Investors have been quick to capitalize on an attractive data center supply-demand mismatch that's sent valuations (and potential returns) through the roof. Layering on the rapid growth of compute-heavy artificial intelligence use cases, it seems like things couldn't be better for the asset class, at least on the surface.
But, it's possible there's now too much of a good thing.
As recently as 2022, 43 percent of investors felt that "data centre technology is relatively future-proof compared to other asset classes,” including the sentiment as one of their top three investment drivers in a DLA Piper survey. Just 24 months later, there are the earliest hints of concern that data centers may be more exposed to tech change than previously thought.
AI's computing requirements have introduced such a step change in data center needs that disruptive hardware innovation is now almost a requirement to meet demand. For a sense of how drastic that disruption might be, consider Larry Ellison's announcement today that Oracle has recently secured building permits for a new larger-than-gigawatt-capacity data center powered by three modular nuclear reactors.
Other hyperscale cloud providers are pursuing similarly aggressive strategies: Microsoft is also exploring small modular reactors and microreactors to support its cloud operations, while Google has invested in novel geothermal power systems.
Traditional data centers, designed for more modest workloads, are struggling to keep pace with the extreme power densities of AI-optimized hardware. New use cases require densities of 50 kW per rack or higher, compared to 5-10 kW for typical enterprise workloads just a few years ago, and some AI training needs could reach as high as 100 kW per rack.
It's not just a question of managing power needs: next-generation hardware requires more sophisticated cooling solutions, which has led to a shift toward new direct liquid cooling technologies. In practice, this means a fundamental change to data center design, from the rack level up to entire facility layouts.
Jordi Sinfreu, JLL's Head of Data Centers for Southern Europe, explains that "higher density implies heavier racks, which affects floorplate loads and footprint, while increased heat generation is resulting in a shift away from traditional air cooling towards various types of liquid cooling."
The upshot is that assets that were recently seen as cutting-edge may now be at risk of accelerated obsolescence. Facilities constructed even five years ago lack the electrical capacity, cooling systems, and physical space to accommodate the latest high-density AI hardware. Retrofitting these data centers is often prohibitively expensive or technically infeasible, leaving owners with difficult choices about decommissioning or repurposing assets.
As demand shifts toward AI-optimized facilities, older data centers may struggle to maintain occupancy and revenue. Even today, many tenants take advantage of lease expirations to upgrade infrastructure by relocating to a new provider. This could cause older facilities to become stranded assets — unable to meet new power and cooling requirements and relegated to lower-value use cases.
On the more bearish end of the spectrum, some industry observers have drawn parallels to previous technological shifts that rendered entire classes of infrastructure obsolete, such as the transition from mainframes to distributed computing.
Long-term tenancy can blunt some of the immediate impact, but the real concern is the erosion of asset value when it comes time to exit. Over the last four years, data center EBITDA multiples have ranged between 25x and 30x, according to data from CBRE, compared to an average of 16x for the wider universe of private infrastructure deals.
A similar premium for future exits is less certain. There is already a clear valuation gap developing between the highest-quality assets and legacy builds. If this trend is developing in the earliest days of AI-led expansion, it's not unreasonable to expect even more drastic (and yet unknown) hardware developments for what is increasingly seen as the main bottleneck for further AI advancement.
In a worst-case scenario, this leads to a situation in which investors purchased assets at nosebleed valuations that could only be justified by future AI-driven demand growth, only to miss out on that upside because the trend they bet on turned out to be too hot to handle.