Feature
Businesses and vendors alike have spent much of the past 18 months talking up what generative AI can do. What's been a smaller part of the conversation is what the development of generative AI is doing to the environment, particularly the amount of energy required to develop and maintain these models.
Generative AI and the Elephant in the Room
A growing body of academic papers and research notes are paying attention to the problem, although it's when people like Elon Musk speak that the issue seeps into the mainstream. At the recent Bosch Connected World 2024, Elon Musk was the first to raise the elephant in the room during an online Q+A. "I've never seen any technology advance faster than this,” he said. “The chip shortage may be behind us, but AI and EVs are expanding at such a rapacious rate that the world will face supply crunches in electricity and transformers next year.”
He added: “Then, the next shortage will be electricity. They will not be able to find enough electricity to run all the chips. I think next year, you will see they just cannot find enough electricity to run all the chips."
While it's always a good idea to take anything Musk says with a bucket of salt, in this case he might be on to something.
The topic also came up during the Davos Summit in January.
Fortune magazine reported Amazon CEO Andy Jassy saying there was currently only enough energy to keep LLMs running.
"These large language models are so power hungry, and there's just not enough energy right now,” he said. "So, we're going to have the dual challenge as a group to find a lot more energy to satisfy what people want to do and what we can get done for society with generative AI.”
Tackling this must be done in a renewable way to ensure that the development of generative AI is carbon neutral, he continued. “It can’t be going back to coal.”
While not directly related, he also posed one of the major questions facing organizations and their digital workplaces now: how they choose to spend their tech dollars.
"I see companies really battling with prioritization. Are they better off continuing with the modernization of their technology platform? Or should they put all their engineering resources into generative AI?"
He added that everyone will decide differently. But without a sound foundation for your technology infrastructure, "you are going to have a hard time being successful with generative AI.”
Related Article: Your Content Has a Carbon Footprint ... and It's a Problem
Demand for Data Centers Is Nearing Capacity
Powering the development of AI isn't just an environmental issue, it's about the future development of AI too. The biggest single question here is data centers and how to power them. The demand for data centers had already been high, but with generative AI, where data is the building block of LLMs — demand has never been greater.
“Demand for data centers has always been there, but it’s never been like this,” Pankaj Sharma, an executive vice president at Schneider Electric’s data center division told ArsTechnica.
By 2030, "we probably don’t have enough capacity available” to run all the required facilities, he continued. And he should know. His unit is working with chipmaker Nvidia to design data centers optimized for AI workloads.
The same article cited Appleby Strategy Group CTO Daniel Golding saying: "At some point the reality of the [electricity] grid is going to get in the way of AI.”
How Much Juice?
So what are we looking at in real terms? Shortly after the 2018 introduction of Google's BERT LLM, which then consisted of 213 million parameters, a group of researchers found that BERT emitted 280 metric tons of carbon emissions, which is the equivalent of five cars' emissions over their total lifetime.
In the five ensuing years, the problem has become decidedly more pronounced as the models have grown. For context: GPT-4 has 1.7 trillion parameters.
A 2022 study found that even using a renewable energy source for training a 2022-era LLM emits at least 25 metric tons of carbon equivalents. If you use carbon-intensive energy sources like coal and natural gas — which was the case for GPT-3 — this rises to 500 metric tons, the equivalent of a million miles driven by an average gasoline-powered car.
Due to their vastness and multitude of characteristics, LLMs also necessitate large amounts of data transportation and computation. Energy-hungry GPUs and CPUs are strained by this, which increases electrical demand, heat generation and calls for more cooling power, according to a recent report from Straits Research.
Water-based cooling techniques can contaminate water and damage ecosystems, adding another layer to the environmental impact math.
Related Article: Is Remote Work Good or Bad for the Environment?
A Growing Issue
The problem is only going to get worse. The data centers that support the LLMs devise most of their energy from fossil fuels, according to a recent article in The State of the Planet, even if there are current efforts to bring renewable energy into the equation.
The world’s data centers account for 2.5% to 3.7% of global greenhouse gas emissions, exceeding even those of the aviation industry. The daily carbon footprint of GPT-3 alone is estimated to be the equivalent of 8.4 tons of CO2 in a year.
Relief may be on the way in this area at least. According to the article, 40% of the energy use is related to the initial model training, while the other 60% is related to inference, or the process of providing answers to millions of queries a day.
