“AI infrastructure capex persists through 2026 because hyperscaler revenue attribution to AI is now large enough to defend the spend internally.”
Reasoning chain
- 01Hyperscaler 2026 capex guides stepped up sequentially across MSFT, GOOG, META, AMZN.
- 02Token volume per active AI product is growing faster than per-token cost is falling, so gross AI revenue at the hyperscaler layer is expanding.
- 03NVDA remains the default training substrate; alternative silicon (TPU, MI300, Trainium) is additive supply, not displacement, at current demand.
- 04Therefore datacenter revenue can grow into 2026 even if unit ASPs compress, as long as units ship.
Invalidation conditions
- Two or more hyperscalers cut FY26 capex guidance in the same quarter.
- Datacenter unit shipments decline QoQ for two consecutive quarters with no supply explanation.
- A credible disclosure that internal AI revenue attribution has stalled at a top-3 hyperscaler.