The AI Energy Bottleneck: Where the Electricity Actually Runs Out

Most explanations of the AI energy story begin with a frightening number: data centres will consume some enormous quantity of electricity by 2030, therefore the world is running out of power.

That framing is wrong, and being wrong about it will cost you money.

The world is not short of electricity. The United States has spare generating capacity, gas is cheap and abundant, and the physics of making electrons has not changed. What the world is short of is the machinery that converts fuel into electricity and moves it to a building. Turbines. Transformers. Circuit breakers. Pylons. Substations. And the people qualified to install them.

That distinction matters because it changes who gets paid. If the scarce thing were energy, you would buy energy producers. Because the scarce thing is equipment, the money flows to a small group of industrial manufacturers with multi-year order books and, for the first time in two decades, genuine pricing power.

This is the first of two articles. This one is about where the chain actually breaks, why each break is hard to fix, and when each one plausibly clears. The second, who is building the power for AI, maps the companies at each layer and asks which of them have a durable business and which are riding a cycle.

What You'll Learn

• Why "the world is running out of electricity" is the wrong way to describe this problem
• The full chain from a gas molecule to a GPU, and which of its nine links are actually binding
• Why only three companies on earth build utility-scale gas turbines, and what their order books really say
• What an interconnection queue is, and why four to seven years is normal
• Why a transformer takes longer to buy than it takes to build the data center it serves
• Which bottleneck cannot be fixed with money at all
• When each layer plausibly clears, from 2026 to 2032
• The honest case that this shortage is smaller and shorter than the headlines suggest

Nobody Is Short of Electricity. Everybody Is Short of a Connection

Start with the number that gets quoted most, and then look at what sits next to it.

The International Energy Agency projects global data centre electricity consumption roughly doubling from about 485 TWh in 2025 to around 945 TWh by 2030. That is a large increase. It also lands at roughly 3% of world electricity demand.

Now put that growth beside the other things happening to electricity demand over the same period.

Data centres come fourth. Air conditioning adds more new electricity demand between now and 2030 than every AI data centre on earth combined.

That is the whole story in one paragraph. An air conditioner adds a kilowatt to an existing wire. A gigawatt-scale AI campus is the electrical equivalent of dropping a mid-sized city onto a field in rural Ohio and asking for it to be energised by 2028.

The scarce commodity in AI power is not the electron. It is speed to power. The industry has even coined the phrase. Speed to powerSpeed to PowerThe industry's term for how quickly a site can go from a signed lease to energised racks. In an environment where electricity itself is cheap and abundant but connections are scarce, speed to power, rather than the price per unit, has become the thing data center developers compete on.See all terms in the glossary is how quickly a site goes from signed lease to running racks, and it is now what data center developers compete on. The price per unit of electricity barely features in the decision.

Part 1: The Chain From Molecule to GPU

Before you can find the bottleneck you have to see the whole chain. Electricity reaching a GPU has passed through nine distinct layers, each built by different companies on different timescales.

Four of the nine layers are hard constraints. Three are tight. Two are not really problems at all.

Notice what that pattern tells you. The binding constraints are clustered in the middle of the chain, in the layers that involve large, custom, slow-to-manufacture electrical equipment and the permission to use it. Neither end of the chain is the issue. There is plenty of gas, and the chip shortage is a separate story with separate causes, which is covered in the real AI chip shortage.

Walk down it layer by layer.

Part 2: Fuel, the Layer With the Most Slack

Start where the constraint is loosest, because it usefully calibrates the rest.

The United States produces enormous quantities of natural gas and continues to increase output. The Energy Information Administration expects production to rise through 2026 and again in 2027, driven by the Appalachian, Haynesville and Permian basins. Nothing about AI demand strains the country's ability to produce methane.

There is a wrinkle, and it is about pipes rather than molecules. Appalachian gas production has hovered in a band roughly between 34 and 36 billion cubic feet per day since 2020, constrained not by geology but by takeaway pipeline capacity. In the Permian, most long-haul pipelines originate at the Waha hub while a growing share of new processing capacity sits in the northern Delaware basin, which lacks the intra-basin pipe to move volumes to those pooling points.

