How Many Data Centers Do Amazon, Google, Microsoft, Meta and Tesla Have — and What Do They Cost to Run?

Every search, video, chatbot answer and online order runs in a data center. The biggest tech companies now spend more on data centers than almost anything else — together roughly $700 billion planned for 2026. How many do they have, and what does it cost to keep them running?

Who owns how many

Exact counts are hard to pin down: companies publish “regions” and “zones” rather than buildings, and each zone can contain one or several data centers.

CompanyWhat they publish2026 capital spending (mostly data centers)
Amazon (AWS)39 regions, about 123 availability zones; leaked documents reported by Bloomberg suggested 900+ facilities including leased and edge sites≈ $220 billion
Microsoft (Azure)70+ regions; Microsoft has cited 400–500+ data centers≈ $170–190 billion
Google43 cloud regions, 130 zones; no official building count$195–205 billion
MetaLarge owned campuses for its own apps (it doesn’t sell cloud), plus leased and partner-built capacity$135–145 billion
TeslaThe Cortex AI training cluster at Giga Texas — about 50,000 Nvidia H100 GPUs at first, planned to exceed 100,000. Tesla shut down its own Dojo supercomputer in 2025Far smaller

Spending figures are 2026 guidance reported by the companies in mid-to-late 2026; each company defines capital spending slightly differently. The large “Colossus” AI data center in Memphis belongs to xAI, a separate Elon Musk company, not Tesla.

How much electricity data centers use

According to Lawrence Berkeley National Laboratory’s report for the U.S. Department of Energy, U.S. data centers used about 176 terawatt-hours in 2023 — about 4.4% of all U.S. electricity, up from 58 TWh in 2014. By 2028 that could reach 325–580 TWh, mainly because of AI.

AI hardware is the reason. An Nvidia DGX H100 server with 8 GPUs draws up to 10.2 kW. A newer GB200 NVL72 rack, with 72 GPUs, draws about 120 kW and needs liquid cooling.

PUE: how much power goes to the computers

Power usage effectiveness (PUE) is total facility power divided by the power used by the IT equipment. A PUE of 1.5 means that for every 1 kW of computing, another 0.5 kW goes to cooling, power conversion and lighting.

  • Industry average: about 1.54 — unchanged for six years, according to the Uptime Institute’s 2025 global survey.
  • Google’s fleet: about 1.09 in 2024 — less than 10% overhead.
  • A typical office server room cooled by ordinary air conditioning is far less efficient.

What it costs to run: a worked example

Take a cluster like Tesla’s Cortex at launch: 50,000 H100-class GPUs, running at 80% of maximum power on average, in a facility with a PUE of 1.3, paying the 2025 U.S. average industrial electricity price of 8.62¢ per kWh. The data center operating cost calculator gives:

  • Average power: about 66 MW for the whole facility
  • Electricity: about 581 GWh a year — the same as roughly 55,000 U.S. homes
  • Electricity bill: about $50 million a year, or about $5,700 an hour — around $1,000 per GPU per year

Scale that to a 100,000-GPU cluster of newer GB200-class chips at a PUE of 1.15 and the bill is about $116 million a year for 1.3 TWh — the electricity of nearly 128,000 homes.

Where the rest of the money goes

  • Hardware: for AI data centers the biggest cost is usually the chips themselves, replaced every few years as faster generations arrive. Spread over their working life, they typically cost far more per year than the electricity.
  • Cooling and water: many sites use evaporative cooling, which trades electricity for water — see how much water data centers use.
  • Power infrastructure: substations, backup generators and batteries (UPS), and increasingly companies’ own power deals.
  • Staff, security, maintenance, insurance and property taxes.

Why it matters to everyone else

Data centers that need hundreds of megawatts compete for grid capacity, can push up local electricity prices and require new power plants and transmission lines. That’s why efficiency (low PUE), off-peak tricks like ice thermal storage, and where data centers are built have become public-policy questions.

Running-cost examples use maximum system power scaled by average load and exclude construction, demand charges and taxes. Real bills depend on local power contracts.

Frequently asked questions

How many data centers does Amazon have?

AWS publishes 39 regions with about 123 availability zones, each containing one or more data centers. Leaked internal documents reported by Bloomberg suggested more than 900 facilities including leased and edge sites; Amazon doesn’t publish an official building count.

Does Tesla have its own data center?

Yes. Tesla’s Cortex AI training cluster at Giga Texas launched with about 50,000 Nvidia H100 GPUs and is planned to exceed 100,000. Tesla shut down its custom Dojo supercomputer in 2025. The Colossus data center in Memphis belongs to xAI, not Tesla.

How much electricity do data centers use?

U.S. data centers used about 176 terawatt-hours in 2023, about 4.4% of U.S. electricity, according to Lawrence Berkeley National Laboratory. That could rise to 325–580 TWh by 2028.

How much does it cost to run an AI data center?

For 50,000 H100-class GPUs at 80% average load, a PUE of 1.3 and 8.62¢ per kWh, electricity alone is about $50 million a year — around $1,000 per GPU. The hardware, replaced every few years, usually costs more per year than the power.

What is a good PUE for a data center?

The industry average is about 1.54 (Uptime Institute, 2025). The most efficient hyperscale operators reach about 1.1 — Google’s fleet averaged 1.09 in 2024.

Sources

Written by Pramod Parnati, Founder, Calculator Suite. Published November 19, 2026. Found a mistake? Every number in this guide can be checked with the linked calculators — and we’ll fix anything that’s wrong.

More guides