L1 / L5 · Capex
Hyperscaler AI Capex Tracker
The four US hyperscalers are guiding combined capex past $300B+ in 2025 and toward the $500B–$725B range in 2026 — roughly three-quarters of it AI infrastructure. Every one of them now describes capacity as power-supply constrained, not demand-constrained — which is why the bottleneck has moved from chips to megawatts.
| Company | Basis | 2024 | 2025 | Note |
|---|---|---|---|---|
| Amazon (AWS) | calendar | ~$83B | ~$100B+ | Largest absolute spend; AI infra the bulk of the growth. |
| Microsoft | fiscal (Jun) | ~$55.7B | ~$80B+ | Azure AI capacity-constrained on power, not demand. |
| Alphabet (Google) | calendar | ~$52.5B | ~$85B (guidance) | TPU build-out + datacenter expansion. |
| Meta | calendar | ~$39B | ~$60–65B (guidance) | Custom MTIA silicon + multi-GW campuses. |
From company earnings releases & guidance, as of early 2026. Figures are total capex (AI infra is the majority of the growth, not the entire base). Oracle and CoreWeave-class neoclouds add materially on top.
Where the capex actually goes
The dollars land on power, land, and silicon. The sourced supply-chain landscapes:
Engineering reference, not investment advice.