 TWh in 2024, roughly 1.5% of world electricity. Its base case reaches about 485 TWh in 2025 and about 950 TWh by 2030 (about 3%), a rough doubling. Longer-range IEA-linked figures vary by source, so read them as approximate.

The US picture is steeper. Lawrence Berkeley National Laboratory (LBNL) puts 2023 consumption at 176 TWh, and the Electric Power Research Institute (EPRI) sees 9-17% by 2030, up 60% on prior projections.

US compound annual growth was 7% from 2014-2018 and 18% from 2018-2023, with 13-27% projected for 2023-2028. LBNL’s 2030 scenarios span 9.5-15.3%.

That width is the point. A bullish equipment or copper thesis is a bet on the upper half of the range, not a certainty. Three forces could pull demand towards the low end:

1. Efficiency gains in chips and cooling.
2. Workload management that shifts computing to quieter hours.
3. On-site generation that bypasses the grid.

Against those, the spending pushes high: hyperscaler capex of about $720-745 billion in 2026.

Copper: the metal underneath every other bottleneck

Copper runs through transformers, switchgear, cabling and generation, so every bottleneck above leans on it. London Metal Exchange (LME) copper moved from about $9,000-10,000 per tonne in 2024 and early 2025 to roughly $13,000-14,700+ in 2026, with record highs and volatility. The ISM survey showed ten consecutive months of price increases.

S&P Global projects global copper demand could rise about 50% to 42 Mt by 2040, leaving a potential 10 Mt annual shortfall, about 25% of demand, without meaningful supply expansion.

S&P calls this unprecedented. Demand drivers include electrification, grid expansion, electric vehicles, renewables, AI data centres and defence.

Why the shortfall is projected

S&P points to declining ore grades, permitting delays, geopolitical risk, cost inflation and a thin pipeline of new mines. Even the full risk-adjusted project pipeline leaves supply materially short. Chambers expects markets to gradually realise shortages of copper, explorers and drill bits, which is his opinion rather than a forecast.

The copper structural deficit is driven as much by supply-side limits as by demand, which is why falling ore grades and slow permitting weigh more heavily on the outlook than any single demand forecast.

What could close the gap

The shortfall is conditional, so watch what could narrow it: recycling, brownfield expansion, aluminium substitution, new design architectures and policy. The International Copper Study Group (ICSG) has alternated between modest refined surpluses (about 96-150 kt in some 2026 updates) and deficits, shaped by disruptions such as Grasberg and scrap availability.

Treat the 2040 gap as a warning, not a prediction. Mine approvals, grade trends and recycling rates are the real signals. The long development timelines of mines are where the structural case is strongest; substitution, weaker demand and policy shifts are where it can be undercut. This is general information, not personal advice.

Weighing the physical ceiling against the risk of overbuilding

The structural view rests on 128-plus week transformer lead times, 3-7 year turbine queues, a US data centre share rising from about 4% towards as much as 12%, and a 10 Mt copper gap. The cyclical view points to OEM expansion, wide demand ranges and efficiency gains. Detailed expert evidence on labour and interconnection queues was not available, so those remain open constraints.

Scarcity benefits whoever already holds capacity or reserves; the main danger is demand arriving below the most aggressive forecasts. Four variables would tell you which way it is breaking:

1. OEM backlog trends.
2. Transformer and switchgear lead times.
3. Hyperscaler capex guidance.
4. Copper supply additions and mine approvals.

Exposure looks more attractive if these stay tight while capex holds, and less so if lead times shorten and guidance softens.

Investors weighing the overbuilding risk will find our full explainer on the AI investment cycle useful for seeing how infrastructure spending phases give way to application returns.

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. Financial projections are subject to market conditions and various risk factors.