In the second quarter of 2026 Taiwan Semiconductor Manufacturing Company spent $15.7bn on plant and equipment. Add the first quarter and the half-year total reaches $26.8bn. On the same day it reported those figures, the company raised its full-year capital budget to a range of $60bn-$64bn, up from the $52bn-$56bn it had guided in January. Management gave two reasons: demand that keeps climbing, and tools now bought at inflation-era prices. Free cash flow fell by NT$60.8bn over the quarter, which the company attributed to capital spending outrunning the growth in operating cash. The quarterly report sets it all out.

Money on that scale moves at the speed of a quarter. The plant it buys will run for a decade and more. And the customers who need it sit, overwhelmingly, in one place: by location of the buyer, North America accounted for 78% of TSMC’s revenue in the quarter, against 8% for the rest of Asia-Pacific, 6% for China and 4% each for Japan and Europe. High-performance computing alone was 66% of sales, up a fifth on the previous quarter.

So the chip is designed in America and made in Taiwan, and that circuit closes in a quarter. The electricity to run it does not.

Here the clock changes. The Lawrence Berkeley National Laboratory publishes an annual study of the queue to connect new generation to America’s grid. For projects that began commercial operation in 2025, the median wait from application to switch-on ran beyond five years. Of all the capacity that filed between 2000 and 2020, only 13% had reached commercial operation by the end of 2025. Three-quarters withdrew; a tenth are still waiting (Queued Up: 2026 Edition).

Five years to plug a generator into the grid. And the box that holds the chips cannot be built without one.

Turbines keep similar time. Delivery of a large gas turbine now runs to around 2029, with enquiries booked into 2032, the Nikkei reported in March 2026; the three main makers took combined orders in the final quarter of 2025 that were 70% higher than a year earlier.

An industry that can move $26.8bn in half a year, and an industry that needs five years to connect a single generator to the grid. The market files both under the same word.

＊

The waiting already carries a price.

Mitsubishi Heavy Industries puts a number on it. Its energy segment order backlog rose from ¥4.9trn in the 2024 financial year to ¥7.0trn in 2025. The group backlog stands at ¥13.2trn, ¥3trn higher than a year before. Orders in the gas-turbine combined-cycle business alone reached ¥2.65trn, up four-fifths on the ¥1.47trn of the year before, and that business now holds a backlog above ¥5trn (2025 results presentation). The company ties the expansion to power demand from data centres, alongside the switch from coal and oil and the role of gas in absorbing the swings of renewable supply.

The same deck carries a figure for gas-turbine demand from McCoy Power Reports: 56.2 gigawatts in 2019, then 40.3, 48.3, 30.6, 36.8 and 34.4 through 2024, and then 96.1 in 2025. Nearly three times the previous year, and more than three times the 2022 trough. One year cleared the whole range of the past six.

The most telling figure in the deck is neither the order intake nor the market size. It is the advances.

Mitsubishi Heavy held ¥2.2trn in customer prepayments at the end of the 2025 financial year. Contract liabilities rose by ¥727bn to ¥2.16trn. Free cash flow of ¥893bn was ¥551bn higher than the year before, and the company says the reason was prepayment cash arriving ahead of delivery, led by the turbine business.

Customers are paying, years before they take delivery. The backlog of large gas turbines under contract stands at 74 units and 33GW, up from 48 units and 21GW a year earlier; of the 35 units ordered in 2025, 19 went to the Americas.

That is the price of the queue. The plant is scarce, so buyers pay to hold their place. Mitsubishi Heavy’s cash flow runs high not because it sold something, but because it sold a position in line.

That prepayment is not revenue, though. It is a contract liability, a debt the company owes until the turbine is delivered. Mitsubishi Heavy states in the same deck that from the 2027 financial year it will draw on that high backlog for delivery, funding the working capital it needs. The ¥2.2trn now in hand goes out again to build the turbines it is owed. Cash arrives first, the obligation stays behind, and that order holds for as long as the queue.

At Hitachi’s grid subsidiary, Hitachi Energy, the length of the wait is drawn as a chart. Its order backlog stood at ¥9.2trn, or $57.9bn, at the end of the 2025 financial year, up two-fifths in yen and by a third in dollars. Revenue for the year was $19.8bn, at an adjusted EBITA margin (operating profit before the amortisation of acquired intangibles) of 13.4%. The results deck plots the lead time from order to revenue across a span marked Year 1 to Year 6, sliding back through products, systems and then services and software.

A backlog of $57.9bn that turns into revenue over as long as six years. The company names data-centre demand, alongside grid renewal, as a reason its orders grew.

＊

So much for the physics. The trouble is that Japanese firms do not touch it evenly. What each one takes from it differs by company.

Advantest reports revenue by region in its 2025 results: ¥1.04trn from Asia, ¥44.5bn from the Americas, ¥25.1bn from Japan and ¥23.1bn from Europe. Of total sales of ¥1.13trn, Asia is 91.8% and the Americas just 3.9%. And in a year when the whole company grew 44.7%, the Americas alone shrank by 5.6%.

Set that against TSMC. Its customers are 78% North American; Advantest, which sells the testers those customers’ chips pass through, draws 3.9% of its revenue from the Americas. Put the two firms’ own disclosures side by side and you can see what sits between them. The North American fabless designer places the order, the Taiwanese foundry and its outsourced assembly and test houses buy the equipment, and that equipment ships from Japan to Taiwan. Whether the American data centre gets built reaches Advantest’s revenue only after passing through both of those layers.

By destination of shipment the concentration is starker still: Taiwan ¥569.5bn, South Korea ¥213.3bn, China ¥180.3bn. Taiwan alone is 50.5%, and in the fourth quarter 57.5%.

Tokyo Electron sits close by in geography. Its 2026 revenue split runs China 34.1%, South Korea 22.3%, Taiwan 20.4%, Japan 9.8%, North America 6.8%, South-East Asia 3.8% and Europe 2.8%. Take out Japan and Asia earns four-fifths of the total, much as at Advantest, if not quite its 91.8%; the centre of gravity sits in the same place.

The profit, though, ran the other way. Tokyo Electron’s operating profit for the year fell 10.4%, to ¥625bn from ¥697bn the year before. Advantest’s operating profit rose 118.8% over the same 12 months. Two firms on the same shelf marked “semiconductor production equipment”, living opposite years. That is the first supply chain. American demand reaches both firms only after passing through the capital spending of Taiwan, South Korea and China; and betting on the same layer does not deliver the same result.

Ibiden stands somewhere else again. Its annual securities report lists sales by customer: of ¥416.2bn in revenue, NVIDIA Corporation accounted for ¥122.5bn, or 29.4%, and Intel Corporation for ¥75.3bn, or 18.1%. The top two together are 47.5%. NVIDIA was ¥75.1bn, or 20.3%, the year before, so its share grew by three-fifths in a year.

Ibiden’s exposure is not even to data centres. It is to the shipment of GPUs themselves. The company will put roughly ¥500bn into its electronics business across the three years to 2028 (announced in February). In its third-quarter briefing it said demand for AI-server IC package substrates, the multi-layer boards that seat a chip and wire it out to the wider board, still ran beyond its own capacity, and that it held a share of 70-80%.

Of the four, Fujikura sits closest to the data centre itself. Its information and telecommunications division earned ¥653bn, 55.2% of the group’s ¥1.18trn in revenue, and ¥152.7bn of operating profit, 80.9% of the group’s ¥188.7bn. Almost all the company’s profit comes from this one business (results).

And Fujikura’s binding constraint is neither fibre nor plant. It is hydrogen. In its results briefing the company said a steep ramp in cable output risked outrunning its supply of raw materials, hydrogen among them, and that it had built the effect conservatively into its plan for the year. It makes most of its own hydrogen but buys the extra volume in, and a supplier’s scheduled maintenance has left that volume short. The company was explicit that this has nothing to do with the Strait of Hormuz (Q&A summary). This is not geopolitics. It is a supplier’s routine maintenance. The speed at which optical cable reaches the AI construction site is set there.

Asked in the same session whether fibre could become the bottleneck on data-centre construction, the company said its soundings among American officials and customers found power, chips and fibre all tight, and that fibre was not the single constraint. Of Fujikura’s stated investment of up to ¥300bn, the ¥40bn at its Sakura site comes on stream in 2030, and the remaining ¥260bn or so after that.

Sumitomo Electric does the same optics in a wholly different shape. Its information and communications segment earned ¥326.6bn in 2025, just 6.4% of the group’s ¥5.11trn. Yet it made ¥77.4bn of operating profit, 18.5% of the group’s ¥418.2bn. The company set out a target of ¥970bn in segment revenue and ¥240bn in operating profit by 2028, treble the present figure (outlook). It also names itself a development partner on NVIDIA’s co-packaged optics, the method that folds optical parts into the package (Q&A summary).

Lay the four exposures out and little is shared. Fab capital spending; NVIDIA’s shipment schedule; an American building site and one company’s hydrogen supply; and an optical business that is still small but growing fastest. The single word “AI” covers all four.

＊

One odd thing is happening around this physics.

In June 2025 Hitachi, with the Southwest Power Pool, the regional grid operator across America’s mid-continent, and with NVIDIA, announced work on an AI tool aimed at cutting the time to study a generator’s grid interconnection by 80%. Hitachi’s grid algorithms run on NVIDIA’s computing platform. The first-phase milestones are due between the winter of 2025 and the winter of 2026.

The release notes that America generates 1.28 terawatts of power, and that more than twice that sits idle in the queue, unable to connect because of problems in the interconnection process. The source it cites for this is the same Berkeley queue study. In the pool’s own territory, it adds, the reserve margin could fall from 24% in 2020 to 5% by 2029.

An electricity shortage made by AI, now to be solved by AI. On its own terms, that is reasonable.

There is another side to it, though. Hitachi Energy’s backlog of $57.9bn built up because transformers and switchgear are scarce. The longer the shortage lasts, the longer the backlog runs, the more the pricing holds and the longer the lead times stretch. Its adjusted EBITA margin improved by 3.9 points to 13.4% in 2025, on the volume that came with rising sales and, the company says, on the improving profitability of the backlog itself.

So Hitachi is a company that profits from the length of the queue, and is selling the technology to shorten it. Speed the interconnection study up by 80% and more generation connects, and more transformers are needed; in the near term its own demand rises. But if the queue truly clears, can the pricing behind a $57.9bn backlog hold at today’s strength? A company betting that a physical constraint will ease, drawing its profit from that same constraint. Hitachi Energy holds both inside one set of accounts.

＊

For the reader the implication is a modest one.

To hold the group of Japanese stocks gathered under the “AI” label as a single basket is to hold four separate supply chains as one. The capital spending of Taiwan, South Korea and China; NVIDIA’s shipment schedule; an American building timetable and one firm’s hydrogen; and an optical business only now getting going. The four are called by the same name. They are not the same thing.

Which is the better one this piece does not argue. That is stock selection, a question that sits outwith the structure it has traced.

What can be said is this: when TSMC puts $26.8bn into plant in half a year, that money settles nothing about the world outside the fab. Outside the fab, the generators are queuing. Median wait, five years.

Nothing here is investment advice.