Why Data Centers Became Infrastructure Assets
Institutional capital buys cash-flow shapes, not buildings. Data centers joined toll roads and pipelines when their leases started looking like bonds — and AI is testing whether they still do.
The common explanation for why data centers became an institutional asset class — sitting in portfolios next to toll roads, pipelines, and airports — is that demand for computing grew until the buildings became too important to ignore. That explanation is almost exactly backwards. Demand growth makes an industry; it does not make an asset class. Plenty of fast-growing, essential physical businesses never attract a dollar of infrastructure capital, because infrastructure capital does not buy growth, importance, or even buildings. It buys a specific shape of cash flow. Data centers were admitted to the club on the day their cash flows acquired that shape — and understanding exactly what the shape is tells you both why the classification happened and where the AI buildout is now quietly breaking it.
Here is the framing we use: the contract, not the concrete, is the asset. When an infrastructure fund underwrites a data center, the physical object is close to incidental — a means of anchoring a lease. What is actually being bought is a stream: long-tenor, contractually fixed payments from a counterparty whose credit approaches that of a sovereign borrower, with escalators that step the payments upward, attached to a facility the tenant would find genuinely painful to leave. That bundle — tenor, credit, escalation, stickiness — is a corporate bond with real-asset tax treatment and an inflation feature, and it is financeable with cheap long-term leverage precisely because lenders can underwrite it the way they underwrite bonds. Every physical asset class that has ever graduated into "infrastructure" — pipelines, towers, terminals — graduated by acquiring this bundle. The physics differ; the shape is identical.
The three traits, and how data centers acquired them
The first trait is counterparty credit, and it arrived with the hyperscale era. A data hall leased to dozens of small enterprises is a real-estate business with churn, credit losses, and re-leasing risk priced in. The same hall leased for a decade to one of a handful of the most creditworthy corporations in existence is a different security altogether — the landlord's risk has been transformed from occupancy risk into single-name credit exposure on names the bond market prices as nearly riskless. This is the quiet engine of the entire re-rating: the tenant roster changed, so the discount rate changed, so the asset class changed.
The second trait is stickiness — the property that renewal, not departure, is the default outcome at lease end. Migrating a live, deeply integrated deployment out of a facility is operationally disruptive and risk-laden enough that tenants overwhelmingly stay put, which converts a ten-year contract into something closer to a rolling perpetuity with repricing options. The third trait is the moat, and here is where the common account errs most. The moat of a modern data center is not the building, which is replicable, but the position: secured megawatts of power, a grid interconnection with a place in a queue measured in years, land with the right zoning, fiber adjacency, and water or cooling rights. As the binding constraint on new supply migrated from capital to power availability, the scarce asset stopped being anything a competitor could construct and became something a competitor must wait for — and a moat made of waiting time is among the most durable in industrial economics.
The illustration: the tower playbook
The cleanest precedent is the communications tower industry, and it is worth recounting as history because data centers followed its playbook almost step for step. A tower is physically trivial — steel, a foundation, a fence. For years towers were owned by the carriers themselves as operating equipment, exactly as early data centers were owned by the companies whose servers they housed. The transformation came when towers were separated from their users and re-packaged as contract streams: carriers signed long leases with built-in escalators, multiple tenants hung equipment on the same steel, zoning friction made new sites hard to permit, and relocating live network equipment was disruptive enough that renewals approached certainty. Institutional capital did not buy the steel; it bought that lease stack, levered it cheaply, and compounded it into some of the best-performing real-asset businesses of their generation.
Every element maps onto data centers: the corporate spin-out from cost center to asset, the credit tenant, the escalator, the permitting and siting moat, the switching cost that converts tenor into perpetuity. But the mapping also exposes the one element that does not carry over, and it is the load-bearing one. A tower's technology risk lived almost entirely in the tenant's equipment — the steel neither knew nor cared which generation of radio hung from it, so successive technology upgrades meant new tenant equipment on the same unchanged asset. A data center is not so indifferent. Power density, cooling architecture, and floor design are coupled to the computing technology inside, and each leap in that technology reaches backward into the building itself. The tower separated technology risk from asset ownership; the data center only ever partially can. That partial separation is precisely where the underwriting now has to concentrate.
