This page is a central point for topics related to reducing the carbon footprint of large-scale HPC and AI infrastructure.
Scale
- “One gigawatt is enough energy to power about 750,000 homes.”1
- “Vantage has started to see deals with tenants in the 100- to 500-megawatt range “that are specifically in support of AI on a dedicated basis for customers,” he said.”2
Impact of AI
OpenAI and Google have both converged on a single ChatGPT-like query consuming as much electricity as streaming 8-10 seconds of Netflix,3 or about 0.0003 kWh. The water consumption is less straightforward.
Locations
I have been cataloging major AI datacenter buildout projects in AI datacenters.
- “Major data-center markets include Northern Virginia, Atlanta, Phoenix and Silicon Valley. But increasingly, development is spreading to secondary markets, including northern Indiana, Idaho, Arkansas and Kansas, according to CBRE.”1
- “there are still many places, including a swath from Pennsylvania to Illinois, with an oversupply of power.”4
- See https://www.theregister.com/2024/08/21/dc_na_boom/ (or really North America Data Center Trends H1 2024 | CBRE)
- CoreWeave has 120 MW being installed near Richmond, VA5
- xAI’s Colossus cluster has a significant carbon footprint in the Memphis area.
Government concerns
Players
- Silver Lake, Vantage, DigitalBridge2
Energy sources
- Nuclear power is a zero-carbon source of energy that provides the base load power required by large AI data centers.
- Natural gas-fired combined cycle is commonly used to power data centers because it is inexpensive and efficient, but it is not clean.