Somewhere in the American Southwest, there is a data center built in the last three years that will be functionally obsolete before its mortgage is paid off. Not because it will stop working — the concrete and steel will be fine — but because the hardware it was designed to cool and power has already been superseded by chips that generate heat in completely different geometrical patterns, at completely different densities, and at power draws its cooling infrastructure was never engineered to handle.
The stranded asset problem is usually discussed in the context of fossil fuels — coal plants that got written down before the end of their useful life as the economics shifted beneath them. The same logic is quietly converging on AI infrastructure, and the timelines are compressed in ways that make the energy sector’s version look leisurely.
Here is the specific tension: the build cycle for a serious hyperscale data center runs roughly two to four years from land acquisition to full operation. The design decisions that matter most — power density per rack, cooling architecture, water availability, grid interconnect capacity — get locked in during early planning. But the AI chip generation cycle has been running faster than that. Each successive generation of accelerators has pushed rack power density upward substantially. A facility designed around one density assumption can find itself either stranded with headroom it cannot monetize, or — more dangerously — unable to physically accommodate the hardware its customers actually want to run.
Liquid cooling is the clearest illustration. Air cooling, which dominated data center design for decades, works adequately up to a point. Beyond that threshold, the physics get unfriendly: you cannot move enough air through a rack to dissipate the heat without the airflow itself becoming a problem. Direct liquid cooling — whether rear-door heat exchangers, direct-to-chip cold plates, or full immersion — becomes necessary. Retrofitting an air-cooled facility for serious liquid cooling is not impossible, but it involves structural changes, different flooring loads, different plumbing, and different facility management expertise. A building designed in one era of thermal assumptions does not gracefully accept the next era’s requirements.
The geography layer compounds this. Data centers were aggressively sited over the past decade partly on the basis of cheap land and available power — which led to heavy concentration in places like Northern Virginia, Phoenix, and parts of the Pacific Northwest. The water stress calculations those locations carry are already attracting regulatory and public scrutiny. But the more underappreciated issue is grid capacity. The transmission infrastructure in many of these regions was not built to absorb the rate of demand growth that AI workloads represent. Interconnect queues in several U.S. markets stretched to years, not months. A data center that breaks ground today based on a grid interconnect commitment made two years ago may find that commitment renegotiated, delayed, or dependent on new transmission infrastructure that is itself on a long build cycle.
None of this means the build-out stops — the demand pressure is real and the capital is committed. What it does mean is that a significant portion of current investment will produce infrastructure that is mismatched to the workloads it eventually hosts. Some facilities will be underutilized because their power density is too low. Others will be stranded by water constraints before their equipment reaches end of life. The operators with the longest runways will be those who designed for adaptability rather than optimization at a single point in time — modular power distribution, over-provisioned liquid cooling stubs, flexible rack layouts — even though that approach looks wasteful in a spreadsheet today.
In infrastructure, flexibility is usually priced as waste right up until the moment it becomes survival. The AI data center boom is about to learn that lesson the hard way, and the write-downs will be informative.