In June I wrote a piece called "Investment outlook for the second half of 2026." Middle East risk versus the AI trade. Three scenarios. Seven things to watch.
It got zero likes.
I understand why now. There wasn't a single line in it that only I could have written. The oil price outlook, the Fed's stance, the P/E on the Nikkei — I had carefully rearranged that day's news. The numbers were borrowed and the judgment was commentary.
But one of those seven items, I remember thinking as I wrote it: this one I can actually explain in my own words.
"Power as the emerging bottleneck in AI capex."
That's all I'm writing about today. I deleted the other piece.
Data center construction moves in lockstep with power plant planning
When you hear "we're building a data center," what comes to mind is probably land, servers, and chips. Most of the news 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 decided is how many megawatts you can actually pull into that location. If the grid has headroom, you connect. If it doesn't, you wait for transmission upgrades, you build your own generation, or you move the project somewhere else.
I buy fuel for a living. In my job, a power plant conversation always starts with when does it run. Until the commercial operation 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 isn't a box that makes electricity. It's a date on which something starts consuming fuel.
That date keeps sliding backwards.
** What's short isn't power plants. It's what goes inside them**
The reason is almost embarrassingly physical. The machines don't arrive.
The gas turbine at the heart of a gas-fired plant comes from an effectively three-company market: Mitsubishi Heavy Industries, GE Vernova, and Siemens Energy. AI data center demand landed on top of that.
The order books show it. GE Vernova signed 21 GW of new gas turbine agreements in the first quarter of 2026 alone, taking total gigawatts under contract from 83 to 100, with backlog rising from 40 GW to 44 GW and slot reservations from 43 GW to 56 GW. Directional lead times on new heavy-duty orders are running around three years. Siemens Energy reported a 62% jump in Gas Services orders to a record, with a book-to-bill of 2.65 — booking work considerably faster than it can ship it.
Then there's the second machine. The transformer.
You can generate all the power you like; without equipment to step the voltage up and push it out, none of it reaches anyone. Large transformers are tight too.
Here I want to be careful with the numbers, because they get quoted loosely. Wood Mackenzie's second-quarter 2025 survey put average lead times for standard power transformers at 128 weeks — about two and a half years — with generator step-up units at 144 weeks, and prices up 77% versus 2019.
Two caveats that usually get dropped. That's the US market, where roughly 80% of large power transformers are imported. And 128 weeks was actually a ten-week improvement over the prior quarter, not a fresh deterioration. More recent reporting through May 2026 puts five-year price inflation nearer 80%, with some specialized orders stretching to four years.
So the picture isn't "getting worse every quarter." It's "stuck at a level that reorders everything downstream of it."
Which is the point. The first thing that could cap AI's growth isn't a shortage of electricity. It's the delivery schedule for 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 grows today. Data centers get built this year and next year. The new equipment shows up several years from now. That gap doesn't close, so everything in between gets absorbed by the fleet that already exists.
The old thermal plant you were going to retire, you don't retire. You shorten the interval between scheduled outages — teiki tenken, in the language of the control room. You push up utilization. Every one of those moves eats into reserve margin.
And on the fuel procurement side, a much longer clock starts running.
You cannot buy the fuel for a plant that starts in 2029 by waiting until 2029. LNG is typically contracted in ten- to twenty-year terms. Years before commercial operation, you're already negotiating the supplier, the pricing formula, and the minimum offtake volume.
Put differently: a single gas turbine ordered today constrains how fuel gets bought into the 2040s.
Now line up the three clocks.
- AI demand forecasts: rewritten in months
- Power equipment lead times: three to five years
- LNG supply contracts: ten to twenty years
Energy procurement is currently making decisions with those three completely out of phase. Capital allocation is hard here not because demand is unpredictable. It's hard because the horizon you can forecast is far shorter than the horizon you have to commit to.
**
"AI is going to run us out of electricity" is a slightly sloppy sentence
Once you see the structure, the familiar phrases get sharper.**
"AI will run us out of electricity." What runs out isn't kilowatt-hours in the abstract — it's supply equipment and its delivery dates. So the response isn't "build more generation," it's "secure the machines" and "decide the queue order for grid connection."
"AI will raise your power bill" works the same way. The path runs less through raw consumption than through equipment cost inflation and the competition to procure it. A 77% rise in transformer prices since 2019 eventually lands on somebody's bill, through transmission tariffs and capital recovery.
And if grid reinforcement can't keep up, sites start bringing their own generation behind the meter. That's the space I'd watch most closely right now.
The counterargument I have to take seriously
There's an obvious objection to everything above, and it comes from inside the industry.
On the same earnings call where GE Vernova reported those order numbers, CEO Scott Strazik said the turbine itself is often not the gating item — that across a three-year project cycle, the EPC buildout, permitting, and fuel availability tend to bind first.
He would know better than I do. And I think he's describing something real: if you're the one selling turbines, the constraint you see is the one your customers complain to you about.
But it doesn't change my answer, for a boring reason. Permitting can be accelerated by policy. EPC labor can be bid up. A turbine slot in 2029 cannot be moved forward by paying more, because the slot in front of it is also sold. And fuel availability — the third item on his own list — is the thing that gets locked in years ahead precisely because the machine date is fixed first.
The turbine may not be the longest pole. It's the one that's hardest to shorten.
``** Wrapping up**
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. That's why nothing was left over.
This time I picked one. The conclusion:
The speed limit on AI right now is set less by semiconductor capacity than by the delivery schedule for generation and transmission equipment. And that schedule is booked out to roughly the end of this decade.
The numbers I've used are mostly US and European. The structure is not. Japan, Ireland, the Nordics, Singapore, Northern Virginia — the queue lengths differ, the concentration of turbine suppliers doesn't.
So here's my question for anyone building on this infrastructure: in your part of the world, what actually arrives last? The chips, the transformer, the grid connection, or the permit?
Next time you read a data center story, try reading it as "when does the machine arrive" instead of "will there be enough power." The view changes a little.
Related, if this was useful: Japan throws away 2.42 TWh of solar power every year — I work in the industry and I think that's fine
I work in fuel procurement in the Japanese energy industry and write about what actually moves energy prices, mostly in Japanese at note.
