Large language models, a form of artificial intelligence many banks have deployed to do things like
"Today's large language models will never be green," said Seth Dobrin, founder of advisory firm Qantm AI and IBM's first global chief AI officer, in a
"If we get rid of those, is that green? What do we do with those? That contributes to the lack of sustainability of these AI systems," Dobrin said. There needs to be a way to recycle some of this hardware, he said.
Other experts agree that the energy-guzzling nature of large language models is a problem that needs to be addressed, but they see hope in the efforts tech companies and banks are making to reduce their carbon-spewing impact.
"Large language models are indeed very power-hungry, and this should be clearly a concern with regards to their environmental impact," said Javier Rodriguez Soler, global head of sustainability and corporate and investment banking at BBVA.
But he also pointed out that some providers, including OpenAI, are producing less energy-hungry versions of their models. OpenAI's GPT-4o reportedly matches the performance of GPT-4 with a more efficient architecture, leading to lower energy demands, as indicated by reduced operational costs, Soler said.
"We've seen a similar trend with the rest of the LLM providers with costs — which is a good proxy of energy consumption — decreasing by a factor of three or four in just a few months," he said.
Nvidia, Google, Meta Platforms and banks that use advanced AI are all taking steps to reduce the energy consumption of their AI projects. Whether these companies are doing enough to make this technology "green" is an open question.
Tech companies trying to make LLMs greener
Large language model providers OpenAI, Anthropic and Meta say they are trying to reduce their models' energy consumption — by shrinking the number of parameters they use and the amount of data they consume, and by making their models work more efficiently.
These kinds of reductions could continue almost indefinitely, Soler said.
"Consider that the human brain, arguably more powerful than the most advanced LLMs, operates on just 20 watts," he said.
The big tech companies like Microsoft, Google and Amazon.com that many banks use to host their large language models — sometimes called hyperscalers — all message eco-conscious efforts.
Microsoft is developing ways to quantify the energy use and carbon impact of AI while working to make large systems more efficient, in both training and application, a company spokesman said.
"We will continue to monitor our emissions, accelerate progress while increasing our use of clean energy to power data centers, purchasing renewable energy and other efforts to meet our sustainability goals of being carbon negative, water positive and zero waste by 2030," the Microsoft spokesman said.
Meta recently partnered with the geothermal energy startup Sage Geosystems. The two companies plan to start generating up to 150 megawatts of carbonless power, enough to support a large data center. Sage said it harvests energy from underground using repurposed oil drilling equipment.
Google is working on a geothermal project in Nevada with Fervo Energy. The two companies and NV Energy hope to generate 115 megawatts of power for Google.
Nvidia, the largest provider of the graphics processor units that are used to run large language models, has said it is working to reduce energy consumption in its chips. The company declined a request for an interview.
What banks are doing to minimize energy consumption
BBVA is in the early stages of exploring large language models and learning how to deploy them at scale. Although its current LLM energy consumption is limited, the bank is anticipating this will change in the near future and considering strategies to mitigate the impact.
One of its strategies is adhering to the principle of data minimization when building AI models.
"By using only the data that is absolutely necessary, we create smaller, more efficient models that require less computational power and consequently less energy," Soler said. This approach aligns with the tenets of the European Union's General Data Protection Rule and embodies an AI development best practice that models should be only as complex as required to achieve their objectives, Soler said.
"In instances where data minimization is not feasible, such as when deploying large language models, we focus on selecting the smallest effective model to minimize energy usage," Soler said. "For some applications we are witnessing two orders of magnitude — close to 100 times — energy reduction without sacrificing performance."
BBVA is also moving all its AI operations, so that they are conducted through hyperscalers that are committed to running on renewable energy sources.
The bank will continue to encourage the LLM providers it works with to explore advances in AI and hardware efficiency, Soler said.
Banks that have a presence in Europe must think more about these issues than U.S.-only companies do, due to rules like the EU's Sustainable Finance Disclosure Regulation and customers who are more focused on climate issues, said Harry Stahl, senior director, enterprise strategy and ESG strategy lead at FIS. He commented for this article as a member of ISITC, an industry organization of financial institutions and technology providers that work to improve the operational effectiveness of the financial industry.
U.S. banks mainly rely on their cloud and generative AI partners to help them keep the energy consumption of their large language models as low as possible.
Some banks are getting interested in small language models, which have fewer parameters and therefore a limited capacity to process and generate text compared with large language models.
"For a lot of the tasks that banks want to apply this technology to, they don't necessarily need the full knowledge base and inputs of a public-facing large language model," said Gilles Ubaghs, a strategic adviser at Datos Insights. "A smaller, more specialized model should in turn lead to lower energy efficiency."
Some U.S. banks, however, have ignored the climate effects of burning lots of electricity.
"Despite all the ESG messaging and public statements on things like climate change, most banks have not taken major steps like divesting from the fossil fuel industry," Ubaghs said. "They'll tout their in-office recycling programs but just ignore the energy consumption of generative AI."
But U.S. banks are starting to feel regulatory and legal pressure to become more carbon conscious.
Banks that do business in California and have annual revenues of more than $1 billion are subject to the state's Climate Corporate Data Accountability Act, which will require them, starting in 2026, to publicly disclose their annual greenhouse gas emissions and pay fees for their direct and indirect emissions.
A rule the Securities and Exchange Commission finalized this year would require some companies to disclose certain carbon emissions. This was put on hold earlier this year due to a court challenge.
"The broader sense from those who are looking at it from a sustainability office or role is, we have to think about this," Stahl said. "The question is, how fast do we have to move?"
The answer is: It depends on which AI model you're using and who the provider is.
Artificial intelligence models are energy hogs. Climate First Bank and UBS are among the very few trying to solve this problem.
Banks must think about where their technology providers are getting their power, Stahl said, and where their large language models are physically running — in their own data centers or in those hosted by third parties. This becomes part of third-party risk management and due diligence.
"If it's to a provider that is aggressively moving to reduce their carbon footprint and move to net zero, you could actually potentially simultaneously shift more into AI and reduce your carbon impact," he said. "But you need to look at the data. What's the current profile and the target profile of your providers?"
Large companies tend to license enterprise versions of large language models, to keep their data secure. But as they are thinking about risk management for a tech provider, they also need to think about its energy-consumption profile.
Banks can also think about whether they really need AI for some projects at all, Stahl said.
"There are organizations like the Green Software Foundation that are thinking about, how do you write software that in fact is designed to consume less power?" Stahl said. "And there are ways to organize to be more energy-efficient in how you build and run software, that are not going to move the overall economy massively, but are worth considering."
Some jobs don't really require AI at all; another technology, like robotics process automation, might work just as well or better.
"There are some places where AI is extra," Stahl said. "It's like the icing on cornflakes, you don't need it."
