Could AI Compute Go Through an Infrastructure Bubble?

The long-term value of AI does not imply that every AI infrastructure investment made today will generate an adequate return.

The world is currently experiencing an extraordinary expansion of AI infrastructure: chips, data centers, power generation, networks, and related financing. The critical question is whether demand and monetization can grow fast enough to absorb this capacity.

Evergreen’s working hypothesis is that AI infrastructure may eventually experience a conventional capital cycle:

High return expectations → capital inflows → rapid capacity expansion → declining returns → localized overcapacity → asset impairment and industry consolidation

This would not necessarily mean that “AI was a bubble.”

Railways, fiber-optic networks, the internet, and cloud computing all experienced periods of excessive investment. Much of the infrastructure was eventually absorbed by the economy. But infrastructure proving useful to society does not mean that the original investors necessarily earned attractive returns.

AI infrastructure faces several specific risks. Improvements in algorithms may reduce the compute required for a given task. Rapid advances in chips and computing architectures can accelerate technological depreciation. Data centers and power assets have long investment lives, while AI technology changes much faster. And if enterprise AI revenues grow more slowly than infrastructure spending, returns on capital will eventually come under pressure.

There is, however, an important counterforce: cheaper compute can stimulate entirely new applications, agents, and business models, creating additional demand.

The central question is therefore not whether AI will create long-term economic value.

It is:

Can AI applications generate cash flow fast enough to keep pace with the rate at which global capital is building compute capacity?

This may become one of the most important questions for understanding the AI investment cycle over the next several years.