Home / Wiki / A / AI Infrastructure

AI Infrastructure

Technology and VendorsUpdated 16 Nov 2026

AI infrastructure is the combination of hardware, facilities and software needed to train and run artificial intelligence models at scale. It includes the specialized chips that perform the calculations, the servers and networks that connect them, the storage that holds data and models, the buildings and power systems that house them, and the software that schedules and manages the work. Most of it sits inside an AI data center.

The term came into general use in financial markets after 2022, when the success of large language models led technology companies to raise their spending on computing capacity sharply. It is used both in an engineering sense, for the systems themselves, and as the name of an investment theme covering the companies that design, build, supply and finance them.

01

Layers of the stack

  • Compute. Training and running models relies on accelerators, chiefly the graphics processing unit and custom chips designed for machine learning, assembled into servers with large amounts of fast memory.
  • Networking. Thousands of chips must exchange data with very little delay, which requires high-speed switches and optical links inside the facility and long-distance fiber, including leased dark fiber, between sites.
  • Storage. Training data, model files and the records of user requests are held on systems built for rapid reading and writing.
  • Facilities and power. Buildings, cooling, electrical equipment and grid connections or on-site generation supply and remove the energy the chips consume.
  • Software. Programming frameworks, scheduling tools and model-serving systems allocate capacity among jobs. Much of the layer is delivered to customers as a cloud service billed by usage.
02

Who builds and supplies it

The largest buyers are a handful of global cloud and internet companies, often called hyperscalers, which build capacity for their own products and rent it to others. Around them are specialist providers that rent accelerator capacity, colocation landlords, developers of AI models that contract for capacity years ahead, and governments funding national computing facilities.

The supply chain is long and in places highly concentrated. A small number of firms design the leading accelerators, and their manufacture depends on a few advanced chip foundries, memory makers and suppliers of production equipment. Further along are makers of servers, network gear, cooling systems, transformers and turbines, together with construction contractors and electricity producers. Because advanced chips are treated as strategically important, the trade is also shaped by export controls, notably the restrictions the United States has placed on sales of advanced chips to China since 2022.

03

Economics

Building the infrastructure is a matter of capital expenditure on a scale more usual in energy or telecommunications than in software. The spending appears on company accounts over time through depreciation, so the assumed useful life of chips and servers has a large effect on reported profit. Chips are superseded quickly, while buildings and power connections last for decades.

Revenue comes from renting capacity or from selling services that use it. A central measure is utilization: how much of the installed capacity is earning money. The cost of producing a given quality of model output has been falling fast as chips and software improve, a trend followed by benchmarking services such as Artificial Analysis. Falling unit costs widen the range of uses, but they also mean that the price a provider can charge for a fixed amount of computing declines over the life of the equipment.

Funding has moved from company cash flow toward outside capital. Operators have turned to bonds, leases, private credit and joint ventures with infrastructure investors, in some cases keeping the debt off their own balance sheets.

04

Investment theme and risks

Investors approach the theme through shares in chip, equipment, cloud and power companies, through property and infrastructure funds, and through thematic investing products such as a specialized exchange-traded fund. The rise and possible slowing of the spending is discussed under AI infrastructure cycle, and one way of sizing it is data center systems spending.

The main risks follow from the features described above. Capacity is being added ahead of demonstrated demand, and a shortfall in revenue from AI services would leave owners with underused assets and lenders with weaker security. Sales are concentrated among a few very large customers, some of whom also invest in or lend to their suppliers. Supply of electricity, chips and skilled labor can delay projects. Rapid technical change can make equipment less valuable sooner than planned. Past build-outs of railways, electricity networks and telecommunications left useful assets behind but produced poor returns for many of the original investors, a comparison often made in discussions of a possible bubble.

Last updated 16 Nov 2026