DEV Community: Luige

The bottleneck on AI data centers isn't semiconductors. It's a machine with a three-year lead time

Luige — Wed, 12 Aug 2026 13:45:52 +0000

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.

I work in fuel procurement in the Japanese energy industry and write about what actually moves energy prices, mostly in Japanese at note.

Japan throws away 2.4 TWh of solar power every year. I work in the industry, and I think that's mostly fine.

Luige — Sun, 09 Aug 2026 11:19:37 +0000

Every year, Japan generates a large amount of solar power and then deliberately refuses to take it.

The usual reaction is that this is a waste. I buy fuel for a living in the Japanese energy industry, and I don't think it is — at least, not in the way most people mean.

You paid for that electricity, and it never got generated

Start with the scale.

In FY2024 (April 2024 to March 2025), curtailed renewable output across Japan came to about 2.42 TWh. The previous fiscal year it was roughly 1.76 TWh, so it grew about 1.4x in a single year.

For a household using 400 kWh a month, that's roughly a year of consumption for 500,000 homes.

Here's the part that makes it personal. Those solar farms are underwritten by a surcharge that appears on every electricity bill in Japan — the renewable energy levy. For FY2026 it's ¥4.18 per kWh (roughly $0.028).

So the structure is:

We collectively pay to buy this electricity, and then it is never produced.

Stated that way, irritation is the reasonable response. I'm still going to defend it.

If you don't throw it away, the lights go out

Electricity has an awkward property that rarely makes it into public conversation: you can't store it.

More precisely, storing it requires expensive equipment. Left alone, power has to be consumed the instant it's generated.

So a grid operator's job is to keep supply and demand matched continuously — in Japan this rule is called dōji dōryō, "simultaneous equal volume."

This is where intuition breaks. Electricity isn't only dangerous when there's too little. Too much is dangerous too.

When the balance breaks, frequency drifts. Generators trip off the grid to protect themselves. Other generators pick up the load, drift further, and trip in turn. The end state is a wide-area blackout.

Curtailment is the operation that stops this one step before it starts.

We're not throwing power away because the system is broken. We're throwing it away because the system is working.

Before you say "just add batteries"

Most people reach the same place at this point: store it in batteries.

As a technical statement, that's correct. My objection is about cost.

Who buys enough storage to catch 2.42 TWh?

Grid-scale batteries are not a cheap purchase. If government subsidies cover them, the source is tax revenue. If they're recovered as transmission network investment, they land in wheeling charges — which land in electricity rates.

Either path ends on the same bill.

So: to avoid curtailing 6%, how much are you willing to raise electricity prices? Saying "what a waste" without doing that comparison is a little dangerous.

There is money on the other side of "waste," too.

Kyushu curtails 6%. Tokyo curtails zero.

The regional split is where the real problem shows up.

Region	Curtailment rate	Why
Kyushu (main island)	~6.2% in FY2024, ~6.1% in FY2025	Very high solar penetration, limited transmission capacity to send surplus elsewhere
Tokyo area	0% as of June 2025	Demand is large enough to absorb midday solar outright

Kyushu adopted solar faster than anywhere else in Japan, and the transmission lines out of the region have finite capacity. Tokyo simply has enough load to swallow it.

Which means: under one national scheme, only generators in Kyushu have their revenue cut.

I won't defend that part. Curtailment itself may be unavoidable. Concentrating the losses on one region's operators is a separate problem, and a worse one.

So what percentage is acceptable?

Honestly: I don't have an answer.

Is 6% tolerable and 10% the point where you should invest? I can't justify where that line goes.

Even working inside the industry, I rarely meet anyone who can state this number clearly. I've come to think that's because it isn't a technical question. It's a values question.

The less you curtail, the more electricity costs. The cheaper you keep electricity, the more you curtail.

Someone has to draw the line somewhere.

And in Japan, we prepay for the right to vote on that line every month, in the form of the levy on our bills.

Where would you draw it?

I'd genuinely like to hear how this looks from other grids. California, South Australia, Germany, Chile — the curtailment numbers differ, but I suspect the argument doesn't.

I work in fuel procurement in the Japanese energy industry and write about what actually moves energy prices, mostly in Japanese at note.

Related, if you read Japanese: