Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain
the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in
Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles
and JavaScript.
Why scientists should lead the shift away from AI mega data centres
Publicly available AI models and local infrastructure can reduce AI’s environmental footprint while giving researchers greater control over the tools they use.
Cassidy K. Buhler is a postdoctoral fellow at the Cooperative Institute for Research in Environmental Sciences (CIRES) and the Environmental Data Science Innovation and Impact Lab (ESIIL) at the University of Colorado Boulder, Boulder, Colorado.
Fernando Pérez is faculty co-director of the Eric and Wendy Schmidt Center for Data Science & Environment and associate professor at the Department of Statistics, University of California, Berkeley, Berkeley, California.
Carl Boettiger is faculty advisor at the Eric and Wendy Schmidt Center for Data Science & Environment and associate professor at the Department of Environmental Science, Policy & Management, University of California, Berkeley, Berkeley, California.
As public opposition grows against the soaring energy and water demands of data centres powering the artificial-intelligence boom, some technology companies are talking about putting these facilities in space. For instance, businessman Elon Musk’s AI and rocket firm, SpaceX, is one of a handful of companies planning to deploy constellations of satellites in low-Earth orbit that act as data centres.
The logic is seductive: such facilities could tap abundant solar energy and avoid opposition from communities. But the premise that massive data centres are a prerequisite for enabling AI-driven scientific advances is incorrect. The infrastructural needs of science are fundamentally different from those of consumer AI platforms built to serve millions of users. Researchers with the necessary technical know-how should champion an alternative vision: one centred on open-weight AI models — those with publicly available parameters — that can be deployed locally while prioritizing the efficient use of computing resources. Such an approach would make the use of AI tools more sustainable and better aligned with public interest.
Data centres have supported the Internet economy for decades. However, those being built to support AI models require a massive amount of power. The world’s data centres used about 485 terawatt-hours of electricity last year, similar to that used by Germany, and the International Energy Agency expects that to double by 2030. Five technology companies — Amazon, Alphabet, Microsoft, Meta and Oracle — are expected to spend a total of more than US$600 billion on AI infrastructure this year; a decade ago, the same five companies spent less than $40 billion. Data centres concentrate this extraordinary energy demand on the electricity grids of the specific communities where they are built, despite concerns about water use, noise and equity.
But this expansion is facing mounting resistance. A poll published by Gallup in May found 71% of Americans opposed the construction of a data centre in their local area (20% were somewhat in favour of it).
As scientists who rely on AI in our own work, we think a more practical solution exists on Earth. Researchers must pioneer the adoption of open-weight AI models that run locally on institutional servers. Here, we outline a vision for a more decentralized approach to AI — one that allows researchers to deploy these tools in a more accountable manner, while reducing reliance on massive data centres.
Decentralize AI
Although precise numbers are difficult to obtain, most of the billions of queries made to AI chatbots each day are currently handled by data centres run by large tech companies. Open-weight models offer comparable capabilities to those of closed-weight, proprietary models in many cases, but using them often requires technical know-how. This use of chatbots has fostered the misconception that advanced AI can operate only in vast, centralized data centres. This is not true, based on our experience.
The history of personal computing offers a useful analogy. Early computers filled entire rooms before shrinking into desktop computers and laptops. The AI era is just a few years old, but signs of a similar shift are already visible.
For example, the chipmaker NVIDIA has announced a new laptop chip, the RTX Spark, to run powerful AI models (such as Google’s Gemma 4) on laptops from Dell, HP and other providers; Apple’s laptop chips have supported a range of local AI models for years. These devices are currently expensive, but the trend is clear. Similarly, a version of Google’s flagship AI model, Gemini, is designed to run within an organization’s own facilities. A rack of servers the size of a mini fridge can support around 50 users simultaneously sending prompts to an AI model and receiving responses.
Many research teams would have enough expertise to set up a similar server loaded with an open-weight model, rather than buying a subscription from Google. Academic institutions are not currently investing in such an approach, but they must, because neither the environmental costs of AI nor the subsidized fees for access can go on indefinitely. Scientific communities, which have long led the adoption of open-source software — with freely available code, training data and model parameters — should help to drive this shift, setting an example for wider public use of AI tools that is more energy efficient and sustainable.
In practical terms, a model that runs on a laptop today can match or exceed the performance that top proprietary systems achieved just a year ago. Epoch AI — a research institute in San Francisco, California, that investigates AI trends — found that open-weight models take around four months to match the capabilities of frontier proprietary systems.
