What is a data centre worth when the tenant leaves?

By Sola J., Investment Manager, AIIM

The previous article in this series asked whether a stranded data centre could become a cold store. The answer was no: alternative uses of a data centre only protect the minimum value the asset reverts to, which is made up of land, shell and captive power. This ends up being worth a fraction of the original investment. The harder question follows, if the floor (i.e. the minimum value) sits that low, what is a data centre actually worth when the lease ends?

This question matters because of how these assets are priced. A stabilised data centre yielding 6% to 7% over a 15-to-20-year lease roughly returns its capital over the term, so a large part of the investment case rests on what the asset is worth after the lease ends. That residual value is usually assumed rather than analysed. The more useful approach is to test what actually supports that value once the original lease falls away.

We start with the assumption that the data centre use remains the best use of the site. In structurally undersupplied markets this is reasonable, since compute demand is what the investment was underwriting in the first place. The goal is to value the asset within its existing use, but conservatively.

The terminal value can then be defined as the value of a fully built, fit for purpose site to the next data centre occupier (we may call this X), less the cost of refurbishing the facility to the standard that occupier requires (we may call this Y). Said differently, terminal value = X minus Y.

X is reasonably well anchored because a data centre is made up of several components that age at different speeds. The land, grid connection or captive generation, the fibre and permits all age slowly. The land appreciates. The shell ages over decades. It is these durable layers that hold the value of X up.

Y is where the difficulty lies because it sits in the component that ages the fastest. Industry estimates 60% - 70% of a data centre’s capex is in its electrical, mechanical and cooling systems. The equipment these systems serve have a lifespan of 3 years while the infrastructure around it is designed to last 15-20 years.

This mismatch is what creates the functional obsolescence. Each new equipment generation does not simply refresh the machines, it arrives with different requirements: rack densities that averaged 5kW to 10kW a decade ago now exceed 100kW for AI workloads, dragging cooling from air to liquid and drawing more power from the grid / captive generation. Artificial intelligence has compressed this cycle further. Moore’s law promised a doubling of computing density roughly every two years, but leading chipmakers now release a new accelerator generation every year, with roadmaps pointing to racks drawing several hundred kilowatts. A facility can therefore be functionally out of step within three years and will pass through several generation equipment in one lease. The tenant absorbs these changes during the lease. But at expiry, the owner inherits whatever mismatch remains.

The refurbishment bill (i.e. the Y) is therefore not artificial. Recent conversions of former crypto mining sites to AI facilities suggest retrofit costs of USD3 – 4 million per MW, roughly a third of a greenfield build, and those were near-empty shells acquired mainly for their power. Refurbishing a fully fitted facility built around a previous generation adds strip-out and demolition first. Depending on how far the market has moved, Y could plausibly reach a fraction of the cost of building new, plus the cost of the supporting engineering change (i.e. converting to liquid cooling, strengthening the floors to carry much heavier high-density racks together with their piping and water, replacing the racking itself, etc).There is also a silent penalty, which is the cost of using the space less efficiently. The constraint in a data centre is power and cooling rather than floor area, so the next occupier deploying dense racks may need only 20 of the facility’s 100 sqm to absorb all the available power. The remaining space earns little, and X, measured across the whole building shrinks accordingly.

Taken together, this arithmetic should discipline the underwriting. If Y can reach a third to a half of a new-build cost, then the terminal value is not 100% of a fit for purpose site but perhaps half to two-thirds of one, before the space penalty. The residual assumed at acquisition should start from the fit for purpose value and then subtract a realistic, generation-adjusted refurbishment cost, with a discount conservative enough to leave room for that retrofit. Underwriting the exit at a full replacement cost assumes that Y is zero, and Y is never zero.

The neocloud segment shows where this logic ends up. Operators renting out GPU capacity (i.e. GPU-as-a-service) have seen rental prices for flagship AI chip more than halve within two years of launch with the hardware typically depreciated over four to six years. Their contracts run for three to five years, not fifteen to twenty, so for the owner the lease-end question is not a second-decade scenario but the first refinancing i.e. the terminal value is a large share of the case, as it arrives sooner, and it carries a larger Y. The faster the tenant’s equipment cycle, the less affordable it becomes to assume the residual rather than analyse it.

The useful consequence is that Y, unlike X, sits partly within the investor’s control, and it is cheapest to control on day one. In practice this means being super-adaptive from year one. Taller halls, stronger floors than today’s equipment requires, modular power blocks, cooling able to serve both air and liquid, and land banked for expansion. This adaptability carries a known, modest premium at entry in exchange for avoiding a large and uncertain refurbishment cost at exit. Any facility built today should be more flexible than today requires, because the equipment it will serve in year fifteen has not been designed yet.

Contracts can carry part of the burden too. Restoration and make-good clauses determine who pays for the strip-out; refresh covenants keep the facility closer to current standards through the lease; renewal options and their pricing decide who captures the site’s scarcity value at expiry. These terms are negotiated at signing, long before they matter, which is when they are cheapest to secure. And in shorter-tenor segments, they matter sooner than the owner tends to assume.

For an investor, it is important to know what the retrofit / refurbishment cost shall be at exit or once the lease ends. Even when faced with two similar tenants, the total investment yield and terminal value could be different if one facility was built with the end in mind (i.e. factoring in a refurbishment cost at the start) and the other was built assuming the same use forever. In the previous article in this series, the question was whether the facility could be used as a cold store, and that answer set only the floor. The better protection, from an investor lens, is a facility that needs little to no retrofitting.

Ultimately, the residual value of a data centre depends on how easily it can serve the next generation of demand. The more adaptable the facility is at the outset, the lower the refurbishment burden is likely to be at lease expiry and the more value the investor can preserve. This means designing and underwriting the asset not only for the tenant occupying it today, but for the occupier that may need to use it fifteen or twenty years from now.