Are we really repeating the telecoms crash with AI datacenters?
a day ago
- The AI datacentre boom differs from the 2000s telecoms crash due to slower GPU efficiency gains, rising power consumption, and accelerating AI demand (e.g., agents consuming 10-100x more tokens).
- In telecoms, exponential supply improvements (e.g., fiber capacity increased 100,000x) met overestimated demand (4x overestimate), leaving 95% of fiber dark; in AI, supply improvements are slowing, and demand may be underestimated.
- AI datacenters face high utilization rates, contrasting with telecoms' massive overcapacity; the risk is timing (early buildout) rather than permanent obsolescence.
- Key risks for AI include agent adoption stalling, financial engineering unraveling, or efficiency breakthroughs changing the math, potentially causing short-term corrections.
- Unlike telecoms, AI infrastructure retains value longer due to slower hardware improvements, and built capacity is likely to be absorbed over time rather than left permanently unused.