Views: 1
Imagine building a football stadium with a capacity of 80,000 people, but only 4,000 attend each match. That's more or less what's happening with the world's most advanced data centers.
Big tech companies are investing record amounts in infrastructure. Capital spending by major US hyper-specialists is expected to rise from $387 billion in 2024 to a staggering $700 billion in 2026. Names like Amazon, Microsoft, Meta, and Alphabet are taking on debt to finance a race no one seems willing to lose.
The problem is that the data doesn't support this euphoria, and there are some figures that no one wants to publicize.
Cast AI, a company specializing in cloud infrastructure optimization, publishes an annual report on the actual state of resource utilization in data centers. And the results are, to say the least, striking.
According to its 2026 edition, the average utilization of GPUs—the most expensive and coveted chips on the market—doesn't exceed 5%. CPUs barely reach 8% utilization. And memory, at 20%, turns out to be the best-utilized resource of the bunch.
In other words, 92% of installed computing capacity is, on average, idle. This is very serious.
So why is investment continuing? The short answer is: fear. Fear of being left behind. Morgan Stanley analysts warn that technology companies have taken on enormous debt to finance their AI expansion, with a real risk of a bubble comparable to that of fiber optics in the 1990s.
There is also a fundamental strategic argument: those who don't build today won't be able to compete tomorrow. Energy, land, grid connection permits, and operational experience are barriers to entry that are very difficult to replicate once others have gotten there first. In that sense, overbuilding can be interpreted not as a mistake, but as a gamble on controlling scarce resources.
The parallels with the telecommunications bubble of the late 1990s are striking: massive debt to build infrastructure based on projections of exponential growth, with the first-mover advantage narrative as the main driving force. Back then it was fiber optic cables; now it's GPU racks.
And yet, some remember that all that "wasted" fiber optic cable in the 1990s ended up being the backbone of the internet as we know it today.
Beyond the financial debate, there is a consequence that is difficult to ignore: the environmental impact. It is estimated that between 2026 and 2030, approximately 100 gigawatts of new data center capacity will come online, effectively creating a sector that doubles its current size in just five years. All of this requires colossal quantities of electricity, water, and land.
When and how will this end? No one knows for sure. Academic analyses anticipate a three- to four-year "turbulence phase," followed by stabilization after 2030, when AI models mature and the actual demand for inference justifies the installed infrastructure.
What does seem clear is that we are in a classic cycle of speculative overinvestment. Perhaps the savvier are seeing something we're missing. Or perhaps no one simply wants to be the last to leave the party.
Veremos lo que nos trae el tiempo.
EN 
ES