Why Scientists Must Lead the Shift Away From AI Megadata Centers

Public opposition is growing over the energy and water demands of data centres built to support the artificial-intelligence boom. In response, some technology companies are exploring the possibility of placing data centres in space. Elon Musk’s AI and rocket company, SpaceX, is among several firms planning satellite constellations in low-Earth orbit that could operate as data centres.

The idea is appealing because space-based facilities could access abundant solar power while avoiding opposition from local communities. However, the assumption that enormous data centres are essential for AI-powered scientific progress is misleading. Scientific research has very different infrastructure needs from consumer AI platforms serving millions of users. Researchers should promote a more sustainable alternative based on open-weight AI models — systems whose parameters are publicly available — that can run locally while using computing resources efficiently. This approach could reduce the environmental impact of AI and align its development more closely with the public interest.

How much energy will AI really consume? The good, the bad and the unknown

Data centres have powered the internet economy for decades, but facilities designed for AI models require unprecedented amounts of electricity. The world’s data centres consumed about 485 terawatt-hours of electricity last year — roughly equivalent to Germany’s annual use — and the International Energy Agency expects that figure to double by 2030. Amazon, Alphabet, Microsoft, Meta and Oracle are expected to spend more than US$600 billion on AI infrastructure this year. A decade ago, the same five companies spent less than $40 billion. This rapid expansion concentrates energy consumption in the communities where data centres are built, raising concerns about electricity grids, water use, noise and social equity.

Resistance to new data centres is increasing. A Gallup poll published in May found that 71% of Americans opposed building a data centre in their local area, while only 20% were somewhat in favour.

As scientists who use AI in our own research, we believe a practical solution can be developed on Earth. Research institutions should lead the adoption of open-weight AI models that run on local servers. This decentralized AI model would enable scientists to deploy AI tools more transparently and accountably, while reducing dependence on hyperscale data centres.

Decentralize AI

Although exact figures are difficult to establish, billions of daily AI chatbot queries are handled by data centres operated by major technology companies. In many applications, open-weight models can deliver capabilities comparable to those of proprietary, closed-weight systems. Their main limitation is that deployment often requires technical expertise. The popularity of cloud-based chatbots has created the impression that advanced AI can function only in vast centralized data centres. Our experience suggests otherwise.

The development of personal computing provides a useful comparison. Early computers occupied entire rooms before becoming desktop systems and laptops. The AI industry is only a few years old, but a similar transition towards smaller, more efficient and locally deployed systems is already beginning.

NVIDIA, for example, has announced the RTX Spark laptop chip, which is designed to run advanced AI models such as Google’s Gemma 4 on laptops from Dell, HP and other manufacturers. Apple laptop chips have supported a range of local AI models for years. Although these devices remain expensive, the direction of travel is clear. Google has also developed a version of its Gemini AI model that can run inside an organization’s own facilities. A mini-fridge-sized server rack can support approximately 50 users who send prompts and receive AI-generated responses simultaneously.

Many research groups have the technical expertise needed to install a comparable server and run an open-weight AI model instead of relying on a subscription from Google or another provider. Universities and research institutions are not yet investing sufficiently in this approach, but they should. The environmental costs of AI infrastructure — as well as artificially low access fees subsidized by technology companies — cannot continue indefinitely. Scientific communities have long supported open-source software, including freely available code, training data and model parameters. They should now help establish a more energy-efficient and sustainable model for AI deployment.

Open-source AI tool beats giant LLMs in literature reviews — and gets citations right

In practical terms, an AI model that runs on a laptop today can match or exceed the performance of leading proprietary systems from just one year ago. Epoch AI, a San Francisco, California-based research institute that studies AI development, found that open-weight models typically take about four months to reach the capabilities of frontier proprietary systems.

Technology companies will continue building large AI data centres, particularly to support proprietary services and collect user data for training future models. However, hyperscale infrastructure is not the only route to advanced AI. Personal computers are designed for efficiency: they consume relatively little power when idle and can be cooled without large quantities of electricity or water. Scientific institutions should help create an AI future that reduces dependence on massive data centres and prioritizes sustainability, efficiency and local control.

The agentic future

The capabilities of large language models (LLMs) increasingly depend on the software, data and digital tools connected to them. This approach is commonly known as agentic AI. Early LLMs often struggled with mathematical problems because they generated one token, or word fragment, at a time using patterns learned from training data. Since these fragments did not always form complete equations, the models frequently made errors. That limitation is beginning to change as AI systems gain access to external tools and structured reasoning processes.

Data centres will use twice as much energy by 2030 — driven by AI