Webinar
Sep
25
CMS Briefing: A Live Look at What’s Next in AI-Driven Platforms
Learn how leading organizations are using AI‑driven tools to publish faster, personalize smarter and stay secure.
Register
Webinar
On demand
Ready or Not: How Data-First Organizations Are Unlocking Agentforce Potential
Learn how to cut through the noise, activate Agentforce and build a Salesforce AI strategy that actually delivers.
Watch Now
Webinar
On demand
AI in Customer Service: Faster Resolutions, Happier Customers
Don’t let rising demand burn out your team. See how to build a smarter, more resilient support org.
Watch Now
Webinar
On demand
From Hype to High-Impact CX Strategies That Actually Scale
Turn buzzworthy AI and outsourcing trends into measurable CX wins with fresh 2025 data.
Watch Now
Webinar
On demand
Insights to Action Rethinking the Contact Center for Real Business Impact
Join our exclusive webinar to hear CX executives share their innovative strategies for transforming service delivery.
Watch Now
Webinar
On demand
Accelerating Healthcare Ops with AI
Sign up to see how your peers are putting AI to work — and making it stick.
Watch Now
The situation is urgent. Recent figures from the International Energy Agency state that electricity consumption from data centers, artificial intelligence (AI) and the cryptocurrency sector could double by 2026.
Data centers are significant drivers of growth in electricity demand in many regions. After globally consuming an estimated 460 terawatt-hours (one trillion watts of power used for one hour) in 2022, the total electricity consumption of data centers could reach more than 1000 TWh in 2026. "This demand is roughly equivalent to the electricity consumption of Japan,” the report reads.
How 'Green' Can a Vendor Be?
Despite heavily marketed green initiatives by many vendors, the figures are not encouraging.
- Microsoft used an additional 1.7 billion gallons of water in 2022 compared to 2021, a 34% year-over-year increase, according to the 2022 Environmental Sustainability Report published in May last year.
- Google used an additional 1.2 billion gallons of water in 2022 compared to 2021, which represents a 20% year-over-year increase. Ninety-three percent of the water Google used in 2022 was used to cool data centers.
- A ChatGPT interaction of between 10 and 50 queries (known as an inference) uses 500 ml of water, according to OECD estimates at the end of last November.
- In March of this year, the EU mandated annual water and energy use disclosures by data center operators beginning September 2024.
- Only a little over a third of data centers traced their water use in 2022, a 12% drop from 2021, according to the most recent Uptime Intelligence report, published in 2023.
The same Uptime report found that while many data center operators are monitoring their centers' energy and water usage habits, they are not monitoring the greenhouse gases they generate.
The executive summary of the 2024 report noted that the years leading up to 2030 will be difficult, with many organizations struggling to meet sustainability goals and reporting requirements.
“There are already signs that monitoring bodies are becoming stricter in assessing corporate sustainability. In August 2023, for example, the UN-backed Science Based Targets initiative (SBTi) removed Amazon’s operations (including Amazon Web Services) from its list of committed companies because Amazon had failed to validate its net-zero emissions target,” the report read.
Related Article: The Business Case for Sustainability: Driving Organizational and Environmental Impact
What Can Be Done?
Acknowledging that there's no turning back in terms of putting generative AI back in the stable, what can be done to mitigate the environmental impact of LLMs?
Demand-side management, user self-regulation and algorithmic advancements will all be necessary. It will involve a number of different strategies adopted by the vendors creating the models, data center providers and end users.
- Energy-efficient hardware: When compared to typical CPUs, the use of energy-efficient hardware, such as GPUs and TPUs, can cut energy consumption by up to 50%.
- Renewable energy: Using renewable energy to power computing resources can cut the carbon footprint of big language models by up to 98%.
- Carbon offsetting initiatives: These are aimed at offsetting greenhouse gas emissions that result from the development and application of big language models. This can assist in lowering the emissions connected to these models, even though it does not immediately lessen the impact on the environment.
- Effective model design and training: Methods such as distillation, quantization and model pruning can reduce the amount of computing power needed to train big language models by as much as 90%. This can lessen these models' negative environmental effects.
- Reusing and sharing pre-trained models: Reusing and sharing pre-trained models can save up to 80% of the total energy used when training multiple models. This strategy lessens its negative effects on the environment while encouraging cooperation within the AI community.