What is an AI data center?

How an AI data center differs

A conventional data center is a building full of general-purpose servers, each drawing a modest amount of power and cooled by air. An AI data center is organized around clusters of accelerators, GPUs or custom chips, linked by very fast networks so that thousands of them can work on one model at once. Packing those chips together is what makes them useful, and it is also what makes the building different: each rack draws many times the power of a conventional one, and the heat is too concentrated for air alone, so liquid is piped to the chips.

The scale has changed too. Data centers used to be described in megawatts; AI campuses are now described in gigawatts. Nvidia’s latest 10-Q describes guarantees for leases covering about 4.25 gigawatts of capacity at a single campus in Ohio, and AMD’s agreements with OpenAI and Meta each cover up to 6 gigawatts of its GPUs.

• Nvidia: The OpenAI campus Nvidia guarantees

  “In August 2026, we entered into guarantees, capped at a total of $105 billion, to provide credit support on a land, power, and shell buildout with affiliates of SB Energy Corp. (SB Energy) on behalf of a customer, an affiliate of OpenAI Group PBC (OpenAI), related to leases for approximately 4.25 gigawatts of IT load in the aggregate at SB Energy’s PORTS Technology Campus in Pike County, Ohio.”

  Nvidia 10-Q, filed August 26, 2026

• AMD: AMD’s OpenAI and Meta agreements

  “In October 2025 and February 2026, we entered into multi-year agreements with OpenAI OpCo, LLC (OpenAI) and Meta Platforms, Inc. (Meta), respectively, under which each customer intends to deploy up to 6 gigawatts of AMD data center GPUs, with the first gigawatt of each deployment powered by our AMD Instinct MI450 series products.”

  AMD 10-Q, filed August 5, 2026

What goes into one, and who sells it

Follow the bill of materials and you have the AI stack. The accelerators come from Nvidia, AMD or custom-chip designers such as Broadcom; the high-bandwidth memory beside them from SK Hynix, Samsung and Micron; the servers from Dell, Supermicro and HPE; the networking from Arista, Cisco and the optics makers. Under all of it, the chips are made by TSMC on machines from ASML, Applied Materials and Lam Research.

The building itself is a layer of its own. Our power, cooling and data-center layer covers 16 companies with $229B of revenue over their latest four quarters, +37% in two years. The largest: GE Vernova (gas turbines, grid equipment and electrification); Quanta Services (builds transmission lines and grid connections); Eaton (electrical equipment from grid to rack); Constellation (nuclear power sold under long contracts to data centers); EMCOR (electrical and mechanical construction inside data centers); Vistra (nuclear and gas generation); Vertiv (power distribution and liquid cooling for racks); Equinix (colocation data centers). See the power and data-center layer.

How much is being spent

The five largest buyers, Microsoft, Amazon, Alphabet, Meta and Oracle, spent $566B on capital projects over their latest four quarters, most of it on AI data centers and the equipment inside them. That is 3.1 times what they spent two years earlier, and it now takes 97% of their combined operating cash flow.

Capex is only part of it. Many AI data centers are leased rather than owned, and the leases don’t appear in capex or on the balance sheet until they begin. Microsoft $329B, Meta $356B, Oracle $260B, Alphabet $131B, CoreWeave $35.5B: these companies disclose leases, mostly for data centers, that have been signed but not yet started. Meta’s describes what they cover.

• Meta: Leases signed, not yet started

  “These lease obligations were approximately $355.73 billion, consisting of data centers, energy-related infrastructure, colocations, and certain network infrastructure, which will commence during the remainder of 2026 through 2036 with lease terms ranging from greater than one year to 30 years.”

  Meta 10-Q, filed October 22, 2026

Power is the constraint

Chips can be ordered in months; a grid connection or a new power plant takes years. That is why power has become the binding constraint on AI data centers, and why the buyers now sign long contracts directly with generators. Microsoft has a 20-year agreement supporting the restart of a Constellation nuclear plant; Meta has 20-year agreements with Vistra for 2,609 megawatts of nuclear power.

For investors this makes the power layer the slowest-moving part of the stack and in some ways the most predictable: backlogs at equipment makers such as GE Vernova run for years, and generators sign contracts lasting decades. It also means timing risk: equipment ordered today is a bet on data centers that open years from now. See GE Vernova, Vertiv and Constellation.

• Constellation: 20-year PPA for a restarted nuclear plant

  “The restart is supported by a 20-year PPA with Microsoft to purchase the output generated from the renewed plant.”

  Constellation 10-Q, filed August 6, 2026

• Vistra: 2,609 MW of nuclear power under 20-year PPAs

  “In January 2026, Vistra announced it had entered into 20-year PPAs with Meta, pursuant to which the Company has agreed to supply Meta with a total of 2,609 MW of carbon-free power and capacity from the Company's PJM nuclear power plants as follows: 1,268 MW of energy and capacity from Perry and 908 MW of energy and capacity from Davis-Besse; and 213 MW of uprate energy and capacity from Perry, 80 MW of uprate energy and capacity from Davis-Besse, and 140 MW of uprate energy and capacity from Beaver Valley.”

  Vistra 10-Q, filed August 10, 2026

Who owns AI data centers

Three kinds of company own them. The hyperscalers build their own campuses and rent out the capacity through their clouds. Neoclouds such as CoreWeave rent GPU capacity in data centers they lease from developers. And colocation and wholesale landlords such as Equinix and Digital Realty, or developers such as Applied Digital, build the shell, power and cooling and lease it to others. See hyperscaler and neocloud.

The choice between owning and leasing matters to investors because it changes where the cost appears. A hyperscaler that builds shows the spending as capex and depreciates it over the building’s and the servers’ lives; one that leases shows rent over the lease term and discloses the commitment in the notes. The economics can be similar, but the reported numbers look very different.

Ownership increasingly runs through joint ventures and leases rather than balance sheets. Meta holds two large campuses in ventures it doesn’t consolidate, backed by residual value guarantees. That keeps assets off the balance sheet, but the obligations remain, and they are disclosed in the notes.

• Meta: $41.5B of residual value guarantees, $68.4B maximum exposure

  “The following table summarizes our maximum exposure to loss in the unconsolidated data center ventures (in millions): September 30, 2026 December 31, 2025 Carrying value of equity investments $4,732 $1,826 Lease commitments 18,476 12,311 Future funding commitments 3,703 3,724 Maximum residual value guarantees 41,493 28,091 Total $68,404 $45,952 Other In addition, we have other types of unconsolidated VIEs of which we are not the primary beneficiary.”

  Meta 10-Q, filed October 22, 2026

AI data center stocks: how to read them

Investors often group the companies above as “AI data center stocks”. The label covers very different businesses: an equipment maker paid on delivery, a construction firm paid by project, a landlord paid rent for decades, a generator paid per megawatt-hour. Their revenue arrives at different points in the build-out, so they rise and fall at different times.

Three numbers say most of it. Backlog or remaining performance obligations shows how much work is contracted ahead. Customer concentration shows how much depends on a few buyers. And capex against cash flow shows how much each is spending to keep up. The layer page puts them side by side, and is AI a bubble? tracks whether the buyers’ spending is running ahead of their revenue.

Related terms

More from the glossary: Hyperscaler, Neocloud, High-bandwidth memory (HBM), Remaining performance obligations (RPO), TPU vs GPU, CoWoS (advanced packaging), Vendor financing, and every other term you need to read AI companies’ filings.