What Are Data Centers Doing? A Mind-Boggling Amount of Math.
nytimes.com
I wrote a piece last week in The New York Times visualizing the mind-boggling number of calculations needed to train the largest AI models.
“In 2026, GPT 6 Astra took nearly an octillion calculations to train — over 50 million times as many calculations as GPT 1 did. Even calling this figure astronomical understates its size: An octillion is a thousand times the number of stars in the known universe.”
An octillion is a thousand trillion trillion, or a number with 27 zeroes. As a comparison, I estimated that it would take an iPhone 17 Pro around 13 million years to do an octillion calculations.
Here’s a link to the piece:
We usually think of adding and multiplying numbers as abstract and immaterial operations. But it takes energy to do math. While working on this story, I started to see computation as a physical raw material that AI companies are consuming in order to obtain increasing amounts of intelligence, similar to how industries burn fossil fuels to obtain energy.
AI researchers have often referred to training AI in similar material terms: Ilya Sutskever, who co-founded OpenAI, has called data “the fossil fuel of AI” and Dario Amodei, the head of Anthropic, has written about how artificial intelligence emerges from a “big blob of compute”.
In other words, AI companies are using exponentially increasing amounts of “compute” to mine data for intelligence. And this ultimately boils down to using more and more energy, as I explain in the article.
Let me know what you think!
Read the article