So gas is abundant, and gas in the specific place a data center wants to burn it is occasionally not. That is a real constraint on siting, and it explains why so many behind-the-meter projects are being announced in West Texas and Appalachia rather than in Northern Virginia. It is not a constraint on the AI buildout as a whole.

Uranium is a different shape of problem and is covered properly in the copper and uranium trade. The short version: the uranium market has a genuine structural supply gap, but it binds on a timescale measured in decades, not on the 2027 to 2030 window that matters for the current buildout. No AI data center is waiting on uranium.

Part 3: Turbines, and Why Only Three Companies Matter

A heavy-duty gas turbineGas TurbineA machine that burns natural gas to spin a shaft connected to a generator. Heavy-duty gas turbines are the workhorse of new dispatchable power generation, and only three companies build them at utility scale: GE Vernova, Siemens Energy and Mitsubishi Heavy Industries.See all terms in the glossary is essentially a jet engine bolted to the ground, burning natural gas to spin a shaft connected to a generator. It is one of the most demanding pieces of machinery humans manufacture. The hot section runs above the melting point of the alloys it is made from, and stays intact only because of single-crystal blade casting and internal cooling passages of extraordinary precision.

Three companies build them at utility scale: GE Vernova, Siemens Energy and Mitsubishi Heavy Industries. That is the entire world market.

This is not a cartel. It is what happens when a product requires decades of accumulated metallurgical know-how, enormous forging presses, a global service network, and a customer base that will not buy from anyone without a thirty-year track record. The barrier is real and it is not going anywhere.

What the Order Books Actually Say

Here is where the reporting usually goes soft, so use the filings instead.

In its second quarter 2026 results, filed 22 July 2026, GE Vernova disclosed that its gas equipment under contract grew from 100 GW to 116 GW in a single quarter, and said it expects to reach at least 125 GW by the end of 2026. In that same quarter it shipped 3 GW.

Set those two numbers against each other. The company is adding contracted volume roughly five times faster than it is delivering it.

Siemens Energy tells a similar story from the other side of the Atlantic. In its third quarter fiscal 2026 results, Gas Services took €10 billion of orders, up 62% year on year, described as a record intake. Group orders of €17.9 billion produced a book-to-bill ratioBook-to-Bill RatioNew orders received in a period divided by revenue billed in the same period. Above 1.0 means the backlog is growing because the company is winning work faster than it can deliver it. Below 1.0 means the backlog is shrinking. It is one of the earliest signals that a capital goods cycle is turning.See all terms in the glossary of 1.57 and lifted the total order backlog to an all-time high of €162 billion.

A book-to-bill of 1.57 means that for every euro of work the company delivered, it booked more than one and a half euros of new work. The queue is not merely long. It is lengthening.

The Detail Most Coverage Misses

There is a distinction inside GE Vernova's 116 GW that changes how you should read it.

More than half of the headline number is queue positions rather than orders. That conversion rate, reservations turning into firm orders quarter after quarter, is one of the cleanest live indicators of whether this cycle is real. It is worth watching more closely than the headline gigawatt figure.

Why Capacity Did Not Get Built Ahead of Time

The obvious question is why three profitable companies with visible demand did not expand sooner.

The answer is that they remember the last cycle. Through the 2010s the gas turbine market was brutally oversupplied. Renewables took share faster than expected, gas plant orders collapsed, and manufacturers carried expensive idle capacity for years. GE's power business was one of the worst industrial value destructions of the decade.

An executive who lived through that does not respond to a three-year demand spike by building factories. They respond by raising prices, extending lead times, and expanding capacity cautiously against contracted volume. Which is exactly what all three are doing. GE Vernova has committed $6 billion of capital expenditure across 2025 to 2028, a substantial figure, but one deliberately sized against contracted demand rather than forecast demand.

Part 4: The Interconnection Queue, Where Projects Go to Wait

Suppose you have solved generation. You still cannot plug in.

To connect anything large to the grid, whether a power plant or a data center, you join an interconnection queueInterconnection QueueThe waiting list to connect a new facility to the electricity grid. Waits of four to seven years are routine in major data center hubs, and up to ten in constrained regions. This, not chips or money, is the sector's binding constraint.See all terms in the glossary. The grid operator studies what your connection does to the stability of everything already attached, calculates which upgrades are required, and allocates the cost. The study process is serial, engineering-intensive, and staffed by a small profession.

