Why scientists should lead the shift away from AI mega data centres

Growing public dissent against the escalating energy and water consumption of data centers powering the AI revolution has sparked discussions about relocating these facilities to space. Companies like SpaceX, led by entrepreneur Elon Musk, are among several exploring the deployment of satellite-based data centers in low-Earth orbit.

This approach promises abundant solar energy and sidesteps community opposition. However, the assumption that large data centers are essential for enabling AI-driven scientific breakthroughs is misguided. The infrastructural requirements of science are markedly distinct from those of consumer AI platforms catering to millions of users.

Researchers with the requisite technical expertise should advocate for an alternative vision: one centered on open-weight AI models—those with publicly accessible parameters—that can operate locally while emphasizing efficient computing resource utilization. Such an approach would render AI tools more sustainable and better aligned with public interests.

To comprehend the scale of AI's energy consumption, consider this: data centers have fueled the Internet economy for years. However, those constructed to support AI models demand considerable power. Last year, the world's data centers consumed approximately 485 terawatt-hours of electricity, roughly equivalent to Germany's entire energy usage, and the International Energy Agency predicts this figure will double by 2030.

Five major technology firms—Amazon, Alphabet, Microsoft, Meta, and Oracle—are projected to invest over $600 billion in AI infrastructure this year; a decade ago, the combined investment by these five entities was less than $40 billion. Data centers exacerbate this massive energy demand by concentrating it within the electricity grids of specific communities, despite concerns about water usage, noise, and equity.

Yet, this expansion is encountering growing resistance. A Gallup poll from May revealed that 71% of Americans oppose the construction of a data center in their vicinity (20% were somewhat in favor). As scientists who depend on AI for their work, we believe a more viable solution exists on Earth. Researchers must spearhead the adoption of open-weight AI models that run on local institutional servers.

This vision entails outlining a more decentralized approach to AI, enabling researchers to deploy these tools in a more accountable manner while reducing reliance on massive data centers. Although precise figures are challenging to ascertain, the majority of billions of queries directed at AI chatbots daily are currently handled by data centers operated by large tech companies.

Open-weight models can match the capabilities of closed-weight, proprietary models in many instances but often necessitate technical proficiency. This widespread use of chatbots has led to the misconception that advanced AI can only function within vast, centralized data centers. This is not accurate, based on our observations. The history of personal computing provides a relevant analogy.

Early computers occupied entire rooms before evolving into desktop computers and laptops. The AI era is relatively young, but indications of a similar transformation are already evident. For instance, NVIDIA, a leading chipmaker, has unveiled a new laptop chip, the RTX Spark, designed to run powerful AI models (such as Google’s Gemma 4) on laptops from Dell, HP, and other manufacturers; Apple’s laptop chips have supported a variety of local AI models for years.

These devices are currently costly, yet the trend is apparent. Similarly, a version of Google’s flagship AI model, Gemini, is engineered to operate within an organization’s facilities. A rack of servers roughly the size of a mini-fridge can accommodate up to 50 users simultaneously interacting with an AI model and receiving responses.

Many research teams possess the necessary expertise to establish a comparable server loaded with an open-weight model instead of subscribing to Google's services. Academic institutions have not yet embraced this approach, but they must, as neither the environmental impact of AI nor the subsidized fees for access can persist indefinitely.

Scientific communities, which have historically championed open-source software with freely available code, training data, and model parameters, should spearhead this shift, setting an example for broader public AI tool utilization that is more energy-efficient and sustainable.