The AI CapEx Trade Is Changing: From Picks-and-Shovels to Power

For three years the AI trade had one name on the marquee. You bought Nvidia, you understood that every model everyone was building ran on its silicon, and you were right. That was the obvious trade, and obvious trades work until the whole market is standing on the same side of the boat.

Here is the thing about a compute boom that nobody put on a slide in 2023. A GPU is useless until it has power. Not “power” in the abstract sense of a wall socket, but hundreds of megawatts of firm, always-on electricity, delivered to a specific building, through a substation that has to be built, connected to a grid that has to agree to carry the load. You can order a rack of H-series chips and have it in weeks. You cannot conjure the interconnection in weeks. In some markets you cannot conjure it in five years.

That gap, between how fast you can buy compute and how slowly you can energize it, is the whole investing thesis here. The bottleneck moved. It used to be chips. Now it is watts. And the companies that sell the watts, the turbines that make them, and the copper that carries them have quietly become the best-performing part of the entire AI complex.

The money is real, and it is enormous

Start with the demand signal, because it is not subtle. Google, Amazon, Microsoft and Meta are collectively guiding to roughly $725 billion in capital expenditures for 2026. That is up about 77% from around $410 billion the year before. Amazon sits near $200 billion, Google near $185 billion, Meta at $125 billion to $145 billion, and Microsoft around $120 billion. Four companies. Three-quarters of a trillion dollars. In one year.

A 77% jump in capex is not a maintenance budget. It is four of the most disciplined capital allocators on earth deciding, simultaneously, that the risk of underbuilding AI capacity is worse than the risk of overbuilding it. You can debate whether they are right. What you cannot debate is where the money physically goes. A data center is not mostly chips by cost once you count the shell, the power, and the cooling. And the part that has become impossible to buy off the shelf is the part that turns a field into a live megawatt.

Why watts, not chips, is the binding constraint

Here is the number that reframes everything. In PJM Interconnection, the largest grid operator in the US, the average time from filing an interconnection application to reaching commercial operation rose from under two years in 2008 to over eight years in 2025. Eight years. For an industry that reprices its entire hardware stack every eighteen months, an eight-year queue for the electricity to run it is not a delay. It is a wall.

And it is not just the queue. The physical gear has its own backlog. Substation transformer lead times now exceed 160 weeks, over three years, up from about 140 weeks in 2023. High-power transformers that took 24 to 30 months before 2020 now stretch to about five years. AI infrastructure projects that entered service in 2025 averaged more than seven years to reach operational status. You do not solve a seven-year timeline by writing a bigger check to TSMC.

This is the part the market took a while to price. Everyone understood that AI needed a lot of electricity. Fewer people understood that the supply chain for delivering that electricity, the turbines and transformers and switchgear and the humans who install them, has a multi-year lead time and cannot be scaled by throwing money at it in a single quarter. When demand is capped by factory throughput instead of by orders, the manufacturer gets pricing power. That is the setup you want to own.

The turbine maker who ran out of factory

GE Vernova is the clearest expression of the thesis. Its gas turbine backlog reached 100 GW in Q1 2026, up from 83 GW at the end of 2025, and it guides to at least 110 GW of combined backlog and slot reservations by year-end. About 20 GW of that, roughly a fifth, is tied directly to data centers and AI.

Now here is the punchline. GE Vernova can only manufacture roughly 10 GW of gas turbines a year through 2030. It is targeting about 24 GW of annual capacity by 2028, but until then, demand is not the problem. The factory is. A 100 GW backlog against a 10 GW annual build rate is a multi-year delivery queue where customers are lining up for slots that do not exist yet. When that happens, you raise prices. CEO Scott Strazik said 2026 orders would be priced 10 to 20 points higher than Q4 2025 orders. That is what scarcity looks like on an income statement.

The electrical side tells the same story. GE Vernova's Electrification segment booked $2.4 billion in data-center equipment orders in Q1 2026 alone, which exceeded its entire 2025 total for that category. It shipped 25 gas turbines in the quarter, a 32% year-over-year increase. This is not a company hoping AI shows up. This is a company that cannot build fast enough to serve the AI that already showed up.

Inside the building: cooling, switchgear, and copper

The turbine makes the power. Something has to distribute it, condition it, and keep the racks from cooking. That is where Vertiv and Eaton live, and their numbers are, if anything, louder than GE Vernova's.

Eaton's data-center orders accelerated roughly 200% in Q4 2025, with sales up over 40% year-over-year, and its data-center backlog is described as equal to about eleven years of 2025 construction. Read that again. Eleven years of backlog. Vertiv's organic orders grew 81%, and Q1 2026 revenue rose 30% year-over-year. These are the companies selling the power distribution units, the busways, the liquid cooling, the uninterruptible power systems. The unglamorous guts of the room.

Then there is Quanta Services, which is the labor and construction layer, the people who actually build the grid connections and the electrification work. Quanta posted Q1 2026 adjusted EPS of $2.68 against $2.03 expected, on revenue of $7.87 billion, up about 26%, with a record $48.5 billion backlog driven by grid and electrification work. When you cannot find enough qualified crews to string the wire, the company that has the crews becomes a chokepoint too.

The demand curve is not slowing down

If you think this is a one-year spike, the forecasts disagree hard. The IEA projects global data center electricity consumption to roughly double from about 415 TWh in 2024 to around 945 TWh by 2030 in its base case, with electricity consumption from AI-focused data centers rising about 50% in 2025 alone. In the US specifically, data centers account for around half of the total projected electricity demand increase through 2030. Half of the entire country's demand growth, from server rooms.

On the power side, worldwide data center demand is projected to rise about 27% in 2026 to roughly 132 GW, up from 104 GW in 2025, and to reach about 290 GW by 2030. Bloom Energy's January 2026 report estimates US data center demand nearly doubling from 80 GW in 2025 to 150 GW by 2028. Goldman Sachs projects global data center power demand up about 165% from 2023 levels by 2030, and one estimate cited in April 2026 pegs AI data-center electrification alone at about $1.4 trillion of needed investment by 2030.

Different forecasters, different methods, same direction and same rough magnitude: the electricity demand from this buildout roughly doubles by the end of the decade. When the demand side is that durable and the supply side has a multi-year lead time, the squeeze does not resolve quickly. It compounds.

Nuclear is the tell

Want to know how desperate the hyperscalers are for firm power? Watch where they are willing to go. Meta signed nuclear agreements in January 2026 with Oklo, Vistra and TerraPower for up to 6.6 GW by 2035, including a 20-year Vistra power purchase agreement for over 2,600 MW and a 1.2 GW Oklo campus in Pike County, Ohio. As of May 2026, every major hyperscaler had signed at least one nuclear deal, with 13 announced projects committing over 9.8 GW.

Nuclear is slow, capital-heavy, and politically fraught. Companies do not sign 20-year nuclear PPAs because it is convenient. They sign them because they have looked at the grid queue, looked at their own load growth, and concluded that waiting for the utility to catch up is not an option. When the buyers of power start underwriting their own reactors, that is not a sentiment indicator. That is a company telling you, with its balance sheet, that firm electricity is the scarce input.

The scoreboard already agrees with me

Here is where the thesis stops being a story and becomes a number. Over the trailing year, GE Vernova returned about 125% and Vertiv about 154%. Year-to-date in 2026, GE Vernova was up about 74% and Vertiv about 92%. Nvidia gained roughly 36% in 2025. The picks-and- shovels play on the chip got beaten, badly, by the picks-and-shovels play on the power.

I am not telling you Nvidia is a bad company. It is an extraordinary one, and it still sells the single most important component in the stack. But the market has largely figured out the chip. Thirty analysts cover it, every fund owns it, the story is fully priced. The edge in investing is rarely in the obvious layer once the obvious layer is on the cover of every magazine. The edge is one derivative out, in the layer that is structurally scarce and less crowded.

How I'd think about the risk

None of this is a free lunch, and I would be doing you a disservice to pretend the setup has no downside. These stocks have already run hard. A 154% year means a lot of the good news is in the price, and names that trade on backlog and forward demand get punished violently the moment growth expectations wobble. If the hyperscalers ever blink on capex, if $725 billion turns out to be a peak rather than a way station, the power names fall faster than the market because their multiples assume the spending continues.

The way I frame it: the demand is real and the supply constraint is real, but the valuations already reflect a good chunk of both. This is not the trade you make because it is cheap. It is the trade you make because the constraint is structural and multi-year, and because the thing being sold, delivered electricity, is the one input in the AI stack that money genuinely cannot accelerate past a certain point. Transformers take five years whether you are Microsoft or a machine shop.

So watch the capex guidance, because that is the demand tap and the hyperscalers control it. Watch the backlogs, because a backlog that stops growing is the first sign the squeeze is easing. And watch the interconnection queues and the transformer lead times, because as long as those stay measured in years, the companies that supply the watts keep the pricing power. The obvious trade was the chip. The one worth studying now is whoever plugs the grid into the building.