Compute is the new steel

Every era runs on a material it can't make enough of. Ours is compute, and data centers are where it's made. Your phone is a window. When you ask an AI a question, the work happens in a data center: each chatbot query 'literally performs a supercomputer-class numerical computation', in the words of UChicago computer scientist Andrew Chien^[5]. Training a major model takes thousands of specialized processors running in parallel for weeks or months^[7].

The numbers today

Data centers used about 1.5% of the world's electricity in 2024^[2] and about 4.4% of US electricity in 2023^[1]. In the United States they account for nearly half of all electricity demand growth between now and 2030^[2]. A typical AI-focused data center uses as much electricity as 100,000 households, and the largest under construction will use 20 times that^[2]. We want these numbers to rise faster, with the power plants to match.

What 1,000x buys

• Science: the IEA says AI led to a 45,000-fold acceleration in mapping protein structures, and can dramatically speed up finding and testing new materials^[2].
• Energy: AI can cut grid outage times 30% to 50% and unlock up to 175 GW of transmission capacity on existing lines^[2].
• Next: models that design drugs, materials and reactors. Tutors for every kid. Robots that build housing. Simulations that make rockets cheap.

That's what going to the stars actually requires.

The plan

1. Build data centers.
2. Build power plants.
3. Build transmission.
4. Repeat until the stars.

Why so many data centers, all of a sudden?

• Because AI changed the math. Training compute for frontier AI models has grown about 5x a year since 2020, and the top models have grown about 10,000x since 2020^[4]. The computing power of AI chips sold each quarter is growing about 2.5x a year^[4]. Every one of those chips needs a building, power and cooling. That building is a data center.
• Here's the fun part: at 5x a year, 1,000x takes a little over four years^[4]. We already did 10,000x once. 1,000x isn't a fantasy, it's a schedule.

Isn't efficiency enough?

The best argument against more data centers comes from Berkeley researchers: smaller open models run about six months behind the frontier, and much of the push for giant facilities is business models, not physics^[6]. Fair. But chips have gotten about 34% more energy efficient every year since 2008^[4], and demand still climbed. Cheaper compute gets used for more things. Efficiency and more data centers aren't rivals. We want both: smarter models running on far more compute.

What will replace data centers?

Nothing soon. Some AI tasks are moving onto laptops and phones, and that's great^[6]. But training models and serving millions of people at once still takes warehouses of chips, power and cooling. Fiber moves data across the country in under 15 milliseconds^[5], so data centers can sit in industrial zones far from anyone's backyard instead of disappearing.

Who owns data centers?

A mix. Cloud and AI companies build and run their own giant sites, Meta's Newton County, Georgia facility for example^[3]. Colocation companies build data centers and rent space to other firms. Enterprises run their own, and edge sites sit close to users^[7]. They all need the same three things: land, power and permits.