The power of AI

AI is highly energy-hungry, requiring massive computing power to train and run models. Training involves setting up the model to learn behaviours, while running the model allows users to interact through prompts and responses – both processes are resource-intensive. Scientific American projects that if current AI trends persist, NVIDIA could ship 1.5 million AI server units per year by 2027. At full capacity, these servers would consume 85.4 terawatt-hours annually – more than the energy usage of many small countries. 

Despite AI’s ubiquity, many overlook the infrastructure powering millions of daily queries. Large language models (LLMs) like ChatGPT-3 handle this load, but the rising demand brings hidden costs. It’s not just electricity – AI’s power use generates immense heat, requiring efficient cooling systems. Most cooling relies on fresh water, yet only 0.5% of the earth’s water is accessible for human use. The University of Oxford reports that cooling methods often consume massive water volumes. For instance, a 1-megawatt data centre, equivalent to powering 1,000 homes, can consume 26 million litres of water annually unless a closed-loop system is used. 

As LLMs grow more popular and companies develop their own models, the strain on the environment will only increase. Still, there is reason for optimism. Awareness is growing, and initiatives to limit AI’s environmental impact are underway – through sustainable data centres, efficient hardware and software, and responsible AI development. Balancing innovation with environmental responsibility is critical.

The rise of small language models (SLMs) 

The future may lean toward small language models (SLMs) tailored for specific industries like IT and customer support. SLMs offer targeted, actionable insights, and deliver real world value with far less computational demand. 

What advantages do SLMs have over LLMs? 

LLMs like GPT-4 automate complex tasks and enhance customer experiences. However, their broad training on diverse datasets limits customisation, demands enormous energy for training and use, and contributes to CO₂ emissions and water shortages due to intensive cooling needs. 

In contrast, SLMs are trained on focused datasets tailored to individual enterprises. This approach reduces inaccuracies and enhances relevance. Fine-tuned SLMs can achieve language understanding comparable to LLMs but with vastly reduced energy consumption. New developments in SLMs are creating energy-efficient, cost-effective AI technologies. Smaller models can run on smartphones, laptops, or personal computers – eliminating the need for massive, water-reliant data centres. 

The CO₂ footprint of SLMs is considerably smaller than that of LLMs. While LLMs display extraordinary capabilities, their environmental impact threatens long-term sustainability. As AI evolves, adopting energy-efficient models and promoting sustainable practices like optimising data centre energy, using renewables, and minimising e-waste is vital. 

Society is shifting toward stronger commitments to corporate social responsibility (CSR). Increasing recognition of AI’s environmental costs is spurring efforts to develop sustainable AI solutions. When choosing between SLMs and LLMs, it’s essential to factor in their potential ecological impacts and advocate for responsible development. By making informed choices, we can leverage AI’s power while protecting the planet. 

Sustainability is a shared responsibility. Employees, businesses, stakeholders, and consumers must work together to shape AI’s future. With greater awareness, we can engage with AI in ways that reflect our environmental values. Both SLMs and LLMs will likely coexist in the future, each suited for specific roles, sometimes overlapping. The key lies in responsible adoption and sustainable innovation. 

Thanks to Andrew Kirkley, The Consultancy World for his contributions and advice on this article.

---

Richard Collins is co-founder of CSR Accreditation and a member of the APPG on ESG's Advisory Board. CSR-A provides Sustainability Consultation, Trining, Reporting and Accreditation.