The growth of AI over the past few years has resulted in soaring power consumption. Technology cloud providers and AI companies are racing to build out power-hungry data centres to house their GPUs used for training AI models and inference.
Presently, AI consumes 4% of total US electricity consumption and will increase to 9.1% by 2030. To help meet this surging demand in the short-term power companies are building new and expanding existing natural gas-powered plants. Cheap natural gas from fracking and the surging energy demands have resulted in a new all-time high of planned gas-powered capacity in 2024.
Technology companies are responding to their surging power needs through various means:
Amazon
Amazon recently announced its purchase of a 960MW data centre from Talen Energy. The data centre is co-located at the Susquehanna Steam Electric Station in Salem Township, Pennsylvania. As part of the agreement, Amazon will purchase nuclear-generated electricity from Talen Energy.
Anthropic
Anthropic has made no direct investments or power purchase agreements.
OpenAI
Training ChatGPT-3 consumed 1,300 mWh of electricity, equivalent to the annual consumption of 130 US homes.
It’s estimated OpenAI is consuming more than 500,000 kWh per day to handle the 200 million requests per day.
OpenAI has made no direct investments or power purchase agreements, but Sam Altman has personally invested $375 million in Helion Energy. Helion Energy is a US-based nuclear fusion company and has agreed to sell power to Microsoft in the future.
Oracle
Oracle has proposed a new 800 MW data centre to house their largest Nvidea GPU cluster. The data centre will be operated by nuclear power from three Small Modular Reactors (SMRs) on site.
Microsoft
Microsoft just signed a 20-year deal to purchase electricity from Constellation Energy’s Three Mile Island nuclear plant. Constellation Energy will restart Unit 1 reactor at Three Mile Island which has been offline since 2019.
Separately, Microsoft also announced a partnership with BlackRock to invest up to $100 billion in AI data centres and energy infrastructure.
xAI
xAI installed at least 18 gas turbines at their Memphis, Tennessee data centre. The data centre requires 150 MW of energy, but the local utility (Memphis Light, Gas and Water and the Tennessee Valley Authority) has so far only been able to provide 50 MW. The remaining 100 MW is presumably being provided by the gas turbines which are operating without a permit.
Conclusion
AI is having an immediate and measurable impact on energy consumption and contributing to carbon emissions. Deep-pocketed players are building or buying nuclear power, but unfortunately others are resorting to carbon emitting power-generating sources. Governments, power generation companies, and AI companies need to work closely to de-couple the growth of AI from carbon emissions.