An AI data center is a computing facility designed for training and running artificial intelligence models. It differs from a conventional data center in the density of its equipment: racks are filled with servers built around accelerator chips, which draw several times the electricity of ordinary servers and need more elaborate cooling. The largest sites are planned around their power supply and are described by their electrical capacity in megawatts or gigawatts, not by floor area.
These facilities are the physical core of AI infrastructure. They have become a subject of financial interest because of the scale of the spending involved, the range of industries that supply them and the demands they place on electricity grids. For investors the term covers both a category of real asset and a theme running through technology, property, utility and credit markets.
Design and operation
The main workload is of two kinds. Training builds a model by processing very large data sets across thousands of chips working together, which calls for fast networking between servers and can be located wherever power is cheap. Inference runs the finished model to answer requests from users and is more sensitive to distance and response time. Both rely chiefly on the graphics processing unit and on custom accelerators designed by the large cloud providers.
Power density drives the engineering. Air cooling is giving way to liquid cooling piped directly to the chips, and sites need substations, backup generation and high-capacity fiber links. Securing a grid connection often takes longer than constructing the building, so access to power has become the main constraint on new capacity in many regions.
Power usage effectiveness
The standard measure of a facility's energy efficiency is power usage effectiveness (PUE):
Here is all the energy entering the facility, including cooling, power conversion and lighting, and is the energy delivered to the computing, storage and network equipment.
A site that draws 130 megawatts in total to run 100 megawatts of computing equipment has a PUE of , meaning 30 percent overhead. A value of 1.0 would mean no overhead at all. PUE says nothing about how efficiently the chips themselves turn electricity into useful work, nor about water use or the carbon footprint of the electricity.
Owners and financing
- Large technology companies. The biggest cloud and internet groups build and own facilities for their own models and for customers, funding them largely from operating cash flow as capital expenditure.
- Specialist AI cloud providers. Newer firms buy chips and rent computing capacity by the hour, often financing the hardware with debt secured on the chips and on customer contracts.
- Developers and landlords. Colocation operators, including more than one real estate investment trust, build the shell and power systems and let them on a long lease.
- Lenders and infrastructure funds. Projects are financed with bank loans, bonds, private credit and securities backed by lease payments, a form of asset-backed security.
Electricity is typically secured through a long-term power purchase agreement or by building generation on site. Operators have contracted for gas, nuclear and renewable energy in order to obtain enough supply.
Investment theme and risks
The build-out reaches well beyond the owners of the buildings. Chip designers and manufacturers, makers of servers, networking gear, cooling and electrical equipment, construction firms and power producers all sell into it, and it is tracked through measures such as data center systems spending. The pattern of rising and falling investment is described under AI infrastructure cycle.
Several risks are regularly raised:
- Demand. Capacity is being built ahead of proven revenue from AI services, and overbuilding would leave assets underused, as happened with telecom networks around 2000.
- Obsolescence. Chips are replaced by faster generations within a few years, so the assumed useful life strongly affects reported profit through depreciation.
- Concentration. A small number of customers and suppliers account for much of the activity, and some financing arrangements link them to one another.
- Power and local impact. Grid constraints, electricity prices, water use and community opposition can delay or block projects.
Whether returns justify the investment is the central open question, and commentators differ on whether the spending resembles a productive build-out or a bubble.