Amazon, Alphabet, Microsoft, and Meta Are Turning AI Capex Into a Return-on-Compute Test

The market used to reward hyperscalers for announcing bigger AI budgets. That framing is fading. The new question is whether Amazon, Alphabet, Microsoft, and Meta can turn huge capex into a sustainable return-on-compute model.

That matters because the input costs are not static. Memory chips are more expensive, electricity is tighter, skilled labor is constrained, and the power stack is becoming a real bottleneck. The same dollar of capex now buys less incremental AI capacity than it did a year ago.

In that environment, capex is not automatically bullish. It is only bullish if the return profile remains strong enough to justify the scale.

Amazon's shareholder letter said it expects about $200 billion of capex in 2026. Alphabet's Q1 call lifted its full-year capex range to $180-$190 billion. Meta is still guiding to $125-$145 billion, and Microsoft remains in the same broad investment race.

That means the relevant unit of analysis is no longer the company. It is the compute stack: land, power, chips, networking, software, and the ability to keep utilization high enough to recover the cost.

The read-through for suppliers is straightforward. If compute is scarce and expensive, then the winners include the memory names and equipment makers that sit closest to the bottleneck: Micron, SK Hynix, Samsung Electronics, ASML, Applied Materials, and Lam Research.

The obvious beneficiaries are chip and infrastructure suppliers. But the less obvious consequence is that every extra dollar of capex now needs to show up somewhere visible: revenue per watt, revenue per rack, or revenue per GPU.

That is why the market keeps returning to memory pricing and power availability. If the cost of building AI capacity keeps rising faster than the monetization curve, the capex surge becomes a margin tax before it becomes a growth engine.

For global investors, the AI spend race is now one of the most important read-throughs on the macro tape because it connects U.S. earnings, Asian semiconductor capacity, and data-center power demand into a single trade.

The hyperscalers are not pulling back. They are doubling down. But the market is no longer rewarding the spend itself; it is rewarding proof that the spend creates durable economics.

That is a better framework for 2026 because it separates winners from merely large spenders.

The next earnings season will matter less for growth optics and more for return optics.