Where AI strains the classification
The AI buildout stresses each trait in turn. Decompose the asset first, because the honest analysis runs on two clocks: the shell — land, structure, power position, interconnection — is genuinely long-lived and technology-agnostic, while the fit-out — electrical distribution, cooling architecture, density provisioning — is coupled to the hardware generation it was designed around, and accelerated computing has been shortening that coupling's useful life while raising its share of total project cost. An asset whose cost is migrating from shell to fit-out is an asset whose economic life is quietly shortening while its financing tenor is not. Second, tenant concentration has intensified: a campus built to suit a single tenant's single workload is a single-name credit instrument with a technology covenant, however excellent the name. Third — and least visible — the renewal logic weakens at the margin: stickiness came from the pain of migrating live, latency-sensitive, deeply integrated workloads, and a training campus is less entangled in those particular switching costs than a cloud region serving thousands of enterprises. None of this means the classification is wrong; it means the classification has become conditional, and the condition is written in the lease: who bears the re-fit cost when the technology turns over, whether tenor matches the life of the fit-out or only of the shell, and what the building reverts to if the single tenant's plans change. Sterling's Embedded Intelligence underwrites these assets by reading the contract stack before the site plan — because the contract, not the concrete, is the asset.
The strongest case against this framework
The serious objection says the contract framing is too clever by half: physical scarcity is doing the real work, and the contracts are merely how it is divided. On this view, secured power in a supply-constrained grid is valuable whatever happens to any individual lease — if a tenant walks from a powered campus in a market where interconnection takes years, the queue of replacement tenants prices the asset, not the departed contract. Concrete plus megawatts is the asset; paper is decoration.
This objection deserves real weight, and the framework should absorb rather than dismiss it: in power-constrained markets, the energized position does provide a floor under the contract, and the strongest assets are precisely those where both layers — scarce position and bond-like paper — reinforce each other. But the objection proves less than it claims, for two reasons. First, the re-leasing floor is itself a market price, and it is procyclical: it holds while compute demand outruns power supply, and it is weakest at exactly the moment a tenant departure is most likely — a general capacity digestion. A floor that exists except when needed is not the thing to underwrite. Second, the floor applies to the shell and the megawatts, not to the fit-out — a replacement tenant re-prices the specialized interior at its value to them, which in a technology transition may be close to salvage. The synthesis is the working rule: physical scarcity sets the recovery value, the contract sets the return profile, and an investor should demand that the label "infrastructure" be earned by the contract while the concrete is priced as the collateral it actually is.
How to apply this framework
- Underwrite trait by trait, never by label. For any data-center exposure — REIT, infrastructure fund, private credit — test the three traits directly: tenor of the contracts, credit and concentration of the tenants, and realistic renewal probability for the specific workload. The label is a summary of past underwriting, not a guarantee of the current book.
- Split every asset into shell and fit-out, and match tenor to the shorter clock. Land, structure, and power position age slowly; density-specific electrical and cooling age on the hardware cycle. Financing or lease tenor calibrated to shell life while economics ride the fit-out is the mismatch that turns an infrastructure yield into a technology bet.
- Read the lease for who bears obsolescence. The decisive clauses are re-fit obligations at technology turnover, escalators against actual power-cost inflation, and reversion terms if the tenant's plans change. A lease that leaves obsolescence with the landlord prices the asset as equipment; one that shifts it to the tenant preserves the bond analogy.
- Treat the power position as recovery value, not return. Secured megawatts and interconnection queues floor the downside in constrained markets, but the floor is procyclical and covers the shell only. Value the contract stream as the return and the energized position as collateral — and be suspicious of any pitch that quietly swaps the two.