The Infrastructure Thesis: What Sustains vs. Disrupts the Trajectory

Now → 18-24 Months: Structural dynamics stable
2027+: Trajectory depends on which forces prove stronger
Beyond: Compounding or disruption?

“The race isn’t just about who has the best models anymore. It’s about who controls the computational substrate on which all models run.”

Table of Contents

Strategic Horizon

Demand Exceeds Supply: TSMC: demand 3-4x capacity through 2027, memory sold out through 2026-27, 7-year grid queues
Efficiency Absorbed (Jevons Paradox): Historically, computing efficiency gains → more use cases, not less infrastructure. DeepSeek gains → 10x more deployments → Same GPU demand
Geopolitical Competition Intensifies: US-China rivalry creates parallel buildout incentives. Neither side can afford to fall behind.
Enterprise Adoption Reaching Scale: Hyperscaler bet: 2026-2027 production deployments. Inference demand could dwarf training demand.
Capital Remains Available: Hyperscalers: multi-year spending without new capital. $1.15T committed 2025-2027 (Goldman Sachs)

Efficiency Revolution: DeepSeek-style 10x improvements change economics. GPT-4 results with fraction of compute.
Alternative Computing Paradigms: Neuromorphic, optical, quantum breakthroughs. Bypass silicon GPU constraints entirely.
Demand Disappointment: Enterprise ROI fails to materialize at scale. $25B services vs $371B infrastructure gap doesn’t close.
Regulatory Intervention: Energy constraints, environmental concerns mounting. Carbon pricing, power allocation restrictions.
Geopolitical Disruption: Taiwan risk (TSMC centrality) = existential supply risk. Conflict, natural disaster, political instability.