Microsoft Goes Nuclear To Run Data Centers

Microsoft is hiring a project manager for nuclear technology as it looks to power its data centers running artificial intelligence (AI) with nuclear energy. Much more energy is needed to run AI-related applications compared with search.

AI infrastructures use a lot of energy to run, especially to support the advertising industry as machine learning and chatbots increase in popularity, and connected television (CTV) and video streaming are more in demand among consumers.

Training one AI model or chatbot can use more electricity than 100 U.S. homes in one year. One report estimates that training Chat-GPT uses 1.287 gigawatt hours of energy, roughly equivalent to the amount used by 120 American homes in a year.

Microsoft co-founder Bill Gates recently invested $4 billion in a project he calls Natrium, a reconditioned coal mine refitted as a nuclear energy plant located in Kemmerer, Wyoming. This is part of his vision for “advanced” energy. The Wyoming plant is scheduled to begin operation in 2030.



Reportedly, the plant will generate 345 megawatts of steady electric generation and includes a power storage component that allows it to “flex” up to 500 megawatts for short periods.

The job description explains the implementation of “a global Small Modular Reactor (SMR) and microreactor energy strategy” to support a statement from Satya Nadella, chairman and chief executive officer of Microsoft, that vows a commitment to helping brands and companies use its platform and tools to innovate in AI.

Concern is mounting over the environmental impact of carbon emissions from technology and generative AI's massive energy use. Academic analysis has estimated that training GPT-3 produced more than 550 tons of carbon dioxide and required 3.5 million liters of water.

Last year, California Governor Gavin Newsom stated plans to sign a sweeping climate bill that would mandate companies to disclose their carbon footprint. The bill is expected to bring light to how companies, brands and retailers doing business in California operate.

The Senate Bill 253, Climate Corporate Data Accountability Act would begin by requiring the state board, on or before January 1, 2025, to “develop and adopt regulations requiring specified partnerships, corporations, limited liability companies, and other business entities with total annual revenues in excess of $1,000,000,000 and that do business in California, defined as “reporting entities,” to publicly disclose to the emissions reporting organization, as defined, and obtain an assurance engagement on, starting in 2026.”

Data from Loop, the software platform for retail returns, found that 56% of consumers are concerned about the environmental impact of returns, while 35% say they have not returned an unwanted item because of environmental concerns.

Microsoft is shifting plans to use nuclear energy to power its AI data centers as it struggles with the fact that huge amount of energy will be needed to run models like ChatGPT.

In 2011, Time magazine reported that Google spent about 0.0003 kWh of energy on an average search query, translating to roughly 0.2g of carbon dioxide.

The act of searching the web 100 times as equivalent to drinking 1.5 tablespoons of orange juice, Google told Time.

When asked how much energy a large data center uses to run large language models for AI, a University of Washington assistant professor said "each processing unit can consume over 400 watts of power while operating. Typically, you need to consume a similar amount of power for cooling and power management as well. Overall, this can lead to up to 10 gigawatt-hour (GWh) power consumption to train a single large language model like ChatGPT-3. This is on average roughly equivalent to the yearly electricity consumption of over 1,000 U.S. households."

Next story loading loading..