What is the environmental cost of the Swiss Alps supercomputer?

Several studies have shown that AI can analyze complex climate models and help predict heat waves, for example. It can also find ways to more efficiently distribute electricity to homes and industry, which will be crucial for the energy transition. The only downside: the development and operation of AI systems requires enormous resources, particularly water and electricity. Enough to create a considerable environmental footprint which could cancel out its advantages.

The International Energy Agency estimatesExternal link that AI technologies could consume ten times more electricity by 2026. Globally, this could represent levels comparable to what countries like Austria or Finland External linkconsume in a year. And this electricity often comes from fossil fuels.

Even Sam Altman, CEO of OpenAI — the company behind ChatGPT — told the World Economic Forum in Davos last year that an “energy breakthrough” would be needed to support the expansion of tools that generate text , images or sound, what we call generative AI. With nearly 4 billion monthly visits, ChatGPT already consumes as much energy as 33,000 households and 6% of the district’s waterExternal link of the United States where its data center is located to cool the servers.

At the same time, the Swiss polytechnics (ETH Zurich and ETH Lausanne) developed Alps, one of the most powerful supercomputers in the world, with the aim of supporting large language models (LLM), such as those on which based chatbots like ChatGPT. The goal is to create public and open-source AI systems for scientific research, particularly in the areas of climate and medicine. But at what cost to the environment?

SWI swissinfo.ch / Kai Reusser

A supercomputer powered by hydroelectricity

Alps’ performance currently places it seventh in the Top500External link global supercomputers. But its environmental impact remains lower compared to its American counterparts, according to Vlad Coroama, independent researcher and founder of the Roegen Center for Sustainability in Zurich.

Equipped with 10,000 graphics processors (GPU), Alps consumes on average seven megawatts of electricity, the equivalent of 10,000 Swiss households, or 0.09% of the country’s total consumption, according to calculations by Vlad Coroama .

Plus

Plus

How Swiss supercomputer Alps aims to advance AI

This content was published on

18 oct. 2024

The new Swiss supercomputer, Alps, must offer scientific projects the means to exploit the potential of AI.

read more How Swiss supercomputer Alps aims to advance AI

“Two wind turbines could cover this demand,” he explains. “This is not negligible, but it is justified if the goal is to advance basic scientific research.” For comparison, the El Capitan supercomputer, the world’s most powerful, consumes more than four times as much energy, or 29 megawatts, also from non-renewable sources.

To meet their growing energy needs, large private technology companies such as Amazon, Google and Microsoft are investing heavily in nuclear energy.

But Maria Grazia Giuffreda, deputy director of the Swiss Center for Scientific Computing (CSCS), believes that it is “unthinkable to continue to build data centers powered by nuclear energy everywhere”. His organization hosts Alps and other major IT systems. “We need to focus on clean, alternative energy sources,” she adds, noting that the publicly funded CSCS has taken this approach.

In fact, Alps relies exclusively on hydropower, meaning its direct CO2 emissions are minimal. While U.S.-based data centers often rely on diesel generators for emergency situations, CSCS has battery backup and a water reserve. This allows critical services like MeteoSwiss, the national weather forecast platform, to operate autonomously for around 20 minutes when supercomputers are down. CSCS also reuses the cooling water from its servers as hot water for the city of Lugano.

Vlad Coroama estimates that the Alps supercomputer represents only 0.013% of Switzerland’s annual CO2 emissions. A negligible quantity, according to him.

The environmental cost of supercomputers

Tech giants like Microsoft are investing billionsExternal link — $80 billion in 2025 aloneExternal link — in AI supercomputers such as “StargateExternal link».

It is estimated that operating this machine will require up to 5 gigawatts, the equivalent of the power output of five nuclear reactorsExternal link.

SWI swissinfo.ch / Kai Reusser

Vlad Coroama envisions and encourages a future focused on smaller, more specialized AI models, like those Swiss scientists intend to develop with the Alps supercomputer. These models require less energy.

But the environmental impact of AI is not limited to the energy used by data centers. “AI data centers are already having significant public health implications,” says Shaolei Ren, associate professor of engineering at the University of California, Riverside.

In a recent studyExternal linkwhich has not yet been peer-reviewed, Shaolei Ren and colleagues estimate that U.S. health costs from air pollution caused by AI technologies could reach $20 billion by 2030, equivalent to the public health burden of all vehicles in California. Air pollution contributes to diseases such as asthma, lung cancer and heart attacks.

The Amazon Web Services (AWS) data center in Ashburn, Virginia (United States). Northern Virginia’s data center cluster is the largest in the world.

EPA/JIM THE BAREFOOT

Sustainable solutions and benefits for society

CSCS has limited these risks by avoiding diesel generators, one of the main sources of air pollution. But Shaolei Ren believes that improvements are possible, because the environmental impact of supercomputers is enormous, even with the use of green energy and recycled cooling water. Indeed, the considerable energy requirements of supercomputers can strain local networks and their extensive cooling systems lead to water loss through evaporation. He compares the trajectory of AI to that of the automobile industry, which was initially very polluting before becoming more efficient.

“We can do better,” he says, suggesting practical steps like locating data centers in regions with abundant, cheap renewable energy and natural cooling resources, such as lakes and cold rivers, snow reserves and underground areas for geothermal cooling. According to Shaolei Ren, the Nordic countries would be ideal for outsourcing local IT capabilities, an option that CSCS Director Thomas Schulthess has already publicly supported.

Maria Grazia Giuffreda believes, however, that this possibility must be considered with caution. “We will have to ask ourselves whether it is really conceivable to move strategic infrastructure to places where we do not have full control,” she observes.

For Vlad Coroama, the success of the Swiss initiative will be measured above all by its societal impact.

“For large scientific and climate simulations, a supercomputer is essential,” he explains, pointing out that generative AI models such as ChatGPT or Google Gemini consume much greater amounts of resources than Alps, with advantages less tangible. “It’s a waste of energy to use these huge, very powerful templates for small tasks like writing emails.”

The Alps supercomputer could offer more, according to him. “Good science costs money, but it is worth it when it provides real value to society as a whole.”

Text reread and verified by Sabrina Weiss and Veronica De Vore, translation by Lucie Donzé