What is the Stargate Project, and why does AI need so many data centers?

Yesterday, January 21, 2025, President Donald Trump stood in the Roosevelt Room of the White House with three business leaders, SoftBank's Masayoshi Son, OpenAI's Sam Altman and Oracle's Larry Ellison, to announce Stargate: a new company that intends to build AI infrastructure in the United States for OpenAI.

The headline number is huge. The partners say they plan to invest as much as $500 billion within four years, with the first $100 billion to be deployed right away. Construction of the first sites has already started in Texas.

Most of what we call "AI" happens on websites and phone apps, so it is easy to forget that it runs on physical machines in very large buildings. Stargate is about those buildings.

Who is involved, and in what role

OpenAI's announcement, co-signed with SoftBank, spells out the structure:

An equity funder puts money in and owns a share of the company. The technology partners supply chips, chip designs, cloud services or know-how. Nvidia makes the graphics processors (GPUs) that do most of the work in training AI, and Arm designs processor technology used across the industry.

CNBC reports that Son had already pledged a four-year, $100 billion investment in the US when he visited the then president-elect at Mar-a-Lago in December. At yesterday's event, Trump suggested that the $500 billion is on top of that earlier promise. He also framed the project as a way to keep AI development in the United States, naming China as a major competitor.

How it works: why AI needs factories of computers

A data center is a building filled with racks of servers, the powerful computers that store and process data for websites and apps. An AI data center is packed with GPUs, chips that can do millions of small calculations at the same time, which is exactly what neural networks need.

AI uses these machines in two ways:

1. Training. To build a new model, a lab runs huge amounts of text, images and code through the network for weeks or months, adjusting billions of internal numbers until the model gets good at predicting and answering. The bigger the model and the more data, the more chips and time it takes.
2. Inference. Every time you send ChatGPT a question, a computer somewhere runs the trained model to produce your answer. With many millions of people asking questions every day, that adds up too, and newer reasoning models, which work through a problem step by step before answering, use more computation per answer.

An analogy helps. If a model is a recipe, training is developing and testing the recipe, and inference is cooking the dish for every customer who orders it. Stargate is about building the kitchens: the buildings, the ovens (chips), the power lines and the cooling that keep them running. That is why the announcement invites companies working on power, land, construction and equipment to get involved.

Microsoft's place in the picture

Microsoft has worked closely with OpenAI since 2019 and supplies computing power through Azure, Microsoft's cloud service. On the same day, Microsoft published a note explaining what changes:

• The core of the deal stays in place through 2030: Microsoft keeps access to OpenAI's models for products such as Copilot, revenue sharing continues, and the OpenAI API (the service developers use) remains exclusive to Azure.
• OpenAI has made a new, large commitment to buy Azure capacity.
• For new computing capacity, Microsoft moves from exclusivity to a right of first refusal: OpenAI must offer Microsoft the chance first, but may build elsewhere if Microsoft declines. Microsoft also approved OpenAI building extra capacity, mainly for research and training.

In plain terms, OpenAI stays close to Microsoft but gains room to build its own supply of computing power, and Stargate is the vehicle for that.

Why it matters, and what remains unknown

As of today, Stargate is a statement of intent backed by some of the largest companies in technology and finance. It signals that the leading labs expect demand for AI computing to keep growing for years, and that governments now see that computing power as a matter of economic and national strategy. The announcement claims the project will bring hundreds of thousands of jobs to the US and strengthen national security.

Several things are not yet public. The partners have not said how much each one is contributing, and OpenAI says definitive agreements are still being finalized. Beyond Texas, other sites are still being evaluated. The $500 billion is a plan spread over four years, not money already spent. It is also fair to note that this announcement arrives weeks after a Chinese lab, DeepSeek, reported training a strong model on a much smaller budget, so the right amount of computing for the next generation of models is an open debate.

What does Stargate change for people who use ChatGPT?

Nothing changes today. ChatGPT keeps running on the same computers it ran on yesterday, and the new data centers will take time to build and fill with chips.

Over the longer term, more computing capacity is what lets a company like OpenAI train larger models, offer heavier features such as reasoning to more people, and keep answers fast when demand spikes. Whether that leads to better or cheaper service for you will depend on decisions OpenAI has not announced.

For now, the most useful takeaway is understanding the machinery. Each prompt you type travels to a data center, runs on specialized chips and comes back as an answer. Stargate is a bet, measured in hundreds of billions of dollars, that people will want far more of those answers.