Are we repeating the telecoms crash with AI datacenters?
8 days ago
- #AI Infrastructure
- #Telecoms Crash
- #Technology Trends
- Comparison between AI datacenters and the 2000s telecoms crash highlights key differences in supply and demand dynamics.
- Telecoms crash was due to exponential supply improvements and overestimated demand (4x), leading to unused infrastructure.
- AI datacenters face slowing supply improvements (GPU efficiency gains slowing) and potentially underestimated demand from agent adoption.
- Telecoms infrastructure became obsolete quickly due to rapid technological advancements, while AI hardware retains value longer.
- AI demand growth could be exponential, especially with agent usage increasing token consumption by 10x-100x per user.
- Current AI infrastructure is already at high utilization, struggling to meet demand, unlike telecoms' dark fiber.
- Datacenter CapEx growth in AI is significant but not as dramatic as some reports suggest, with rebranding of existing compute as 'AI'.
- Forecasting AI demand is challenging due to long lead times for datacenter construction and GPU orders, leading to potential overbuilding.
- Key risks for AI datacenters include slower agent adoption, financial market instability, or unexpected efficiency breakthroughs.
- Unlike telecoms, any overcapacity in AI infrastructure is likely to be absorbed over time rather than remaining permanently unused.