Five companies are spending more money in a single year than most countries produce in a decade, on infrastructure they cannot plug in.

In Q1 2026 earnings, Amazon guided $200 billion in capital expenditure. Alphabet guided $175 to $190 billion. Meta raised its range to $125 to $145 billion. Microsoft is tracking $120 billion or more. Oracle committed $50 billion. Combined: $700 to $725 billion in 2026. In 2024, the same companies spent $200 billion. Tripled in two years. Amazon’s capex alone exceeds the entire US energy sector’s annual spending. Capital intensity has hit 57% of revenue at Oracle, 45% at Microsoft. These are no longer technology companies. They are infrastructure companies building a new physical layer of civilization on top of a grid that cannot support it.

The binding constraint is no longer chips, capital, or demand. It is electricity. Microsoft disclosed an $80 billion backlog of Azure orders it cannot fulfill because the power grid cannot deliver the electricity to run them. Goldman Sachs documents a US data center capacity shortfall exceeding 11 gigawatts today, widening to 40 gigawatts by 2028. The US interconnection queue, the line of projects waiting to connect to the grid, holds 1.4 to 2.6 terawatts of pending capacity with median wait times of four to five years. Power transformer lead times have stretched to 128 weeks. The companies spending $700 billion on AI compute are building data centers faster than the grid can energize them.

The revenue is real and accelerating. Microsoft’s AI annual run rate hit $37 billion, up 123% year over year. Google Cloud surged 63% to $20 billion. AWS grew 28%, its fastest in 15 quarters. But the revenue-to-capex ratio remains below 0.3. For every dollar of AI revenue, more than three dollars of capex is deployed. The question is whether this is the early phase of a general-purpose technology cycle where returns diffuse over decades, or the largest capital misallocation in market history.

Q4 2026 earnings will be decisive. If the revenue ratio lifts above 0.3 with margin expansion and sustained 2027 guidance, the bear case collapses. If ratios compress and free cash flow stays negative, the cycle confirms as the largest overbuild since the 1990s telecom buildout.

But one trade survives both outcomes: power. If AI succeeds, electricity demand explodes for decades. If AI fails, the data centers are already built, the power purchase agreements already signed, and hundreds of billions in committed infrastructure will consume electricity regardless of whether the compute generates revenue. The GPU may depreciate. The kilowatt-hour does not.

China controls 69 to 70% of global rare earth mining and 85 to 90% of processing. The magnets inside power transformers, the components of grid infrastructure these data centers depend on, flow through Beijing. China added 543 gigawatts of generation in a single year. The United States has a 2-terawatt queue and a five-year wait. The Strait of Hormuz blockade is raising energy costs during the largest electricity demand surge in modern history. Goldman forecasts global data center power demand reaching 1,350 terawatt-hours by 2030, up 220% from 2023, with the US consuming roughly 60% of that total. That forecast assumes energy prices that no longer exist.

The screen says $700 billion in AI investment. The grid says the electricity is not there. The gap between those two numbers is the trade.