A single query to a large language model consumes roughly ten times the energy of a standard Google search. Multiply that across billions of daily interactions and the numbers become staggering. Data centres already account for one to two per cent of global electricity consumption, with cryptocurrency mining adding another quarter on top.
In Ireland, where tech giants have concentrated their European operations, data centres now consume 17% of the country’s total electricity. And the problem extends beyond power. Cooling these servers requires enormous quantities of water, adding another environmental pressure point.
The question facing the technology sector is whether innovation can outpace consumption or whether the AI revolution will create an environmental burden that dwarfs its benefits. Microsoft has begun reactivating retired nuclear plants while Amazon and Google are signing long-term power purchase agreements for renewable energy measured in gigawatts. The scramble for clean electricity has become as competitive as the race for AI talent.
“If we want to really scale with AI, we need to solve the energy problem,” says Ami Badani, chief marketing officer at Arm, the chip architecture company whose designs power most of the world’s smartphones. “Today, it doesn’t really look like we’re in a horrible position but, with the scaling laws, if we don’t solve the energy problem, we’re going to be in a tremendously bad position over time.”
The arithmetic is daunting. Training GPT-3 required approximately 1 300 megawatt-hours. GPT-4 demanded roughly 50 times that amount. Each generation of models grows more capable but also more and more hungry for energy. Jensen Huang’s presentations at Nvidia showcase chips that are a thousand times more powerful than their predecessors from eight years ago, and the trajectory shows no signs of flattening.
The solution requires optimisation at every level
Badani argues that no single intervention will solve the problem. The answer lies in attacking energy consumption across the entire technology stack simultaneously.
“You need optimisations everywhere,” she says. “It’s not only a hardware problem and it’s not only a software problem. It’s also not only an energy problem in terms of looking at renewables. It’s across the entire spectrum.”
The tech giants are responding by taking unprecedented control of their hardware.
For years, companies such as Microsoft, Google, and Amazon focused on applications while treating chips as commodities. That approach no longer works. Achieving meaningful efficiency gains requires vertical integration, which means designing custom silicon tailored to specific workloads rather than relying on general-purpose processors.
Roughly 50% to 70% of current energy consumption in AI systems comes from specialised chips and the cooling required to keep data centre racks from overheating. Arm, which built its reputation on power-efficient chips for mobile phones, now works with Microsoft, Google, Amazon, and Nvidia to bring that efficiency to data centres. The company claims its architecture delivers more than 40% power efficiency gains compared to alternatives.
“Power efficiency has been in our DNA,” Badani says. “We started back when you had the Apple Newton in the 1990s. Our first application where Arm really came to life was in mobile phones, which were all battery powered. And now it’s coming more and more to the forefront in data centres because of AI.”
Open-source models are changing the efficiency equation
The release of DeepSeek in early 2025 sent shockwaves through the AI industry. A Chinese laboratory had produced a model rivalling the best Western systems at a fraction of the training cost and computational requirements. The achievement demonstrated that clever engineering could substitute for brute-force scaling, upending assumptions about the inevitability of ever-larger, ever-hungrier models.
Peter Sarlin, co-founder and corporate vice president at AMD Silo AI, sees this as a fundamental shift. His company, Europe’s largest private AI laboratory, has built all its foundation models using 100% hydroelectric power. The excess heat from its operations warms neighbouring buildings, creating a net positive environmental impact.
“If we think about where AI will create value, where we will eventually consume, compute, and eventually then consume energy, that’s where it will happen,” he says. “But you don’t always need a general-purpose model. For many use cases, you can actually do very well with a small, distilled model.”
The implications extend beyond individual companies. Open source functions as digital infrastructure, allowing businesses anywhere to build on existing models rather than training from scratch. A start-up in Johannesburg can fine-tune an existing model for local languages and contexts without replicating the massive energy expenditure of the original training run. This means that the democratisation of AI need not mean the democratisation of its environmental costs.
