Big Tech Earnings Week: Why Capex Guidance Now Moves the Tape More Than Revenue

Four of the largest technology companies report this week, and the line that will move the tape is not revenue or operating margin. It is capital expenditure guidance for 2027. Combined hyperscaler spending has crossed $400 billion annualised, roughly double the level of two years ago, and the market has spent three quarters rewarding each upward revision. That relationship is now conditional: investors have started asking about payback periods, and the first company to guide capex flat will be treated as a signal rather than a discipline.

The clearest crack appeared this month when Microsoft withdrew from the lease on what would have been the UK's largest AI data centre. The company framed it as a portfolio decision, and it may well have been one, but the timing matters. Two years of ordering accelerators ahead of demand has left the industry with a power problem that no amount of capital can solve on a quarterly cadence: grid interconnection queues in Northern Virginia, Ireland and parts of Germany run five to seven years, and gas turbine slots are effectively sold out through 2028.

That is why the constraint has migrated from wafers to watts. Chip supply has loosened as capacity came online through 2025 and 2026; what has not loosened is the ability to connect a building to a substation. The practical consequence is that capex is being spent on things that do not show up in the AI narrative — substations, transformers, cooling retrofits, long-duration storage — and those line items carry lower margins and longer depreciation schedules than accelerators. The accounting effect arrives with a lag, but it arrives.

Depreciation is the mechanism. A GPU depreciates over five to six years on an accelerated schedule; a substation depreciates over thirty. Shifting the mix toward infrastructure raises the capital base while lowering the revenue attached to each dollar of it, which compresses free cash flow even if operating income holds. That is the arithmetic behind the selloff in the sector three weeks ago, and it is why guidance matters more than the quarter: the market can model revenue, but it cannot model a capex number that changes by $20 billion between calls.

The read-through extends well beyond technology. Semiconductor orders depend on the capex number directly; utilities and electrical equipment makers depend on it indirectly, through the interconnection queue; and industrial gas turbine backlogs are now a proxy for AI demand. A single hyperscaler cutting 2027 guidance by 10% would take the Philadelphia semiconductor index down several percent within a session, not because AI demand has disappeared but because the marginal buyer of semis is a data-centre budget that just got smaller.

The test this week is therefore narrow. Any company that raises 2027 capex while lowering free cash flow guidance is telling investors the build-out is demand-led and the payback is long; the stock can still fall on the second half of that sentence. Any company that holds capex flat is telling them the opposite, and will be read as the first to blink. Neither outcome settles whether the spending is justified — only the next two years of revenue attached to those assets will.