As an AI researcher studying the environmental impacts of this technology, every time I tell someone what I do for a living, I wince a little, expecting to hear a strong opinion about AI’s rapid expansion. Increasingly, the crux of the debate around AI’s sustainability has been focused on data centers, which make the nebulous concept of AI very concrete with their massive, humming warehouses full of servers and huge energy requirements.
In the public discourse, data centers have largely been cast as the villains of the AIboom—voracious consumers of natural resources that power our chatbots but threaten our climate progress. But in fact, there are clear, actionable ways we can design, build, and operate data centers that align with both our climate goals and societal values.
Thinking local
The benefits of data centers are global. Once a data center comes online, anyone with an Internet connection and enough money to burn can connect to it. But their impacts are very local, since the energy and water that is consumed and the emissions generated all take place in the municipality where the data center is located.
Currently, data centers are highly concentrated in a handful of regions: Virginia, Ireland, Texas, Singapore. This geographic clustering puts severe strain on already fragile energy grids and aquifers.
The current build out of data centers needs so much energy on such a short time scale (think 100,000 houses popping up in the span of a year) that the solution that many developers are turning to is “behind the meter” energy generation—turbines and generators that are powered by fossil fuels (and are intensely polluting). In the meantime, grid operators scramble to upgrade their infrastructure and make existing customers foot the bill.
Instead of these kinds of knee-jerk solutions (which inexorably provoke valid community pushback), we need to design data centers with smarter strategies that account for local constraints. For instance, we need to look beyond the usual suspects in terms of regions and spread data centers across new, diverse geographies, powered by novel combinations of electricity sources.
And there are many ways in which data centers can be powered: in Nevada, Google is developing a geothermal power project that is set to provide its data centers with 24/7 renewable energy. Regulatory frameworks can accelerate this. Ireland, for instance, has instituted a mandate requiring 80% new renewable energy for new data centers. Given the fast pace of data center buildout, we will need a mix of regulatory approaches and corporate commitments to make it as green as possible.
Conscientious construction
But we should also be tracking and reducing the environmental impacts of data centers all across their lifecycle – for instance, during their design and construction, by choosing more sustainable materials, like timber and low-carbon concrete to reduce overall emissions. And instead of clearing land for new facilities, we should prioritize reusing old factories and industrial buildings that stand vacant. Many of these sites are connected to energy grids and water infrastructure, and already have zoning statuses that make permitting faster, without the need to clear or reclaim land.
We can push the envelope further by recycling the massive amounts of heat generated by data centers and using it to heat buildings. This is already the case in West London, where the Old Oak and Park Royal Development Corporation is recovering 17 megawatts of waste heat generated by local data centers to warm up to 10,000 homes and businesses; in Norway, they’re using it to heat a trout farm. By thinking more holistically about how data centers are built and operated, we can also make them more beneficial to the communities that host them and to the planet as a whole.
Emphasizing sustainability
But in order to truly improve the sustainability of data centers, we need to focus on what’s driving their breakneck growth: AI itself. Instead of adopting the “bigger is better” philosophy, we should be doing more with what we already have.
This means developing smaller, more efficient AI models through techniques like model distillation (training smaller models to mimic larger ones) and quantization (reducing the precision of the numbers used in computation). My recent research showed that choosing the right-size models for specific tasks—instead of the biggest, most generic models out there—can reduce the quantity of energy needed by a factor of 33.
Transparency will play a huge role in catalyzing this change. Individuals and companies want to choose AI models with sustainability in mind, but currently lack the numbers they need. Imagine a little counter that tells you how much energy each one of your ChatGPT queries uses and how much carbon it emits. That could help users and companies take this into account in their everyday AI usage.
Developers need to start exposing energy and carbon footprint data directly in AI model interfaces and APIs, empowering both users and developers to make ecologically-informed choices about the AI they build and use. Code Carbon, a software package I helped develop, enables measuring energy and carbon for open-source models, but we still need buy-in from technology companies to integrate these measurements into widely-used AI tools.
The path forward
AI is a profoundly important, transformative technology; because of this immense potential, it is absolutely critical that we get it right. As a society, we cannot simply throw our hands up and accept skyrocketing carbon emissions and surging energy prices as the unavoidable price of technological progress. Nor should we blanket reject data centers or AI as inherently unsustainable.
By centering the conversation around data centers on sustainability, we can also make it more nuanced. By thinking locally, building conscientiously, and operationalizing sustainability in our deployment choices, AI users, companies, and developers can also help ensure that the infrastructure powering our future doesn’t cost us the earth.