The Data Center Capital Stack in the AI Buildout

McKinsey estimates that meeting projected demand for compute will require roughly $6.7 trillion of data center investment worldwide by 2030, with most of it going to AI-ready facilities. No balance sheet, not even the largest cloud operators, can carry that alone. The money is being split into layers: land and power, the building shell, the powered shell with electrical and cooling plant, and the tenant fit-out with racks and servers. Each layer has a different owner, a different risk, and a different return. Understanding that stack is the first step to investing in it.

Four layers, four different risks

A data center is not one asset. It is a stack of assets that happen to share an address, and capital enters at each level on different terms.

Land and power. The bottom layer is a site with an interconnection agreement, a substation path, water access and entitlements. Its value depends almost entirely on how many megawatts the utility has committed and when they arrive. According to the Lawrence Berkeley National Laboratory's 2024 report for the US Department of Energy, data centers consumed about 4.4% of US electricity in 2023 and could reach between 6.7% and 12% by 2028. Power, not real estate, is the binding constraint, and sites with secured capacity trade at multiples of comparable industrial land. The investor here is taking development and utility-timeline risk, often with no building and no tenant.

Source: https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers

The shell. Above the site sits the building: structure, roof, slab, security perimeter. On its own, a data center shell is a specialized industrial box. The risk profile is construction risk plus the possibility that the market moves before the box is leased. Shell developers are often traditional real estate sponsors who know how to deliver a large building on schedule, and they typically sell or lease to a specialist operator rather than run the facility themselves.

The powered shell. This is where the money concentrates. Transformers, switchgear, generators, uninterruptible power supplies, chillers and the distribution to the data hall can cost more than the building that houses them. For AI workloads with liquid cooling, the mechanical and electrical plant represents the majority of the development cost per megawatt. Powered shell owners usually sign long leases with a hyperscaler or a colocation operator, and the income resembles a triple-net lease on critical infrastructure. This is the layer that most resembles a utility: contracted, long-dated, and sensitive to the creditworthiness of a small number of tenants.

Fit-out and compute. The top layer is the tenant's own equipment: racks, networking, and the GPUs that consume the power. It has the shortest useful life and the fastest obsolescence. It is also the layer where returns are highest on paper and least certain in practice, because the equipment financed today competes against next year's chips. Compute financing, GPU-backed loans and capacity offtake agreements all live here, and they are priced more like equipment finance than like real estate.

Where the returns actually sit

Operators sometimes describe data centers as a single return profile. They are not. A rough institutional view of the stack looks like this:

| Layer | Typical owner | Primary risk | Return character |
|---|---|---|---|
| Land and power | Developer, infrastructure fund | Utility timeline, entitlement | Development gain |
| Shell | Real estate sponsor | Construction, lease-up | Core-plus real estate |
| Powered shell | Infrastructure fund, operator | Tenant credit, plant cost overrun | Contracted infrastructure |
| Fit-out and compute | Tenant, equipment lender | Obsolescence, utilization | Equipment finance |

The important observation is that the middle of the stack, the powered shell, is where most of the dollars go and where the cash flows are most predictable. It is also where institutional capital has concentrated. Infrastructure funds, insurers and sovereign investors have been the natural owners because the checks are large and the leases are long.

The top of the stack carries the most headline risk. Compute capacity contracts and GPU-backed lending have grown quickly, and the underwriting questions are real: what happens to a five-year loan against hardware whose rental rate falls by half in two years, and who bears the utilization risk if demand shifts to a new chip architecture. Investors who want data center exposure without that risk should know which layer they are buying.

The minimum-check problem

Hyperscale deals clear at institutional size. A single AI campus can require more than a billion dollars of powered shell before a tenant moves in, and the equity slice alone runs into the hundreds of millions. That has kept the investor set narrow: a few dozen infrastructure managers and the largest insurers and pension plans.

The result is a structural gap. The asset class generates contracted, inflation-linked cash flow with a credit tenant, which is exactly the profile that family offices, mid-sized institutions and high-net-worth allocators want. Yet those investors have had almost no direct route in. They can buy a listed data center operator, which bundles the whole stack together with corporate leverage, or they can commit to a closed-end infrastructure fund with a ten-year lock and a seven-figure minimum.

This is the problem that digital capital markets structures address. When a powered shell is held in a special purpose vehicle and interests in that vehicle are issued as digital securities, the minimum check can fall from millions to thousands without changing the underlying lease. Distributions can be paid to holders on the same schedule the tenant pays rent. The lease, the utility interconnection, the plant commissioning reports and the tenant's payment history can all be published as verifiable data that an investor, or an investor's software, can check directly rather than reading a quarterly PDF. Commertize's platform is built for exactly this kind of real-asset issuance, where the operating data is as important as the offering document. Our pillar piece on tokenization for energy and digital infrastructure covers the structural mechanics in more detail.

What changes when the layers are investable separately

Three things happen when the stack is unbundled into investable pieces.

Pricing becomes layer-specific. Today a listed operator trades at one multiple that blends development risk, contracted income and equipment exposure. Separate instruments for the powered shell and for the compute layer let the market price each on its own merits. Investors who want the utility-like layer stop paying for obsolescence risk they did not want.

Secondary liquidity becomes possible at the asset level. A closed-end fund interest is transferable in theory and illiquid in practice. An interest in a single powered shell with a published lease and verifiable cash flows is easier to price, and therefore easier to trade, because the buyer can see the asset rather than a fund-level NAV. Settlement on-chain compresses the transfer from weeks of paperwork to a same-day transaction. How the marketplace handles this is described on our how it works page.

New capital enters at the right layer. Developers of land and power sites are often capital-constrained at the moment their asset is most valuable: after the interconnection is secured and before construction begins. A structure that lets them raise against a secured-power site from a broader investor base, rather than selling the site outright to an infrastructure fund, keeps more of the development gain with the people who did the work. Equally, investors who could never write a nine-figure check can now take a position in a contracted powered shell through the marketplace.

The underwriting questions that do not go away

Unbundling the stack does not remove the risks. It makes them legible. Any investor in a digital infrastructure instrument should still be asking the same questions an infrastructure fund asks.

The AI buildout is large enough that every layer of the data center capital stack will need more capital than its traditional owners can supply. McKinsey's $6.7 trillion figure is a demand forecast, not a funding plan. The investors who do well will be the ones who know which layer they are buying, what risk it carries, and whether the data behind the cash flow can be verified rather than assumed.

Source: https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers

Prefer email? Send the offering memo or a short asset summary to deals@commertize.com. We reply within one business day, with no obligation to engage us.