“I would treat open source as a kind of digital infrastructure driving sovereignty in every country,” Sarlin says. “It implies that any business in that specific region can tap into existing models and build on them.”
The energy supply side of the equation matters too
Vattenfall, the century-old Swedish energy company, powers data centres across Europe and has spent the past decade restructuring its portfolio to produce almost exclusively low-emission electricity. Chief executive Anna Borg sees AI’s energy demands as a challenge, but not an insurmountable one. The key lies in taking a long view rather than extrapolating from current technology.
“You cannot assume that technology will look the way it does today and then just stack it and say that’s the future,” she says. “There will be a lot of development in this value chain.”
Her company matches data centre demand with renewable supply around the clock, optimising between wind power when conditions favour it and hydroelectric when they do not. The excess heat generated by servers gets channelled into district heating systems, warming nearby homes and offices rather than dissipating into the atmosphere. What was once waste becomes a resource.
The relationship works both ways. AI helps energy companies operate more efficiently. Vattenfall uses machine learning to forecast solar radiation on specific streets and rooftops, enabling trades on intraday energy markets that would be impossible with human-speed analysis.
“We knew how to do that but we just couldn’t do it fast enough,” Borg says. “There isn’t time to communicate with the trader within those 15 minutes. So we communicate with an algorithm that does the trading, and then we just have someone monitoring that.”
The optimism may be premature
The technologists working on these problems express remarkable confidence. Badani, Sarlin, and Borg all rate themselves at ten out of ten on an optimism scale when asked about AI’s energy future. They point to innovation cycles, efficiency gains, and renewable energy expansion as reasons to believe the problem will solve itself.
“I believe that AI will be a force for good,” Badani says. “I’m not worried. I think we’ll solve a lot of these problems, and I’m glad that we as a global community are thinking about energy as a problem. Now that we have a microscope on the problem, I believe that we’ll find solutions.”
But that confidence sits uneasily alongside the scale of what lies ahead. The energy transition requires not just powering AI but electrifying transport, heating, and industrial processes currently running on fossil fuels. Data centres are competing for clean energy with every other sector trying to decarbonise. Nuclear plants take decades, wind and solar remain intermittent, and grid infrastructure needs massive upgrades to handle the load.
“My mood is actually not on an anxiety scale but rather on a focus scale,” Borg says. “And on the focus scale, I’m a ten out of ten, because I think we need a lot of focus on this. We have the technologies needed to solve this. We have the knowledge and the capital is willing to move in but we need to increase the pace.”
The urgency behind the innovation
Leah Seligmann has spent her career at the intersection of environmental advocacy and corporate transformation. Before leading The B Team, an organisation of pioneering chief executives (including Richard Branson) committed to sustainable business practices, she served as chief sustainability officer at NRG Energy, helping decarbonise one of America’s largest coal-fired power companies. Her message to technologists is direct: climate stability is not guaranteed, and innovation alone will not save us without intentional leadership.
“The urgency of this moment is that we’ve exceeded what was a planetary goal of 1.5 degrees,” she says, referring to the target set by the 2015 Paris Agreement to limit global warming and avoid the worst climate impacts. “Humanity has evolved and thrived in a pretty climate-stable world. As these emissions grow and these temperatures rise, we’re getting to instability, to an experiment that we have never lived through before.”
She points to Ikea as evidence that sustainability and profitability can coexist. Under chief executive Jesper Brodin, who chairs The B Team, the company reduced emissions by 30% while growing revenue by 24% since 2016.
“When you centre around doing less harm and doing more good, you attract people and connect with people,” Seligmann says. “You don’t do everything right, but you are able to really push forward.”
For technology leaders uncertain how climate fits their work, she offers a challenge: examine the entire lifecycle of your products, from manufacturing through use to disposal. That lens reveals not just climate risks but innovation opportunities and cost savings.
“The kids are pissed,” she says. “They do not feel like we are responding with the urgency we need to this crisis that they are inheriting. If you’re thinking about your future employees, your future investors, and your future customers, you need to be thinking about climate.”
Gallery
