Tackling the AI energy hog - from nuclear power to intelligent chips and smart data inputting

As pressure on electricity supplies grows, driven by rising use of AI, attention is turning to other ways to maximize the efficiency of AI systems and reduce their demand on our power supply. While much of the focus is on data center design and location, there are other measures that could reduce demand in the first place.

Nuclear is one of these, with various tech vendors signing deals for this carbon-free energy resource. Microsoft has signed a 20-year power purchase agreement (PPA) with Constellation Energy to supply its US data centers. This will involve re-starting the Three Mile Island Unit 1 nuclear power station and lead to an additional 835 megawatts (MW) of carbon-free energy on the grid.

Amazon has signed a similar PPA with Talen Energy to supply it with 1,920 MW of carbon-free electricity until 2042. This will support AI and other cloud technologies at Amazon’s data center campus in Pennsylvania. The firms also plan to explore building new Small Modular Reactors (SMRs) in the region.

Google, meanwhile, has signed a deal with Elementl Power, which will see the former provide early-stage capital to help the latter prepare three potential US sites for advanced nuclear projects, each aiming for at least 600 megawatts of capacity.

These nuclear power deals are an important part of supporting the rising demand for AI, but a lot more energy is still needed. According to Goldman Sachs, 85-90 gigawatts (GW) of new nuclear capacity would be needed to meet all of the data center power demand growth expected by 2030. However, it expects that well less than 10% will be available by the end of the decade.

Small Modular Reactors

Bentley Systems’ Chief Value Officer Dave Philp is a critical infrastructure expert, whose current focus is the upcoming nuclear builds the UK government is undertaking, including the cheaper Small Modular Reactors. He notes that governments favoring a digital-first approach and wanting to gain the benefits of AI will lead to more energy-intensive data centers. Philp says:

An analog power grid is going to struggle to cope with 21^st-century demands. We need to think about different ways of doing it. We're seeing already that there's going to be a grid strain and the existing UK National Grid, we know it already faces capacity challenges, especially in key data center areas. We know that if you add new AI-driven loads, it's just going to exasperate the problem. I don't think we are there yet in terms of what reliable 24/7, 365 reliable base load looks like yet.

Reaching a reliable AI energy supply requires a different approach that takes account of not only the rise in demand for power, but also new, resilient ways to create a clean base load that can consistently deliver. Philp explains:

It's not just, can we create more resilience in the grid, can we create more demand, but can we do it in a way that's going to be cleaner and actually think about some of the things we've got just now in terms of the journey to Net Zero.

This is where small modular reactors come into play, offering a clean energy supply, which can help to cover the base load required for AI data centers, and compared to traditional nuclear, are much quicker to build. Philp notes:

It doesn't require as much lead-in as its bigger counterparts. Plus, the good thing is you can build your data center off grid and co-locate your small modular reactor site right next to it. So with the co-location of large data center campuses or AI hubs, it's going to create great opportunities for dedicated power, so small modular reactors can reduce strain on the grid as well. Co-locating the small modular reactors next to the data center is also going to reduce transmission losses, as power generated is closer at the point of use.

These campuses are generally located outside of big cities, giving an opportunity to make use of more regional and rural areas, and brownfield sites, which helps create jobs and area regeneration.

Energy-efficient chips

Another option is tackling energy use via the hardware element of AI technology. proteanTecs has developed a runtime monitoring solution for advanced semiconductor chips, which it claims offers 14% in power savings without sacrificing performance by eliminating the energy waste baked into traditional design. According to Noam Brousard, proteanTecs VP, Solutions Engineering, the shift to generative AI has outpaced the infrastructure it runs on, specifically the power and reliability of semiconductors. He notes:

Training just one large AI model can consume up to 10 times more electricity than a Google Search and a 16,384-GPU run experienced hardware failures every three hours without the cause being apparent. As gen AI workloads surge, we are burning through chips as they strain under test-driven model assumptions, with worst-case design guard bands and inefficiencies that quietly bleed energy across data centers and eat up precious resources.

By embedding intelligent monitoring agents into the chips themselves during the design phase, it’s possible to avoid energy waste from inefficient workloads and hardware failures. Brousard adds:

These agents generate new, predictive, previously inaccessible data that help you optimize AI model execution, enhance power efficiency, and prevent chip failures, while the systems are running.

AI star rating

Providing transparency over the energy consumption of different AI models is another way to tackle the problem. Earlier this year, Salesforce launched the AI Energy Score, an AI development benchmarking tool designed in collaboration with Hugging Face, Cohere and Carnegie Mellon University.

The catalyst for creating the benchmark was to encourage uptake of more efficient AI models, leading to the development of more sustainable technologies. Salesforce likens it to the Energy Star label for appliances and electronics, with the intention it will establish a clear and trusted benchmark for AI model sustainability.

Once an AI model has been run through the benchmarking tool, an AI Energy Score Label is generated listing details about the test and how it performed, including a rating of one to five stars, with five stars indicating the highest efficiency. This enables developers to identify and choose more sustainable models.

Superior data

Ensuring that AI systems are fed high-quality data can go some way to reducing energy requirements. Poor-quality or biased data can lead to inefficient models that require repeated training, wasting time and energy; well-structured, unbiased data, enriched with trusted metadata, lets AI algorithms operate faster and more efficiently. Levent Ergin, Global Chief Strategist for Climate, Sustainability & Gen AI at Informatica, notes:

This can reduce the computational overhead and energy consumption associated with developing and maintaining AI systems.

Training AI on specific data tailored to enterprise use cases is another way to reduce energy demands. An AI model for financial forecasting, trained on sector-specific banking data, for example, uses less energy for relevant tasks than a general model, as it avoids unnecessary computations. Ergin adds:

This approach is vital for enterprises, where proprietary data can optimise AI for industry-specific needs, lowering operational costs and encouraging adoption. The International Energy Agency notes AI's potential to optimise energy use in industries, further supporting this strategy.

My take

By combining some or all of the above measures with smarter data center design, there’s a chance that AI’s current growth trajectory will be sustainable. The alternative could see the planet’s dwindling resources funnelled away from communities in favor of powering AI systems, just so we can keep saying please and thanks to ChatGPT. That’s certainly not a future I want.