The compute buildout enters its capital-intensive phase

In the space of one earnings week, four companies guided to as much as 630 billion dollars of capital expenditure for 2026, about 62 percent above a 2025 total that was itself a record. Amazon alone plans around 200 billion. The technology commentary on this has been thorough. The macroeconomic commentary has not, and the macroeconomic consequences are larger.

ChartThe step change in capital commitmentCapital expenditure, billions of US dollars, calendar year. 2026 figures are guidance or the midpoint of a guided range.

Source: Hyperscaler 2026 capex guidance from the February 2026 earnings round. Aggregate guidance of up to about 630 billion dollars represents roughly a 62 percent increase on 388 billion in 2025. Figures are total company capex, not data centre capex.

Three properties that make this different from a normal capex cycle

It is concentrated

Four balance sheets are making a correlated bet on the same three inputs at the same time: chips, power and floor space. Correlated bets create correlated downside. If enterprise adoption disappoints, the entire stack re-rates simultaneously, because there is no diversification across the buyers.

It is front-loaded relative to revenue

Capital spending enters output the moment it is incurred. The revenue it is meant to generate arrives over subsequent years. That asymmetry means the cycle flatters measured growth on the way up, and it means the turn does not require spending to fall. Growth accounting operates on changes, so a programme that simply stops rising contributes zero to growth in the following year while remaining enormous in level terms.

The binding constraint is physical, not financial

Executives are describing themselves as capacity constrained while spending at this scale, which tells you that capital is not what is scarce. Power availability, grid interconnection queues and construction timelines are. That has two consequences: the spending converts into productive capacity more slowly than the headline implies, and the competition for electricity shows up as a price for everyone else on the same grid.

Indicators to monitor in the macro data

Equipment and software in the investment accounts. This is where the spending physically lands. Watch the contribution to quarterly growth, not the level.
Regional divergence. Data centre construction, electrical equipment manufacturing and semiconductor packaging are geographically concentrated. Which regional economies grow fastest is now partly a function of where the buildout sites are.
Retail electricity tariffs in the affected interconnections. Cost recovery is regulated and lagged, so the consumer price effect arrives after the load does.
The second derivative of guidance. The single most informative series for 2027 growth forecasts is whether the next guidance round is above or below this one.

Assessment of risk

Nobody knows whether the return justifies the spending, including the people doing it. The clearest rationale offered is not that the return is proven but that being short of compute is the one error none of them can afford. That is a defensible corporate strategy and a poor basis for a macroeconomic forecast, because it means the spending path is set by competitive dynamics rather than by realised demand. Forecasts that assume capex tracks AI revenue are assuming a discipline that the participants have explicitly said they are not applying.

Sources

Related long-form work. Developed two weeks later in QR-07, Capital Before Revenue.