Four to seven years is normal. In constrained regions it is longer.

PJM as the Worked Example

PJM is the grid operator for much of the eastern United States, including the world's densest concentration of data centres in Northern Virginia. It runs a capacity marketCapacity MarketAn auction where a grid operator pays generators simply to be available in future years, separately from paying for the electricity they actually produce. It is designed to make sure enough supply exists to meet peak demand. PJM, which covers much of the eastern United States, runs the largest one.See all terms in the glossary, an auction that pays generators simply to be available in a future year, separately from paying for the electricity they actually produce.

That auction is where the strain becomes a number.

In the base residual auction for the 2026/2027 delivery year, capacity cleared at $329.17 per MW-day across the entire PJM footprint, which was precisely the FERC-approved price cap. PJM procured 134,311 MW of unforced capacity and demand response, and noted that its forecast peak load for the delivery year had risen year on year by more than 5,400 MW, driven largely by data center expansion, electrification and economic growth.

Those costs land on everyone connected to the grid, which is why data center power has become a live political issue in Pennsylvania, Ohio, Virginia and New Jersey. Retail bills are rising in states where voters did not ask for a data center. That political friction is itself a constraint on the buildout, and a risk factor that rarely appears in investor decks.

The Queue Is Not a Clean Measure of Demand

One honest caveat, which matters and which most coverage skips.

Interconnection queues overstate real demand, sometimes badly. A developer scouting locations will file requests with several utilities for what is functionally the same project, then withdraw all but one. Utilities have reported very large increases in large-load interconnection requests, and some meaningful share of that is the same gigawatt counted three times.

So treat queue volume as a signal of intensity, not as a forecast. The load that actually shows up will be smaller than the load that applied. Nobody knows by how much.

Part 5: Transformers, the Real Chokepoint

If you want one physical object that explains why the AI buildout is slower than the capital wants it to be, it is the high-voltage power transformerTransformerA device that steps electricity from one voltage to another. Power flows long distances efficiently at very high voltage but equipment can only use low voltage, so every path from a power plant to a server rack passes through several transformers. They are custom-built, slow to make, and currently the tightest link in the chain.See all terms in the glossary.

A transformer changes electricity from one voltage to another. Power travels long distances efficiently at very high voltage, but no equipment can use it at that voltage, so the path from a power plant to a server rack passes through several transformers stepping it down. There is no substitute for one, no way around one, and no software that removes the need for one.

They are also custom-built. A large power transformer is engineered to the specific requirements of the site it will serve. It is not a catalogue item.

The Numbers

POWER Magazine, reporting on the state of the market in 2026, gives the clearest published picture of quoted lead times and price moves.

Roughly two and a half to three years to take delivery of a power transformer, against a pre-2020 norm nearer two. Wood Mackenzie has warned of a shortfall of about 30% for power transformers and 10% for distribution units across the national fleet.

The demand side of that gap is stark. Since 2019, demand for generator step-up transformers is reported up 274%, and for power transformers up 119%.

Why This One Is So Hard to Fix

Three reasons, and none of them respond quickly to money.

Manufacturers are responding. POWER Magazine's tally of announced North American expansions comes to nearly $1.8 billion, including Hitachi Energy investing over $1 billion continentally with a South Boston, Virginia plant expected around 2028, Siemens Energy expanding in Charlotte, North Carolina, and Eaton committing $340 million in South Carolina for 2027.

Read those dates carefully. The relief arrives in 2027 and 2028. It does nothing for anyone trying to energise a site in 2026, and by the time it lands, the demand it was sized against will have moved.

Part 6: The Last Hundred Metres

Inside the fence, the electricity passes through switchgearSwitchgearThe assembly of circuit breakers, switches and protective relays that controls and isolates electrical circuits. It is what allows a section of an electrical system to be shut off safely for maintenance or during a fault, and it sits between the substation and the equipment inside a building.See all terms in the glossary, circuit breakers, and a UPSUPS (Uninterruptible Power Supply)A battery system that carries a facility's electrical load for the seconds or minutes between a grid failure and the backup generators reaching full output. In a data center it is what prevents a momentary voltage dip from crashing thousands of servers.See all terms in the glossary system that carries the load for the seconds between a grid failure and the backup generators reaching full output.

