The US$2.5T Bet: Why AI Capital Will Mostly Reward Users, Not Builders

Why It Matters

The capital‑intensive AI buildout creates a mispricing risk: builders face thin returns, while firms that deploy cheap AI can capture outsized cost savings and revenue growth. This shift reshapes where investors should allocate capital in the AI era.

Key Takeaways

• AI capex projected at $560‑$680 B in 2026.
• Inference cost per million tokens fell from $12 to $0.40.
• Historical infrastructure booms saw builders lose value to users.
• Hyperscalers self‑finance AI buildout, reducing external financing risk.
• Companies with high labor costs stand to gain most from cheap AI.

Pulse Analysis

The surge in AI infrastructure investment mirrors past megaprojects such as railroads and fiber‑optic networks, but with a crucial twist: the assets—GPUs, data‑center racks, and networking gear—depreciate in two to five years instead of decades. This rapid turnover forces hyperscalers like Amazon, Google, and Microsoft to reinvest continuously, consuming cash flow that could otherwise support shareholder returns. While the capital is largely self‑financed, the sheer scale—over 2% of U.S. GDP—means any misallocation will reverberate across markets, making the cost curve dynamics a central focus for analysts.

At the heart of the AI boom lies a steep cost‑curve paradox. Per‑token inference costs have plummeted by more than 30‑fold, yet total enterprise AI spend is climbing as organizations expand token consumption to fill the new price space. This mirrors the bandwidth explosion after fiber‑optic deployment, where cheaper capacity unlocked new applications and drove exponential usage. The result is a deflationary pressure on labor‑intensive sectors—professional services, finance, healthcare, and software development—where AI can replace cognitive work, delivering margin expansion for firms that can integrate the technology effectively.

For investors, the historical capital cycle offers a clear lens: firms that pour capital into capacity often see diminishing returns, while downstream users capture the bulk of value. Consequently, valuation premiums on hyperscalers may be overstated, and equities of companies with high labor cost exposure and strong AI adoption roadmaps could outperform. However, the oligopolistic control of top‑tier models by a handful of cloud giants introduces a counter‑risk; if they retain monopoly rents, the value capture could stay upstream. Monitoring model openness, competitive GPU‑cloud entrants, and enterprise ROI metrics will be essential to gauge which side of the AI value chain will ultimately reward investors.