What is the hidden cost of your GenAI research?

New Delhi:

A traditional Google search consumes an average of 0.0003 kilowatt-hours (KWh) of energy, according to a 2009 Google report. This energy can power a household light bulb (9 watts) for about 2 minutes. Google averaged about 8.5 billion searches per day in 2023, which translates to 2,550,000 KWh of electricity per day, which is about 2,000 times more electricity than an average Indian consumes (1,255 KWh) during a whole year.

In May this year, Google announced that the company will integrate AI into its search engine, which will be powered by its most powerful AI model – Gemini. According to Alex de Vries, a Dutch data scientist who spoke to The New Yorker on the subject, a single Google search with integrated AI will consume 10 times more energy (3 KWh) than a traditional Google search. This is 20,000 times more than the average Indian consumption in a year.

Alex de Vries, who is also the founder of Digiconomist – the organization responsible for the Bitcoin Energy Consumption Index – said that Google’s energy consumption will reach around 29 billion terawatt hours (TWh) per year if they proceed with the AI integration into your search. This value is equivalent to Ireland’s electricity consumption and more than Kenya’s.

Why are AI systems so power hungry?

AI systems require a lot of computing power to run complex algorithms to process a large corpus of increasingly large data. When you enter a prompt in ChatGPT, it is processed by the chatbot using its servers hosted in data centers. These hubs alone are responsible for 1-1.5 percent of all global electricity consumption, according to the International Energy Agency.

“I don’t think we’ve yet assessed the energy needs of this (AI) technology,” said Sam Altman, CEO of OpenAI, at a public event in Davos in January this year. Altman expressed the immediate need for a “breakthrough” technology like nuclear fusion to power AI operations, given the frontier technology’s growth projections. This is a clear indication that the industry leader in Large Language Model chatbots is looking for ways to sustain current levels of electricity consumption and safeguard its future demand.

AI Carbon Emissions – Setback to UN Net Zero Emissions Target for 2050?

Data centers, which power cloud computing as well as AI systems, produce 2.5 to 3.5 percent of global greenhouse gas emissions, according to The Shift Project, a French organization nonprofit that works to reduce energy dependence on fossil fuels. This equates to the same levels of greenhouse gas emissions as the entire aviation industry.

The power consumption and subsequent carbon footprint of different AI models vary significantly. For example, the BigScience BLOOM project, which is an AI model with 76 billion parameters (internal variables that a model learns during training) consumes 433 megawatt-hours (MWh) of electricity.

In contrast, OpenAI’s 2020 GPT-3, which had a comparable number of parameters at 175 billion, consumed 3 times more electricity at 1287 MWh, according to data from the Artificial Intelligence Index Report 2024, published by the Stanford Institute of Human-Centered Artificial Intelligence (IRAS).

CO2 equivalent emissions (in tonnes), that is, total greenhouse gas emissions expressed in terms of carbon dioxide, for BLOOM were 25 tonnes and for GPT-3 they were 20 times more, at 502 tons.

The report also states that there is a serious lack of transparency from AI creators about the environmental impact of their models, with most not publicly disclosing their carbon footprints.

Google’s recently released 2024 Sustainability Report also reflects the power hunger of nascent AI technology. The company has recorded an increase of almost 50% in carbon emissions in the last 5 years, due to the reinforcement of its new AI technologies. Microsoft’s 2024 Sustainability Report shows similar trends, with a 29% increase in CO2 emissions in 2023 compared to the previous year.

According to SemiAnalysis, an independent US-based AI research and analytics company, AI will drive the increase in data center electricity consumption to 4.5% of global power generation by 2030. Another estimate from the International Energy Agency Energy suggests that total data center electricity consumption could double from 2022 levels to 1000 TWh (equaling Japan’s current electricity consumption). India has around 138 data centers, with 45 more scheduled to be operational by the end of 2025. The United States has the highest number of data centers, with 2,701 of them.

Lawmakers are beginning to assess the situation. The European Union took note and adopted a new regulation in March this year. Under the scheme, all data center operators are required to report their energy and water consumption (used for cooling systems). They are also mandated to provide information on efficiency measures that are being implemented to ensure reduction.

In February, US Democrats introduced the Artificial Intelligence Environmental Impact Act of 2024. The law proposes the creation of an AI Environmental Impact Consortium comprised of experts, researchers, and industry stakeholders to address the environmental impact of AI. AI.