What is funding TSMC’s expansion and Taiwan’s GDP growth? This scribble traces the money back to its source: a historic capital expenditure cycle originating in the West. The scale of this investment has drawn comparisons to the buildout of the electrical grid or the railroad boom.
The Stargate Scale
The J.P. Morgan Private Bank outlook highlights the $500 billion U.S. “Stargate” project and Europe’s €200 billion InvestAI program as sovereign anchors for this spending. These are not private sector bets; they are state-directed industrial policy initiatives designed to secure “sovereign AI” capabilities.
- Historical Context: In prior general-purpose technology cycles, investment peaked at 2-5% of GDP. The current AI investment boom is tracking towards this, but historical precedent suggests it could double from here. This implies that 2026 is not the peak, but the “mid-cycle expansion” phase.
The Shift from Training to Inference
A critical shift occurring in 2026 is the transition of workloads from “training” (teaching the AI) to “inference” (using the AI).
-
Adoption Rates: Over 300,000 enterprises are now customers of Anthropic, and 45% of businesses pay for LLM subscriptions. The “inference” market is overtaking “training” as the primary consumer of compute.
-
Implication for Hardware: Inference at scale is less tolerant of latency and high power consumption. This shifts the hardware requirement from raw FLOPS (floating point operations) to memory bandwidth. The model must be loaded into memory to respond instantly to a user query. This specific technical requirement is the catalyst for the “Memory Supercycle”.
The Beta Relationship
The correlation, or beta (