DEV Community: Luige

What's Slowing Down AI Data Centers Isn't the Chip. It's a 128-Week Wait for a Transformer

Luige — Fri, 11 Sep 2026 18:01:28 +0000

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.

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: