The AI Money Has Stopped Going to Chips. It's Going to Concrete — That's the Signal

The number that matters in markets right now is almost too big to feel: the five largest U.S. cloud and AI providers — Microsoft, Alphabet, Amazon, Meta, and Oracle — of capital spending in 2026, nearly double the roughly $380 billion they spent the year before. That is not a technology story anymore. It is a construction story, an energy story, and, underneath it all, a capital-flow story about where the giants of the industry have concluded the binding bottleneck actually lives.

For decades, when you thought of a tech buildout, you thought of silicon. In 2024 the scarce thing was an Nvidia H100. That was the easy period. The hard constraint has moved three layers down, into the dirt, the wires, and the transformer yards — the physical world. An AI model is, in the end, a machine that needs a building, an electric grid that can feed it, and equipment that no factory overnight can replace. The agents doing the "thinking" run on concrete.

The constraint walked off the chip

Here is the cleanest way to see the shift. In 2025 the race was about compute supply. By 2026, as one market roadmap put it, the scarce resource became the grid connection to power the GPUs: Gartner now projects that 40% of AI data centers will be power-constrained by 2027, and approvals for new grid capacity in major U.S. and European markets run 24 to 36 months. Microsoft has disclosed an $80 billion backlog of Azure orders it could not fulfill because of power constraints — not demand constraints, power constraints.

The evidence inside the supply chain is even more stark, because it is dollar-denominated. Delivery times for the electrical backbone of a data center have stretched from months to years: medium-voltage switchgear at roughly 80 weeks, transformers at 50 weeks and climbing; generator step-up transformers in North America have exceeded 160 weeks in early 2026. Utilities are now buying equipment five years ahead for substations and pre-paying suppliers to hold production slots. When the people who run electric grids — the most conservative buyers on earth — start placing five-year orders, that is not a narrative. That is a physical shortage being priced in real time.

Where the money actually lands

This is the part that most retail coverage skips, and it is the point. It is easy to treat a capex number as a single lump of "AI spending," but the spend does not all land in the same place. The $7 trillion that McKinsey sees flowing into data centers globally by 2030 explicitly excludes semiconductors. The chips are the glamorous line item; the transformers, switchgear, cooling systems, and generators are the physical layer where much of the money actually lands, and where supply is most inelastic.

That scarcity is turning a mature, boring business into a pricing-power business. Wood Mackenzie estimates that data centers' share of the U.S. electrical equipment market could climb from just under 2% in 2020 to 40% in an accelerated scenario. Transformer prices are already being quoted up 4% to 10% over the next year. Put simply: a company that makes a constrained physical component, has years of booked orders, and can pass through price increases is in exactly the position investors describe as an inflection. That is the economics hiding inside "it runs on concrete."

And here is where the macro lens matters. What the hyperscalers are doing is a textbook build-ahead: spend first, monetize later. The reason the financial press keeps calling it a bubble is that the spending dwarfs the revenue. The pure-play AI companies have combined projected 2026 revenue of less than $35 billion against that $660 billion of capex — a mismatch that, taken at face value, looks absurd. But it is the same shape as every prior infrastructure supercycle, from railways to fiber to 5G. You buy the capacity at the start of an exponential curve, before usage catches up, because if you wait for the demand to prove itself you arrive two years late.

The demand side, for now, is not a fantasy. The agent economy is the reason the buildout is accelerating rather than peaking: Goldman Sachs Research expects agentic AI to drive a 24-fold increase in token consumption by 2030, and token processing has already beaten upwardly revised forecasts. Data center power demand is forecast to rise 165% by 2030, with U.S. data center capacity going from about 24 gigawatts today to 110 gigawatts. That is the exponential age compressing time — intelligence, compute, and capital formation all moving faster than the institutions built to supply them.

What this means for an investor

Step back and you can see the whole map at once. The cycle used to be about who built the fastest silicon; it is now about who can pour concrete, pull power, and make a transformer least slowly. Value is migrating down the stack, from Nvidia-class margins toward the physical layer where constrained supply meets effectively pre-committed demand. That is the single most useful repositioning to understand from this number — it is not a rotation into commodities, but a recognition that the constraint on the entire AI trade is now physical.

The discipline, though, is the other half of the story, and it is the half the hype tends to drop. A five-year buildout with two-year lead times means the payoff is back-loaded and the risk is front-loaded. If token demand eventually bends lower, or the power supply fails to arrive, or a financing crack ripples through the balance sheets funding this, all of those booked orders become sunk concrete — and the same physical layer that compounds in a build will mark down hard in a pause. The cycle can be real and still be cyclical.

The lesson is not "buy the transformer makers and nothing else." It is to hold the two horizons at once, the way you have to: the secular force — an exponential technology whose constraint has moved into the physical world — is genuinely enormous, and it will still draw down. Locate where the money is actually going, respect that it moves in multi-year physical rhythms rather than quarterly ones, and do not mistake a stunning capex number for a completed monetization. The agents run on concrete; the concrete runs on years. That gap between them is where the opportunity — and the risk — live.