This report analyzes how token commoditization and GPU depreciation are reshaping the economics of the global AI infrastructure cycle. Following three years of 98%+ price compression in AI tokens and accelerated hardware obsolescence, the sector faces a pivotal transition from exuberant growth to sustainable normalization. As AI compute becomes cheaper yet physically constrained by power, semiconductor, and labor bottlenecks, understanding the new equilibrium between efficiency, cost, and scalability is essential for investors.

‍

Our research projects AI infrastructure CapEx to expand 15–25% annually through 2027, driven not by speculative demand but by enduring infrastructure requirements. Value creation is concentrating in a few vertically integrated hyperscalers—Microsoft, Amazon, Google, and Nvidia—which together capture more than 85% of total economic returns. Meanwhile, pure-play LLM providers and GPU-rental “neoclouds” face collapsing margins, liquidity stress, and accelerating consolidation, as validated by CoreWeave’s elevated bankruptcy indicators and Amazon’s $2.22 billion depreciation reversal.

‍

For investors, the path forward is clear: the AI build-out continues, but selectivity matters. The next phase rewards those positioned in infrastructure bottlenecks and vertically integrated platforms, not generic “AI exposure.” The cycle is maturing—not bursting—and disciplined capital allocation will define the winners.

‍