What Is an AI Data Center?
How an AI data center differs from an ordinary one — power density, liquid cooling, and tenant concentration — and why each difference changes the investment case rather than just the engineering.
An AI data center is a building whose job is to supply electricity, cooling and network connectivity to computers that someone else owns. That is the whole business. The distinctive thing about the AI variety is not what the computers do; it is how much power they draw per square foot, and the consequences of that number run straight through to the investment.
The one number that changes everything: density
A conventional data center rack — web servers, databases, corporate email — draws something in the region of 5 to 10 kilowatts. A rack of AI accelerators built for training large models draws several times that, and the current generation of high-density designs is built for far more.
That has three immediate consequences.
Air stops working. Below roughly 20–30 kW a rack you can move enough air across the chips to carry the heat away. Above it you cannot, and the building has to deliver liquid to the rack — direct-to-chip cold plates, or immersion. Liquid cooling is a plumbing system, with pumps, manifolds, heat exchangers and leak detection, retrofitted into a building type that was designed around fans.
Floor space stops being the unit of value. If one rack draws 100 kW, a 10 MW facility is only about a hundred racks. You do not need a vast building; you need a vast electrical service. This is why the industry prices in kilowatts and megawatts rather than square feet, and why an offering document that talks mainly about acreage is describing the wrong asset. See how data center leases work.
Power becomes the binding constraint. The scarce input is not land, steel or chips. It is a utility willing and able to deliver tens of megawatts to a specific point on the map, within a timeframe. That is the subject of its own page: why power is the constraint.
What the owner provides, and what the tenant provides
In the common model — and in most retail AI data center offerings — the split is:
| The owner provides | The tenant provides |
|---|---|
| Land and building shell | Servers, GPUs, storage |
| Electrical service, switchgear, distribution | Operating software |
| Backup generation and UPS | Its own staff or remote hands contract |
| Cooling plant, including liquid loops | The workload |
| Physical security, fire suppression | |
| Network entrance and cross-connects |
This matters because it defines what you are exposed to. You are not investing in AI models, chips or the companies training them. You are investing in a utility-like landlord whose rent happens to be paid by companies doing AI. If AI demand stays strong but your tenant's own business fails, you have a vacant building full of expensive plumbing. If AI demand cools industry-wide, you have a vacant building and no replacement tenant at the rent you underwrote.
Retail colocation, wholesale, and build-to-suit
Three shapes, with quite different risk:
Retail colocation rents space and power in small increments — a few racks to a few hundred kilowatts — to many tenants. Revenue is diversified; rates per kilowatt are the highest; the operator does more work. CBRE's pricing series tracks a 250–500 kW band precisely because that is the retail unit.
Wholesale rents a whole hall or building, typically 1 MW and upward, to one tenant or a handful. Rates per kilowatt are lower, terms are longer, and the operator does less. Tenant concentration is the risk: one lease can be most of your revenue.
Build-to-suit builds for a named tenant against a signed pre-lease. The construction risk is real but the leasing risk is largely gone — if the lease is actually signed before the money is spent. Many retail offerings describe themselves in build-to-suit language while disclosing, further down, that no tenant has signed. That distinction is the single most important thing to check in an offering circular.
Why AI density raises the stakes on exit
A conventional data center is a reasonably generic building. A high-density AI facility is more specialised: the electrical service is sized for a particular load profile and the cooling is plumbed for particular hardware. Specialisation is good when your tenant wants exactly that and bad when you need a different tenant in a hurry.
In valuation terms, a building with a long lease to a strong tenant is worth a multiple of its income. The same building, empty, is worth something closer to the cost of the land plus whatever the next operator will pay for used infrastructure they may need to re-engineer. Private AI data center deals usually have no public mark between those two states, which is why the risks page spends as long on tenanting as it does on construction.
How the generations differ
Data center designs are often discussed in terms of tiers and generations. For an investor the useful distinction is simpler:
- Pre-AI stock — built for 5–10 kW racks, air-cooled. Still perfectly rentable for ordinary workloads, but it cannot host dense AI training without substantial retrofit.
- AI-retrofit — an existing building with liquid cooling added to part of the floor. Cheaper and faster than new build; constrained by the electrical service that already exists.
- Purpose-built AI — designed around high density and liquid cooling from the start, sited where large power is obtainable. The most expensive per megawatt and the most dependent on the power timeline arriving as promised.
An offering that claims purpose-built AI economics using retrofit costs, or that assumes purpose-built rents in a market that prices pre-AI stock, is telling you something about the sponsor.
Going deeper on the engineering
Everything above is deliberately shallow on the technical side, because an investor does not need to specify a cooling distribution unit. If you want the real depth — cooling topologies, power trains, network design, commissioning — The Definitive Guide to AI Data Centers is a free vendor-neutral reference of 166 chapters, and it is the right place to go.