This layer is tight but not severe, and the difference is instructive. Switchgear lead times of roughly 44 weeks are long by historical standards and short compared with transformers. Prices are up around 50% for medium-voltage switchgear and 47% for circuit breakers.

Why is this layer more tractable? Because it is more standardised, more modular, and served by more manufacturers. Eaton, Schneider Electric, ABB and Vertiv all compete here, along with several regional players. Capacity can be added by adding assembly lines rather than by building forging capability and training winders for years.

That is also why this layer is less interesting as an investment. Constraints that are easy to relieve produce temporary pricing power, not durable pricing power. The distinction between a temporary and a durable advantage is the whole subject of understanding economic moats, and it is the lens to bring to Part 2 of this series.

Part 7: Cooling, Solved in Principle and Constrained in Practice

Every watt of electricity a chip consumes becomes a watt of heat that must be removed. As rack densities have climbed, air has stopped being sufficient and liquid coolingLiquid CoolingCirculating coolant directly across chips through pipes instead of blowing cold air at them. Necessary above about 30 kW per rack, and reported to cut energy overhead by up to 30%. It is why older data centers cannot simply be filled with AI hardware.See all terms in the glossary has moved from exotic to standard.

The dominant approach is direct-to-chip, where cold plates sit directly on the processors and carry away the majority of component heat. The technology works. There is no scientific barrier here.

The constraint is commercial. Qualifying a new cooling vendor into a hyperscaler's design takes roughly two to three years of testing, because a leak above a rack of accelerators is a catastrophic event. That qualification cycle is a barrier to entry, which is good news for the incumbents who are already inside it, and it means cooling capacity expands at the pace of qualification rather than the pace of manufacturing.

So cooling is tight, will stay tight for a couple of years, and is not the thing that stops a project. It is the thing that constrains which suppliers capture the growth.

Part 8: The Bottleneck You Cannot Order From a Catalogue

Everything discussed so far can, eventually, be manufactured. This one cannot.

Somebody has to install it. High-voltage electricians, commissioning engineers, controls specialists and substation technicians are all in short supply, and this constraint is the least responsive to money of any in the chain. You cannot import a licensed high-voltage electrician the way you can import a transformer. Apprenticeships run years, licensing is jurisdictional, and the existing workforce is retiring.

The scale of the gap is genuinely contested. Estimates circulating in the industry range from roughly 340,000 to nearly 500,000 unfilled data-centre-related construction roles, and the sourcing behind those figures is inconsistent enough that no single number deserves to be quoted as fact. What is better established is the direction and the shape: contractors report turning away work for lack of crews, and senior executives at both NVIDIA and Microsoft have publicly named electrician availability as a constraint on expansion.

The reason this layer matters most is arithmetic. Manufacturing capacity for turbines and transformers is expanding on announced, dated plans. Skilled trade capacity expands at the speed of apprenticeship pipelines, which is to say slowly and with a lag of years no amount of capital compresses. If the equipment shortage clears around 2028 and the labour shortage does not, labour becomes the binding constraint by default.

Part 9: When Does Each Layer Actually Clear?

Pulling this together into a timeline requires estimates, so treat what follows as a structured judgement rather than a forecast, and note what would prove each line wrong.

The shape to take away: the equipment bottleneck is a late-2020s problem that is already being solved, and the labour and transmission bottlenecks extend past it.

Part 10: The Honest Counter-Argument

A guide that only makes the case for scarcity is a sales document. Here is the case against.

That is the question Part 2 takes up: given a shortage that is genuine now and uncertain later, which companies have businesses that survive the answer either way.

Sources

• GE Vernova, second quarter 2026 Form 8-K, filed 22 July 2026
• Siemens Energy, Earnings Release Q3 FY 2026
• Eaton, first quarter 2026 results, filed with the SEC
• PJM, base residual auction results for the 2026/2027 delivery year
• POWER Magazine, Transformers in 2026: Shortage, Scramble, or Self-Inflicted Crisis?
• Bloom Energy, 2026 Data Center Power Report, survey of 152 decision-makers conducted November 2025
• International Energy Agency, Energy and AI

Frequently Asked Questions

Is the world actually running out of electricity because of AI?

No. The world is running out of the machinery that converts and moves electricity, which is a different problem with different investment implications. The IEA projects data centre consumption roughly doubling from 485 TWh in 2025 to about 945 TWh in 2030, around 3% of world electricity. Over the same period industrial demand grows by roughly 1,936 TWh and electric vehicles by 838 TWh. The shortage is in turbines, transformers, grid connections and electricians.

What is the single biggest bottleneck in powering AI data centers?

High-voltage transformers and the interconnection process. POWER Magazine reported quoted lead times of about 128 weeks for power transformers and 144 weeks for generator step-up units, against a pre-2020 norm nearer two years. Interconnection studies routinely run four to seven years. Neither shortens with money, because the constraint is manufacturing capacity, specialised steel, supplier qualification and engineering staff.

Why can't manufacturers just build more gas turbines?

They are, from a small base and cautiously. GE Vernova is on track for 20 GW of annual output from the third quarter of 2026, 24 GW in 2028, with actions to reach 30 GW in 2030. Against 116 GW already under contract, even the 2030 rate is about four years of production. The deeper reason for caution is memory: the 2010s gas turbine market was severely oversupplied and destroyed enormous value, so incumbents expand against contracted volume rather than forecasts.

How long is the wait to connect a data center to the grid?

Commonly four to seven years in constrained regions. Bloom Energy's 2026 survey of 152 industry decision-makers found utilities projecting delivery timelines roughly 1.5 to 2 years longer than hyperscalers and colocation providers expect, with the gap widening in Northern Virginia, the Bay Area and Atlanta.

What is a slot reservation agreement and why does it matter to investors?

It is a payment to hold a place in a turbine manufacturer's production schedule before a firm order exists. It matters because it is weaker than an order. Of GE Vernova's 116 GW under contract at the end of Q2 2026, 63 GW were slot reservations and 53 GW were firm backlog. The rate at which reservations convert into orders is a better health indicator than the headline number.

Could the AI power shortage turn out to be exaggerated?

Yes, in two specific ways. Interconnection queues double count, because developers file the same project with several utilities. And announced gigawatts consistently exceed energised gigawatts. The equipment shortage is verifiable in audited order books today. The demand forecast that extends it into the 2030s is an estimate that has been revised repeatedly.

Key Takeaways

1. The shortage is of equipment, not energy. Electricity is cheap and abundant. Turbines, transformers, grid connections and electricians are not. That distinction determines who gets paid, and it is why industrial manufacturers rather than energy producers are the direct beneficiaries.

2. Four of the nine layers are hard constraints, and they cluster in the middle of the chain: generation equipment, interconnection, transmission, and substation transformers. Fuel at one end and rack power distribution at the other are not problems.

3. Transformers are the tightest single link. Roughly 128 to 144 weeks quoted, prices up 77% since 2019, and three compounding causes that money does not fix quickly: specialised steel, supplier qualification, and a small ageing winding workforce.

4. Read GE Vernova's 116 GW carefully. Only 53 GW is firm backlog. The other 63 GW are slot reservation agreements, which are queue positions rather than orders. The conversion rate between them is the metric worth tracking.

5. PJM clearing at its price cap is the cleanest evidence of genuine scarcity, because a market at its administrative ceiling would have paid more for capacity that does not exist. It is also why data center power has become a retail electricity bill issue and therefore a political risk.

6. Supplier discipline extends the cycle. Manufacturers burned by the 2010s oversupply are expanding against contracted volume rather than forecasts. This keeps the imbalance in place longer, supports margins, and makes a violent overshoot less likely than in the last cycle.

7. Labour is the constraint that outlasts the others. Equipment capacity is expanding on dated, announced plans. Apprenticeship pipelines cannot be compressed. If the equipment shortage clears around 2028 and skilled trades do not, labour becomes the binding constraint by default.

Part 2 of this series, who is building the power for AI, maps every layer of this chain to the companies supplying it, examines their order books and margins, and works through how the hyperscalers are routing around the bottlenecks entirely.