392 meters. That is how deep the Rogfast tunnel drops beneath the North Sea. For most of us, that number is an abstraction, but for the engineers digging it, it represents a brutal fight against hydrostatic pressure and the sheer weight of the ocean. It is a feat of civil engineering that makes a standard data center build look like a Lego set.
The odd thing is that we are seeing this obsession with extreme physical infrastructure at the exact same time the industry is pushing “flexible” data centers. As highlighted in a recent MIT Tech Review report, there is a growing push toward modularity—the idea that we can just snap data centers together like shipping containers to meet the insatiable demand for compute.
It is a seductive narrative. The pitch is that we can bypass the years of zoning permits and concrete pouring by deploying pre-fabricated units. (Though, personally, I suspect this is mostly to keep shareholders from panicking about the lead time on new sites). But let’s be honest: calling a data center “flexible” is a bit of a lie. You can move a rack of H100s in a container, sure, but you cannot move the 100 megawatts of power required to keep them from melting into a puddle of silicon.
The obsession with modularity is a cope for the fact that our electrical grids are crumbling. We are pretending that the bottleneck is the building, when the bottleneck is the transformer.
This is where the friction hits the fan. You can build the most flexible, modular, snap-on data center in the world, but if you have to wait eighteen months for a utility company to run a high-voltage line to your plot of land, your “flexibility” is a fantasy. It is like trying to put a jet engine in a golf cart; the engine is impressive, but the chassis and the fuel delivery system cannot possibly support the load.
Why are we pretending that a modular rack solves a power grid failure?
The reality is that the physical layer has become the only thing that actually matters. We spent the last decade treating the cloud as a nebulous, ethereal thing—an “abstraction layer” that lived somewhere in the sky. Now, we are suddenly reminded that the cloud is actually just a series of very hot, very loud warehouses that require an obscene amount of water and electricity. The Rogfast tunnel is a reminder of what actual infrastructure looks like: slow, expensive, and permanent.
The industry is currently in a state of denial. We want the speed of software deployment (push to prod in five minutes) but we are bound by the speed of civil engineering (dig a hole in the ground over five years). The “flexible” data center is an attempt to bridge that gap, but it ignores the physics of power distribution. You can’t “flexibly” deploy a substation.
The winners of the next phase of the AI build-out won’t be the labs with the cleverest optimization tricks or the most efficient attention mechanisms. They will be the people who own the power contracts and the land with existing high-voltage access. The “compute” part is becoming a commodity; the “plug” is the actual moat.
I suspect the pivot will happen sooner than the C-suite likes to admit. By Q4 2026, we will see at least one major cloud provider announce a “sovereign cluster” strategy that abandons the modular dream in favor of buying existing industrial brownfields with legacy power infrastructure. They will stop trying to build “flexible” sites and start buying old steel mills and chemical plants because that is where the electricity already lives.
It is a blunt reality. The era of the abstract cloud is over. We are back in the era of concrete, copper, and deep holes in the ground.
Hardware is destiny.
