The US$7T Bet: Why the AI Boom Looks a Lot Like the Dark-Fibre Crash

Why It Matters

The mismatch between massive AI capex and weak commercial returns threatens a costly overcapacity cycle that could strain corporate balance sheets and the broader U.S. economy.

Key Takeaways

• AI compute demand projected 200 GW by 2030; US needs 100 GW
• Goldman Sachs estimates $720B global grid bill through 2030 for AI
• US hyperscalers spent $413B on AI infrastructure in 2025, $600‑700B forecast 2026
• Bain finds AI industry $800B short of revenue needed to cover capex
• 95% of enterprise generative‑AI pilots produce no measurable financial return

Pulse Analysis

The surge in artificial‑intelligence compute mirrors a historic infrastructure frenzy, but the scale is unprecedented. By 2030, global AI workloads will demand roughly 200 gigawatts of power—enough to light up a small nation—while the United States alone must add 100 gigawatts of new capacity. This translates into massive data‑centre construction, with modern hyperscaler sites topping 500 megawatts and some proposals reaching 2 gigawatts, comparable to the power consumption of entire cities. Energy analysts warn that the grid will need to absorb an additional $720 billion in electricity costs, a burden that could outpace the sector’s ability to generate revenue.

Financially, the AI boom is a classic case of capex outpacing cash flow. In 2025, the four U.S. hyperscalers—Microsoft, Amazon, Google and Meta—spent a combined $413 billion on AI infrastructure, a figure set to climb to $600‑$700 billion in 2026. Yet Bain & Company’s modeling shows the industry will be $800 billion short of the annual earnings required to sustain such investment, even under optimistic adoption scenarios. The gap is stark: AI‑related spend could dwarf the $2 trillion in global data‑centre investment projected for 2030, while 95 percent of enterprise generative‑AI pilots fail to deliver measurable returns, leaving a massive pool of underutilized compute.

The parallels to the early‑2000s dark‑fibre crash are striking. Back then, telecoms over‑built miles of fiber based on inflated traffic forecasts, leaving a legacy of “dark” capacity that only later found purpose. Today, AI firms risk creating a similar surplus of idle GPU power, which may eventually be repurposed but will likely cost investors heavily in the interim. Companies that can monetize excess compute at a discount or pivot to new workloads may emerge as winners, but the current trajectory suggests a period of financial strain and strategic reassessment for the AI ecosystem.