Sustainable approaches and innovation: how to reduce the ecological footprint of AI?

AI and data centres carry a heavy environmental cost. That is why the industry is actively looking for ways to shrink its footprint. It focuses on energy efficiency, water conservation, and better use of resources. These approaches answer the challenges we described in the previous part.

One example is rethinking how we cool. A Google study suggests that over-cooling components, especially hard disk drives (HDDs), can backfire. Temperatures that are too low cause mechanical and electrical problems, which paradoxically raises the failure rate.

These findings support the trend of running data centres warmer. Google, for example, operates some of its centres at temperatures up to 27 °C. That saves energy without hurting equipment reliability.

When data centres tap the climate around them, they can cut energy use sharply:

Beyond efficient cooling, attention is also turning to the waste heat data centres produce:

Wider corporate commitments and investments matter just as much.

The goal is water positivity. Tech companies such as Microsoft and Google have committed to being "water positive" by 2030, meaning they will return more water than they use.

The other direction is renewable energy. Data centres increasingly run on solar and wind power to cut the carbon footprint of electricity generation.

Beyond hardware and infrastructure, optimizing the AI models themselves plays a big role. It can cut resource use sharply.

It lowers the precision of model computations, for example from 32-bit to 8-bit. That reduces compute, energy use, and the carbon footprint without meaningfully hurting result quality.

Smaller, efficient models (students) learn to imitate the behavior of larger ones (teachers). The result is models with high accuracy and far lower resource demands.

It uses only selected parts of the model for a given task. That cuts the number of computations and the energy used.

It stores and reuses frequently repeated parts of a prompt, which sharply reduces latency and compute cost.

OpenAI rolled out prompt caching in its APIs and cut costs by up to 50%, with faster prompt processing on top.

It removes less important neurons or connections in the model, which shrinks the model and lowers compute requirements.

Studies show that careful pruning can shrink a model, and with it the energy used, by up to 90% with barely any drop in model performance.

This method speeds up text generation. A smaller, faster "draft" model proposes several tokens ahead, and a larger "verification" model then checks and, if needed, corrects them.

Because tokens are processed in parallel, inference speeds up sharply without retraining the model.

An open-source library that optimizes inference of large language models with the PagedAttention algorithm. PagedAttention manages memory efficiently by splitting keys and values into smaller blocks.

That gives up to 24-fold higher throughput than traditional libraries, without changing the model architecture.

Systemic change and technological innovation are key to reducing AI's impact. But we also play a big role as individuals and as a society. Each of us can help build a more sustainable digital environment:

These approaches, both across the industry and in our own actions, show that the field takes AI's environmental impact seriously and is heading toward more responsible, more sustainable technology. Sustainability in AI is not just a technical task, it is a shared responsibility. And every thoughtful step counts.

What other innovative approaches do you see as key to reducing AI's ecological footprint going forward?

Looking to learn more about AI's environmental impact? Check out these related articles: