We Have Seen This Before
The scale of AI infrastructure investment in 2025 is staggering. Hyperscalers are spending tens of billions on data centers. Sovereign wealth funds are backing compute clusters. GPU supply chains are being redesigned from the ground up. The natural question is obvious: is this overbuilding?
Almost certainly yes. And if you study technology history, that is actually the bullish signal, not the bearish one.
The Infrastructure Overbuild Pattern
Railroads in the 1860s. Fiber optic cable in the 1990s. Cloud infrastructure in the 2010s. Each followed the same pattern: massive capital deployment, excess capacity, a bust that wiped out many of the investors who funded the build-out, and then an explosion of value creation enabled by the cheap, abundant infrastructure left behind.
The dot-com bust did not mean the internet was a bad idea. It meant too much capital chased too few viable business models too quickly. But the fiber optic cables remained in the ground, and the cheap bandwidth they provided enabled YouTube, Netflix, and cloud computing. The infrastructure builders lost money. The infrastructure users made fortunes.
What This Means for AI
- Compute will get cheaper, faster than anyone projects. The current constraint on AI adoption is not ideas or talent. It is compute cost and availability. When the infrastructure overbuild results in excess capacity, compute costs will drop precipitously. The AI applications that are marginally uneconomic today will become viable, and entirely new categories of applications will emerge.
- The winners will not be the infrastructure builders. Just as the railroad fortunes went to the businesses that used the railroads, not the railroad companies themselves, the AI value will accrue to the companies that use cheap, abundant AI infrastructure to build products and transform operations. Investing in AI infrastructure companies at peak valuations is a bet on the overbuild. Investing in AI applications is a bet on the aftermath.
- Timing is everything. The overbuild phase is a terrible time to invest in AI infrastructure companies at current valuations. It is an excellent time to build AI applications and capabilities that will benefit from the coming compute abundance. Position for the world after the overbuild, not during it.
The Enterprise Implication
If you are an enterprise planning AI infrastructure investments, consider this: the compute you are buying today at premium prices will be significantly cheaper in 18 to 24 months. Design your architecture for flexibility. Use cloud and API-based models now. Invest in your own infrastructure only when your workload scale genuinely demands it and when prices have normalized.
The infrastructure is being built. The question is whether you position yourself to use it wisely, or to have overpaid for it.