Inc. Magazine
Smarter AI use could slash energy demands by up to 90 percent. New methods from UCL show breaking large language models into smaller tools, shortening prompts, and reducing precision can cut power use sharply without losing accuracy. As AI grows, energy efficiency is crucial for both business and the planet.
3 Innovations That Can Dramatically Reduce AI Energy Costs
Author: Bruce Crumley
New methods for shrinking the large language models (LLM) that power most AI apps can slash their energy consumption by up to 90 percent, according to a new study released this month. That’s a cost-saving opportunity businesses should jump on. Energy prices are projected to skyrocket under Trump’s new tax bill. And though AI is already an insatiable glutton for power, it’s only going to get worse.
The report—which was released by the United Nations Educational, Scientific and Cultural Organization (UNESCO) and uses research by partners at University College London’s (UCL) Computer Science department—identifies three main methods for shrinking the size and operating scope of LLMs.
Whether you’re building your own LLM, shopping around for one, or working to help tech giants cut costs, these findings are worth keeping in mind.
Why it matters
Reducing AI’s energy appetite will be essential to its future cost and environmental viability, particularly as more businesses integrate AI to improve productivity by automating myriad tasks employees currently handle.
The International Energy Agency expects AI electricity consumption to place 160 percent more demand on already struggling grids by 2030. Meanwhile, an analysis last year by Berkeley Lab calculated that AI platforms accounted for just 2 percent of total U.S. energy consumption in 2017, and 18 percent of the share in 2023. That share may reach 27 percent by 2028.
The reason for that growth, per data cited recently by the World Economic Forum, is that “the computational power required for sustaining AI’s rise is doubling roughly every 100 days.” It added the popular ChatGPT app consumes 33 times more energy to execute a command than other software or search engines—an excess that adds up fast with an estimated billion people using it each day.
The increased carbon emissions that generating all that extra electricity produces is just one negative result of AI’s appetite. Multiplying media reports have also warned of data centers recently built by Meta, Google, Microsoft, and other tech giants sucking up nearly half a million gallons of water for cooling systems each day—occasionally exhausting aquifers surrounding residents rely on, or leaving remaining water too sediment-filled to consume.
What to do about it
So how can enterprising tech businesses help sector giants combat those enormous challenges? By using UCL’s research discovery to break LLMs into smaller pieces that radically diminish AI’s energy consumption.
“Our research shows that there are relatively simple steps we can take to drastically reduce the energy and resource demands of generative AI, without sacrificing accuracy and without inventing entirely new solutions,” noted Ivana Drobnjak, a professor of computational healthcare at UCL and co-author of its report.
Enterprising geeks willing to shift their tech business activity to help sector giants reduce their AI’s energy use can find UCL’s full specifics here and here. Awaiting that deeper dive, the three main methods researchers suggest they do that are:
- Divide the LLMs that most AI platforms are built on to answer the widest variety of questions in a single app. That breaks them into smaller models dedicated to specific tasks like research, transcribing, translating, and others. Not only were the platforms UCL researchers parsed out just as accurate in their performance as the bigger systems they came from. They also consumed up to 90 percent less energy by focusing on a narrower range of information and functions. Meanwhile, they could also be packaged alongside other individualized apps to operate in tandem when called upon, offering a similar level of convenience that integrated LLMs provide.
- Reduce the number of words used in both prompts and answers. Researchers say that generates shorter but equally accurate responses that LLMs do. Less verbose inputs require AI to analyze smaller volumes of the vast resources they need to draw on to produce correct, yet less wordy replies. And by making ChatGPT less chatty, UCL said, energy use is typically cut 50 percent per task.
- Scale back the ambition and complexity of AI platforms. Doing so leaves apps capable of producing the accurate answers or task performance most individual and business users need, without providing the deeper expertise many models bake into them. That process starts with quantization, which reduces decimal places in making calculations. Quantization or similar data methods aim to produce a rounding-off effect that enables faster replies that are somewhat less minute in their precision. Despite that, answers maintain at least least 97 percent accuracy rates compared LLM apps, while reducing energy use by 44 percent.
The UNESCO report makes it clear that along with policy makers and big tech companies developing AI, smaller sector businesses can also play a key role in using UCL’s methods to reduce app energy use, and limiting the environmental impacts of additional electrical production demand.
Entrepreneurs can also be vocal in raising awareness in the public and among companies adopting AI, reminding users that switching to sustainable versions of the tech may be essential to avoiding potential energy overload and increased environmental damage.
“Though some AI platforms are already exploring and implementing solutions such as the ones we propose, there are many others besides the three that we looked at,” Drobnjak said, adding that work by tech giants and startups alike can help accelerate the pivot from LLMs to sustainable smaller models. “Wholesale adoption of energy-saving measures as standard would have the greatest impact.”
Credits: TCA, LLC.