Efficient Chips Alone Will Not Solve AI’s Energy Problem

Artificial intelligence may seem intangible, but it relies on deeply physical infrastructure. Each query activates processors, memory, networks and cooling systems housed in data centres. All of this requires electricity, generates heat, consumes water and depends on materials whose extraction and manufacturing also leave an environmental footprint (International Energy Agency [IEA], 2025; Li et al., 2025).

In a recent energy consumption measurement published by Google, a median text prompt in Gemini consumed 0.24 watt-hours, generated 0.03 grams of CO₂ equivalent and used 0.26 millilitres of water. However, these results reflect a specific system, infrastructure, and methodology; consumption varies across models (Elsworth et al., 2025).

The difference between tasks can be enormous. Generating a short text response is not equivalent to producing an image, analyzing thousands of pages, creating a video or running an agent that performs numerous operations before responding. The International Energy Agency notes that certain reasoning tasks, video generation and autonomous systems can consume hundreds or thousands of times more energy than a simple text query (IEA, 2026).

AI’s footprint grows with demand

In 2024, data centres consumed approximately 415 terawatt-hours of electricity, around 1.5% of global electricity consumption. The International Energy Agency projects that this figure could reach approximately 945 TWh by 2030, slightly more than Japan’s current electricity consumption. AI will drive most of this growth, and demand from data centres specifically optimized for AI could increase that number exponentially (IEA, 2025).

Where that electricity comes from matters. A task processed on a grid powered mainly by hydroelectric, wind or nuclear energy will generally produce fewer operational emissions than the same task performed where coal or natural gas predominates. Renewable energy could meet close to half of the additional electricity required by data centres through 2030, but coal and gas could still provide more than 40% of that increase (IEA, 2025).

Water use is equally dependent on context. Data centres consume water directly for cooling and indirectly through electricity generation, while semiconductor manufacturing requires ultrapure water. Research led by Shaolei Ren estimated that GPT-3 could consume a 500-millilitre bottle of water while producing approximately 10 to 50 medium-length responses. That does not mean every 100 words consume a bottle of water. Climate, location, time of day, cooling systems and the electricity mix all change the result (Danelski, 2025; Li et al., 2025).

Magnetic chips can help, but efficiency has limits

Magnetic chips generally refer to spintronic technologies use both the charge and spin of electrons and can retain information without continuous electricity. Some designs reduce the constant movement of data between memory and processor, one of modern computing’s most energy-intensive operations (Marrows et al., 2024; Puebla et al., 2020).

The potential is significant. A review published in npj Spintronics described prototypes that reduced energy use by 89% to 90% in certain associative-memory operations. But this does not mean that a large language model or an entire data centre will consume 90% less energy. (Marrows et al., 2024).

If magnetic devices reduce server electricity use, they could also generate less heat and indirectly reduce the energy and water needed for cooling. Yet their own production requires energy, ultrapure water and minerals. Their environmental value must therefore be measured across their entire life cycle, not only while they are operating (Li et al., 2025; Marrows et al., 2024; Puebla et al., 2020).

There is also a larger problem: the rebound effect. When efficiency makes computing faster and cheaper, companies and users tend to consume more of it. A chip can use less energy per operation while total energy demand continues to rise because the number and complexity of operations grow even faster.

Three conditions for responsible AI in Canada

Canada’s cold climate and clean grid, with over 83% low-emission power, make it an attractive hub for data centres, but expanding them risks straining local resources and crowding out other industries (ISED, 2026a).

To ensure sustainable growth, Canada should require operators to meet three core conditions:

· Fund their own power: Large facilities must finance the clean energy, storage, and grid upgrades they require, rather than passing costs to consumers.

· Protect local resources: Operators should scale back non-urgent workloads during peak demand, prioritize water-scarce regions less, and repurpose waste heat.

· Ensure transparency: Companies must publicly track their energy, water, and carbon footprints, while AI developers should guide users toward right-sized, efficient models.

While emerging hardware like magnetic chips can improve efficiency, technology alone cannot offset AI’s environmental footprint. Lasting sustainability requires combining hardware innovation with strict regulation, political commitment, and responsible use.

Updates

Get your tech and AI news summarized in minutes.

Alerts

© 2025. All rights reserved.