The media and some clients keep asking me the same question: What happens when the data center boom outruns demand?

After decades of watching technology booms come and go, I no longer ask whether the industry will overbuild. I ask what happens next.

Right now, the hyperscalers and their growing list of new competitors are spending unprecedented amounts of money on AI data centers.

Overcapacity is not a subject the industry seems eager to discuss publicly. But history argues that we should.

In the late 1990s, telecom companies invested in millions of miles of fiber expecting internet traffic to double every couple of months. It did not. At least, not fast enough. In 2001, much of this capacity was “dark,” and Global Crossing and WorldCom went bankrupt under debt pressure.

However, that is not the whole story. That excess bandwidth enabled YouTube, Netflix, cloud computing, and smartphones to emerge. In other words, the investors were burned. The economy gained the necessary infrastructure. The 19th-century railroad boom offers a similar lesson. Companies built ahead of demand, and many went bankrupt. But a failed railroad company did not necessarily mean a useless railroad. The lesson is that a bad investment for the original backers can still leave something valuable behind.

But This Time Is Different In One Big Way

Fiber can remain useful for decades. A GPU cluster faces a much faster cycle of technological change. A new generation of silicon can push older GPUs out of the most demanding workloads long before those chips stop being useful.

AI data center chipsets may be commercially valuable for only a couple of years before a new generation of silicon supersedes them and makes older versions uneconomical for cutting-edge work.

The physical infrastructure is a long-term investment. But silicon is perishable. If demand falls short of expectations, an underused data center creates two problems: too little revenue and equipment that keeps losing value.

For this reason, financing matters a great deal. Companies that invest in the buildout using their huge cash flows can better afford to misjudge demand. However, those which depend heavily on debt or leases tied to just a handful of customers are highly vulnerable.

How Overbuilding Could Unfold

Overbuilding will likely unfold in multiple stages.

First, prices drop. Compute becomes cheaper to rent. Customers who once waited for capacity now have more leverage. Smaller companies and startups that can’t afford capacity today gain room to experiment.

Second, the weaker players face financial hardship. Expect to see distressed asset sales, consolidation, and bankruptcies.

Third, the used-equipment market grows as bargain hunters, including the larger companies themselves, buy capacity at a discount. Older GPUs will go to less demanding use cases, like inference, video processing, and scientific computation.

Fourth, the infrastructure legacy will remain. Even if some data centers shut down, the power lines, substations, and fiber optic connections installed to serve them will not just vanish. The regions that have upgraded their grid will have an edge when the next round of demand hits.

The oversupply scenario I am describing would make sense to investors. Their returns would be much higher than they would be otherwise. This doesn’t even account for the possibility that a surplus of infrastructure will attract new demand and start the next cycle.

Why Cheaper Computing Could Sustain Demand

An important counterargument deserves attention: More efficient computing can lower costs and encourage greater use. If that additional demand outweighs the efficiency gains, total consumption rises—a phenomenon known as the Jevons paradox.

Again and again, cheaper processing and storage have opened the door to new uses. Every time we have made our processing or storage much cheaper, we have invented more uses for it.

Of course, efficiency improvements in AI model development could reduce the compute requirements for a particular task. However, cheaper intelligence will also open AI to uses that previously cost too much. Running AI models—a process known as inference—may eventually dwarf the workloads driving the current buildout.

Consequently, a glut could be temporary. Capacity may arrive before the applications that will use it.

The Real Danger Is Timing

The danger isn’t having too much capacity forever. It’s getting far enough ahead of demand to bankrupt the companies building it.

Being right about the destination and wrong about the timeline has killed quite a few smart companies.

If I were advising an executive or an investor, here are the questions I would focus on:

  • Who finances the buildout, and how much of it is debt?
  • How concentrated is each operator’s customer base?
  • How quickly do the chips lose value relative to the revenue secured by customer contracts?
  • How secure is the operator’s access to power, and at what cost?

Oversupply would hurt, but it wouldn’t end AI. It would be a reset, painful for some, beneficial for others. The lesson I take from past booms is that infrastructure is a means to an end, not the end itself.