Massive AI data centres will continue to be built by technology firms, particularly to support proprietary services and harvest user data to train future models. But this is not the only path forwards. Personal computers are designed for efficiency: they consume little power when idle and can be cooled without large amounts of energy or water. Scientific communities must focus on building a future that consciously rejects the need for an ever-greater reliance on massive data centres.
The agentic future
The performance of large language models (LLMs) increasingly depends not just on the models themselves, but also on the external software and tools surrounding them, a process often described as agentic AI. Early LLMs struggled with mathematics tasks because they were limited to generating one token, or word fragment, at a time, based on patterns learnt from the training data. Because these were incomplete segments of equations, they made mistakes. That is changing.
Facts Only
* Cassidy K. Buhler is a postdoctoral fellow at CIRES and ESIIL at the University of Colorado Boulder.
* Fernando Pérez is a faculty co-director of the Eric and Wendy Schmidt Center for Data Science & Environment and an associate professor at the University of California, Berkeley.
* Carl Boettiger is a faculty advisor at the Eric and Wendy Schmidt Center for Data Science & Environment and an associate professor at the Department of Environmental Science, Policy & Management at the University of California, Berkeley.
* Elon Musk’s SpaceX is planning to deploy satellite constellations in low-Earth orbit acting as data centers.
* The world’s data centers used approximately 485 terawatt-hours of electricity last year.
* The International Energy Agency expects data center electricity usage to double by 2030.
* Five technology companies—Amazon, Alphabet, Microsoft, Meta, and Oracle—are expected to spend over US$600 billion on AI infrastructure this year.
* A Gallup poll in May found 71% of Americans opposed the construction of a data center in their local area (20% were somewhat in favor).
* Open-weight models have taken approximately four months to match the capabilities of frontier proprietary systems, according to Epoch AI research.
Executive Summary
Publicly available AI models and local infrastructure offer a way to reduce the environmental footprint of artificial intelligence by giving researchers greater control over their tools. The article discusses the growing opposition to large data centers fueling the AI boom, noting that these facilities concentrate energy demands on specific communities while raising concerns about water use, noise, and equity. Some entities, such as SpaceX, are exploring deploying satellite constellations in low-Earth orbit as data centers. A central argument is that focusing on open-weight AI models, which can be deployed locally, aligns better with scientific needs and sustainability than relying on massive centralized data centers.
The energy demand of data centers supporting AI has increased significantly; the world's data centers used 485 terawatt-hours of electricity last year, and this is expected to double by 2030. Five major technology companies are projected to spend over US$600 billion on AI infrastructure this year, compared to less than $40 billion a decade prior. However, the premise that massive data centers are prerequisites for scientific advancement is challenged; infrastructural needs for science differ from those of consumer AI platforms.
The proposed alternative involves decentralizing AI by deploying open-weight models on local institutional servers. This approach seeks to make AI use more sustainable and align it with public interest by reducing reliance on large energy consumers. The shift is supported by the analogy of personal computing evolution, suggesting that efficient, localized systems are feasible for AI deployment.
Full Take
The narrative presents a structural tension between the established infrastructure supporting consumer AI and the fundamental requirements for scientific progress. The central conflict lies between centralized, energy-intensive data center models and a decentralized approach centered on open-weight models running locally on institutional servers. This framework subtly shifts responsibility: moving from a model optimized for serving millions of users to one optimized for specific, accountable research demands.
The argument about decentralization taps into a powerful historical analogy—the evolution of personal computing—suggesting that localized, efficient computation is the natural trajectory. The move toward agentic AI performance, which depends on external tools, further solidifies the critique against monolithic infrastructure by pointing toward a necessary shift in how computational resources are managed.
The resistance facing data center expansion, evidenced by public opposition and equity concerns regarding energy and water use, frames the technological debate as an ethical and civic one. The underlying implication is that the current infrastructure structure creates externalities—concentrating environmental costs onto specific communities—that must be re-evaluated in pursuit of a more sustainable path for scientific innovation. The failure to invest in local, open-source deployment instead of massive centralized build-outs risks cementing an unsustainable paradigm where access and sustainability are decoupled from accountability.
Bridge Questions: What are the specific governance mechanisms that could effectively incentivize academic institutions to prioritize localized, open-weight deployments over proprietary cloud services? How can the principles of distributed computing be practically scaled to handle the massive computational demands of cutting-edge scientific modeling without sacrificing security or interoperability? What role should external regulatory bodies play in assessing the environmental and equity costs associated with decentralized versus centralized AI infrastructure?
Sentinel — Human
This analysis presents a coherent, well-structured argument drawing on established trends in AI infrastructure and environmental science, exhibiting characteristics typical of informed journalistic or academic synthesis.