A few months ago I wrote a piece called "Investment Strategy for the Second Half of 2026." Middle East risk next to the AI trade, three scenarios, seven things to watch.
It got zero likes.
I know why now. There wasn't a single line in it that only I could have written. The oil price outlook, the Fed's posture, the P/E on the Nikkei — I had tidily rearranged that day's news. The numbers were borrowed and the judgment was commentary.
But one of those seven items was different. While I was writing it I thought: this one I can actually explain in my own words.
"The power bottleneck in AI capex becomes visible."
That's the only thing I'm going to write about here. I deleted the other article.
Data center construction moves in lockstep with generation planning
When people hear "we're building a data center," what comes to mind is land, servers, and silicon. Most of the coverage is about who gets the GPUs.
Look at the same project from the power side and the order is different.
The first thing that gets settled is how many megawatts you can actually pull into that location. If there's headroom on the transmission network, you connect. If there isn't, you wait for a grid upgrade, you build your own generation, or you move the project somewhere else.
I buy fuel for power plants. In my job, a plant always enters the conversation as a date — the commercial operation date. Until that date is fixed, you can't decide how many years of fuel to buy, under what contract structure, or from where. So to us, a power plant is not a box that makes electricity. It's the day the fuel consumption starts.
That date keeps sliding to the right.
What's missing isn't the power plant. It's what goes inside it.
The reason is almost embarrassingly physical. The machines don't show up.
The gas turbine at the heart of a gas-fired plant comes from essentially three companies: Mitsubishi Heavy Industries, GE Vernova, and Siemens Energy. Demand from AI data centers poured into that oligopoly all at once.
Here is where the order books stood as of late summer 2026, from each company's own reporting:
GE Vernova closed Q2 2026 with 116 GW of gas power equipment backlog and slot reservation agreements, up from 100 GW a quarter earlier. It expects at least 125 GW under contract by December, and it is taking reservations for 2031 delivery.
Siemens Energy ended its fiscal third quarter (June 30, 2026) with a 69 GW gas turbine backlog.
Mitsubishi Heavy Industries reported a 35 GW large-frame backlog on August 6, 2026, up from 23 GW a year earlier. Orders booked in that quarter are scheduled for delivery between 2028 and 2030.
Do not add those three numbers together. I see people do it and get 220 GW, and it's wrong, because none of them is counting the same thing. GE Vernova's 116 GW mixes firm equipment backlog with paid slot reservations that haven't converted to orders. Siemens' 69 GW is firm only. Mitsubishi's 35 GW covers large-frame machines and leaves out its aeroderivative and mid-size lines. If you take one thing from this article about how to read energy equipment news, take that.
The way the market itself is sized has also shifted. At its results briefing in May 2026, Mitsubishi Heavy put the year's gas turbine market at roughly 70–100 GW, above the ~70 GW a year it had been assuming.
And there's a second machine.
The transformer.
You can generate all the power you want; if you don't have the equipment to step the voltage up and push it out, none of it reaches the place that needs it. Large power transformers are short too.
The clearest data I know of is American, so let me label it as such rather than pretend it's global. In Wood Mackenzie's Q2 2025 US transformer supply survey, power transformers averaged 128 weeks of lead time — about two and a half years — and generator step-up units averaged 144 weeks. Some specialized orders run to four years and beyond. Prices for power transformers are up about 77% since 2019. Demand over the same period rose 119% for power transformers and 274% for generator step-up units.
Japan has its own version of this. Trade press in the Japanese power sector has been warning that the shortage of large transformers could become the binding constraint on both new data center capacity and renewable buildout here.
So the first thing likely to cap AI's growth is not a shortage of electricity. It's the lead time on the machines that make it, change it, and move it.
What happens when lead times run in years
This is where it gets awkward from the operating side.
Demand arrives today. Data centers get built this year and next year. The new equipment arrives several years from now. That gap does not close, which means everything in between gets absorbed by the fleet that already exists.
Old thermal plants you had planned to retire don't get retired. Maintenance intervals get compressed. Utilization goes up. Every one of those moves eats into reserve margin.
And on the fuel side, a much longer clock starts running.
You cannot wait until 2029 to buy the fuel for a plant that starts up in 2029. LNG is typically contracted over ten to twenty years. Years before commercial operation, you're already negotiating the supplier, the pricing formula, and the minimum offtake.
Put another way: a gas turbine ordered today constrains how someone buys fuel into the 2040s.
Which lines up three clocks:
Horizon
AI demand forecasts rewritten in months
Generation equipment lead time 3–5 years
LNG long-term contracts 10–20 years
Decisions are being made with those three completely out of sync. That's the actual state of fuel procurement right now. Capital allocation here isn't hard because demand is unreadable. It's hard because the horizon you can read is far shorter than the horizon you must commit to.
"AI is running out of electricity" is a sloppy way to put it
Once the structure is clear, some familiar phrases sharpen up.
"AI is running out of electricity." What's running out isn't kilowatt-hours. It's supply equipment and the calendar it ships on. Which means the response isn't "build more generation," it's "secure machines" and "decide who gets connected to the grid, in what order."
"AI is driving up power bills." Same correction. The path runs less through consumption than through equipment cost inflation and the competition to procure it. A 77% increase in transformer prices does not stay inside the utility. It comes down to household bills eventually, through transmission tariffs and cost recovery.
And if grid reinforcement can't move fast enough, the facility builds its own generation on site. That — generation inside the data center fence line — is where I think the most movement is happening right now. I'll be honest that I don't know how far it scales. Behind-the-meter generation solves the interconnection queue; it does not solve the turbine queue, because it's competing for the same three order books.
The takeaway
In June I listed seven things to watch. I didn't dig into any of them, and I wrote all seven in other people's words. So nothing was left.
This time I picked one. Here's the conclusion:
What sets the ceiling on AI's growth rate, for now, is not semiconductor capacity. It's the delivery schedule for generation and transmission equipment. And that schedule is booked out at least to the end of this decade.
Next time you read a story about AI and data centers, try reading it as "when does the machine arrive?" rather than "will there be enough power?" The view changes.
The US, the UK, Germany, Ireland, Singapore — the interconnection rules differ, the fuel mix differs, the numbers differ. I'd be curious whether the ordering is the same. If you've worked on a data center project in your country, was the binding constraint the grid connection, the turbine, or the transformer? I'd like to know which one showed up first.
Related, in Japanese:
「AIで電気代が上がる」と言われて1年
臭いあの物質が、脱炭素の主役になる
I work in fuel procurement in the Japanese energy industry and write about what actually moves energy prices, mostly in Japanese at note.
Top comments (0)