AI data centres are consuming record amounts of energy. Explore the scale of the problem, the nuclear revival, and solutions like Google's TurboQuant.
Every time you ask an AI model to write an email, generate an image, or summarise a document, a data centre somewhere consumes electricity. A lot of electricity. In March 2026, the energy consumption of AI infrastructure has become one of the most urgent conversations in technology, with implications that extend from corporate boardrooms to national energy policy.
AI's energy crisis is not a future concern. It is happening right now. Data centres are being built faster than the power grid can expand, nuclear power plants are being restarted to feed AI demand, and the tension between AI's transformative potential and its environmental cost is shaping every major technology decision being made today.
This article examines the scale of the problem, the solutions being pursued, and what individual developers and organisations can do to build AI responsibly.
Understanding the scale of AI's energy consumption requires looking at multiple levels.
AI energy consumption breaks down into two categories:
While training gets the most attention, inference is actually the larger concern at scale. A model is trained once but serves millions of requests per day. The International Energy Agency (IEA) estimates that a single ChatGPT query consumes roughly 10 times the electricity of a Google search.
The data centre industry's energy consumption has been growing at an extraordinary rate:
In the United States alone, data centres are projected to consume 12% of the country's electricity by 2028, up from 4% in 2023.
Key Takeaway: AI's energy problem is not about individual queries. It is about the aggregate impact of billions of queries, thousands of training runs, and an industry building new data centres faster than renewable energy can keep up.
The race to build AI infrastructure has become one of the largest construction efforts in technology history.
Amazon, Microsoft, and Google are each investing tens of billions of dollars annually in data centre construction:
The fundamental challenge is that data centres need reliable, high-capacity power, and many electricity grids are not ready to provide it. In Northern Virginia (the world's largest data centre market), new facilities are waiting years for grid connections. In Ireland, the national grid operator has effectively paused new data centre approvals due to capacity constraints.
This is creating a geography problem. Data centres are being built wherever power is available, not necessarily where it makes the most sense for latency, talent, or environmental reasons.
Perhaps the most surprising development in AI infrastructure is the revival of nuclear power. In 2026, nuclear energy has gone from a politically toxic topic to one of the most discussed solutions for AI's energy needs.
Nuclear power has three characteristics that make it uniquely attractive for AI data centres:
Several landmark agreements have been announced:
Nuclear power remains controversial. Concerns about waste storage, construction costs, and safety persist. Critics argue that the tech industry's embrace of nuclear is driven by its own energy needs rather than by sound climate policy, and that the billions being spent on nuclear deals could be more effectively invested in renewable energy and storage.
Supporters counter that the climate crisis demands every zero-carbon energy source available, and that modern nuclear technology (particularly SMRs) addresses many historical safety and cost concerns.
Key Takeaway: The AI industry's embrace of nuclear power is reshaping energy policy worldwide. Whether you view it as a pragmatic climate solution or a distraction from renewables, it is now a major factor in how AI infrastructure is being planned and built.
While the supply side (building more power) gets the most attention, some of the most impactful work is happening on the demand side: making AI models more efficient.
In February 2026, Google published research on TurboQuant, a quantisation technique that achieves 6x memory compression with zero accuracy loss on large language models. This is a remarkable result that could significantly reduce the compute (and therefore energy) required for AI inference.
Quantisation works by reducing the precision of the numbers used to represent model weights. Instead of using 32-bit floating-point numbers, quantised models use 8-bit, 4-bit, or even lower-precision representations. The challenge has always been that reducing precision degrades model quality. TurboQuant's innovation is a mixed-precision approach that identifies which weights are critical for accuracy and preserves their precision while aggressively compressing less important weights.
The practical impact is significant. A model that previously required eight GPUs to run could potentially run on one or two, reducing energy consumption by 75% or more for the same quality of output.
The energy conversation can feel overwhelming at an individual level, but developers and teams make decisions every day that affect AI's energy footprint.
Not every task needs a frontier model. For many applications, a smaller, more efficient model delivers adequate quality at a fraction of the energy cost. Before defaulting to GPT-5 or Claude Opus, ask whether a smaller model (GPT-5-mini, Claude Haiku, or an open-source model like Mistral 7B) would suffice.
Simple engineering practices can dramatically reduce energy consumption:
You cannot improve what you do not measure. Tools like CodeCarbon, ML CO2 Impact, and cloud provider carbon dashboards allow you to track the energy consumption of your AI workloads.
Cloud providers vary significantly in their energy sourcing. Google Cloud, for example, matches 100% of its electricity consumption with renewable energy purchases. When choosing infrastructure, consider the provider's energy and sustainability commitments.
Key Takeaway: Every developer can reduce AI's energy impact through model selection, inference optimisation, and conscious infrastructure choices. These individual decisions add up to significant aggregate impact.
AI's energy crisis sits at the intersection of technology, climate, and policy. There are no easy answers. AI has the potential to help solve climate change (through materials discovery, grid optimisation, and climate modelling), but its own energy consumption risks making the problem worse.
The responsible path forward requires honesty about trade-offs. Not every AI application justifies its energy cost. A model that saves a developer 30 minutes has a different calculus from one that generates memes. As an industry, we need frameworks for evaluating when AI's benefits justify its resource consumption.
For a deeper exploration of the ethical dimensions of AI development, including environmental responsibility, see our article on responsible AI and ethics.
The energy challenge will shape AI development for the next decade. The companies, developers, and policymakers who take it seriously will determine whether AI becomes a net positive or net negative for the planet.
The AI Forest program covers the full landscape of AI's societal impact, including sustainability, energy, and responsible development practices. Understanding these issues is not just good citizenship. It is increasingly a professional requirement as regulation and corporate sustainability mandates grow.
The question is not whether AI and the planet can coexist. They must. The question is whether the AI industry will rise to the challenge with the same ingenuity it applies to model architecture and product development. The clock is ticking.
Start with AI Seeds, a structured, beginner-friendly program. Free, in your language, no account required.