Does ChatGPT use a data center? Yes. Here's where, and what runs without one.

When you type a question into ChatGPT, nothing on your phone answers it. Your words cross the internet to a data center, usually in a county you've never visited, where racks of accelerator chips work out the reply. Then the answer travels back. So yes, ChatGPT uses a data center for every prompt. Some AI runs without one, and that's at the end of this page.

The short answer

ChatGPT, Claude and Gemini all run in data centers. The apps on your devices are remote controls. These models are too big to run on a phone, so the work happens on special hardware in a building the provider owns or rents. You send text and their servers send text back. With no connection, the chatbot can't answer, which is why "is ChatGPT down?" trends whenever one of those buildings has a problem.

Where ChatGPT’s data centers are

As of September 2026, OpenAI’s flagship data center is the Stargate campus in Abilene, Texas, which runs on Oracle’s cloud (OpenAI). OpenAI has announced more sites in Texas, New Mexico, Wisconsin and Ohio (OpenAI). Another, near Saline, Michigan, is being built now (Oracle). In August 2026, OpenAI agreed to secure about 8 gigawatts at a planned campus in Pike County, Ohio (OpenAI).

Microsoft is still OpenAI’s main cloud partner (Microsoft). It says its Fairwater AI data centers in Atlanta and Wisconsin are built to support OpenAI, among others (Microsoft). OpenAI’s list of cloud suppliers for business customers names Microsoft, Oracle, CoreWeave, Google, Amazon and Cerebras, in the U.S. and about two dozen other countries (OpenAI).

Abroad, a Stargate cluster is being built in Abu Dhabi, with its first 200 megawatts due in 2026 (The National). In India, OpenAI will start with 100 megawatts at Tata data centers (OpenAI). OpenAI paused its UK plan in April 2026, citing energy costs and regulation (The National). Microsoft took over a Norway site OpenAI had planned to use (DCD).

The path a single prompt takes

It feels instant, so it's easy to think the answer comes from nearby. It doesn't. Here's the round trip for one message:

1. Your device. You hit send. The app packages your text and ships it off; the device itself does almost no work.
2. The network. Your prompt hops across the internet (your router, your ISP, the backbone) to the provider's front door, often hundreds of miles away.
3. The load balancer. At the data center, a traffic cop decides which server should take your request, steering it toward hardware that's free.
4. The GPU rack. This is where the answer is made. A row of accelerators holds the model's weights and runs the math that turns your words into the next words. It's the step that costs the most money and power.
5. The trip back. The generated answer retraces the route, server to network to your screen, and streams in token by token, which is why you watch it type.

That's five stops, and four of them are out of your hands. Repeat it for hundreds of millions of prompts a day and you can see why the buildings keep getting bigger.

What that building costs

A data center sits on land, draws power from the grid, and runs water through cooling systems so thousands of chips don't overheat. Those costs fall on the places that host them, and more of those places are pushing back.

Water is what people noticed first. The Guardian reported on June 7, 2026 that data centers had consumed 264 billion gallons of water as drought gripped nearly 63% of the US, and a day later, on June 8, that a majority of the country's new AI data centers are slated to be built on drought-hit land. On June 2, one headline said Americans don't quite know how to fight AI, so they're fighting data centers instead. Erin Brockovich launched a national tracking map for them in late May.

The numbers behind the worry are public. UC Riverside's "Making AI Less Thirsty" research estimated that training GPT-3 alone evaporated roughly 700,000 liters of water, and that a stretch of a few dozen queries can run through about a 500ml bottle. Google's own 2024 Environmental Report put the company's 2023 water use north of 5 billion gallons. The IEA's "Energy and AI" report projects that data-center electricity demand will more than double by 2030.

Training a model gets the headlines, but the steady cost is inference, the everyday work of answering prompts. MIT Technology Review reported in 2025 that inference is about 80–90% of AI compute. So the round trip above is where most of the demand comes from, one question at a time.

The version that never leaves your desk

A model can also run on the chip already inside your computer. There's no round trip: your own hardware answers the prompt, the way a video editor renders a clip without a server.

That's what Outlier is: a Mac app that keeps the model files on your disk and runs them on your Apple Silicon chip. It's one signed download, with no account to create and nothing to set up in Terminal. Nothing leaves your Mac, so it works with the Wi-Fi off, on a plane for example. Each answer uses a little of the power your Mac already draws.

The trade-off: a laptop won't match a rack of accelerators on speed, and the strongest models still run in those buildings. For most everyday work, like drafting an email or fixing code, a local model is enough, and the answer never leaves your desk. Outlier's patent-pending paged engine can even run a model bigger than the Mac's memory: its 397-billion-parameter Plus model answers on a 64 GB Mac, slowly, at about 1.6 tokens a second.

Frequently asked questions