The Impact of AI Growth on Sustainability Goals: The Strain of Data Centers

The proliferation of artificial intelligence (AI) is accelerating across industries and geographies. Self-driving cars, computer vision algorithms for medical image analysis, text-to-image generative AI, and algorithmic high-frequency trading are just some examples of emerging technologies poised to disrupt civilization as we know it. While corporations and governments worldwide are scrambling to stake their claim to what’s expected to be a lucrative next technological era, one critical narrative surrounding AI has been relatively muted: its potential to destroy the planet. The amount of electricity required to power data centers is expected to exceed 2,200 terawatt hours per year by 2030.

The Hidden Environmental Cost of AI Infrastructure

Carbon costs associated with AI run deep underground and beyond sight, hidden by glossy UIs and world-class customer service teams. Artificial intelligence training datasets and services require hyperscale data centers, industrial facilities that commonly cover dozens if not hundreds of acres and contain millions of computers cooled by large-scale versions of the technological equivalents of air conditioners and refrigerators. These data centers have massive carbon and water footprints: They emit greenhouse gases both directly and indirectly via the electricity they consume, even at data centers that are 100 percent supplied by renewable energy. And they require extraordinary amounts of water to keep their cooling systems online.

For context, Google Corporation consumed 5.6 billion gallons of water across its data centers in 2022. Meta and Microsoft report similarly high consumption. Each of these facilities also represents massive amounts of embodied carbon released during their construction: the energy that goes toward building concrete and steel, transporting those materials to construction sites, and completing builds. Where data centers proliferate, freshwater becomes more scarce. As freshwater becomes more scarce, questions of environmental justice and equity become paramount.

AI Growth vs. Sustainability Commitments

AI’s growth also conflicts with electricity sustainability targets on another front. In short, most of the companies driving AI development are also among the biggest electricity consumers on the planet with demand that’s increasing rapidly year-over-year. Large AI companies get the most traction when they talk about their climate commitments. And yet emissions growth from companies like Microsoft is allowed to run unchecked year after year.

The tension at the core of sustainable AI development that nobody wants to talk about: Right now, the companies leading AI development are two of the largest and fastest-growing electricity users and emitters in the world. No sustainability solution will be “big enough” to matter if AI development continues at its current pace, for three reasons.

First, AI adoption is happening much faster than previous tech revolutions. Adoption of the internet, mobile phones, and personal computers happened slowly over decades. But AI adoption is hitting all sectors at once, and much faster than anyone could have predicted even a few years ago.

Second, there’s a major temporal disconnect between how quickly energy transitions can happen vs how quickly AI is growing. Renewables build-outs, transmission upgrades, and long-duration storage projects take years. Meanwhile, we’re asking AI companies to plan for the next few years at most.

Finally, greenwashing is still greenwashing, no matter how many times you say “carbon neutral.” For a while now, AI companies have been issuing climate commitments left and right. Meanwhile, most companies are increasing their emissions year over year. Microsoft reported a 30% increase in emissions just three years after it released its Climate Commitment.

The Explosive Rise of AI Electricity Demand

AI has many stories, but perhaps none are more important or less understood than electricity consumption. Gigaton-equivalent expenditures on electricity carry huge ramifications for energy security, climate policy, geopolitics, inequality, and economic productivity. Taken together, these numbers aren’t simply big they reshape the world.

Take data centers. Global data center electricity usage is currently estimated at around 200–250 TWh per annum. Relative to S&P Global’s estimate of 2,200 TWh by 2030, that means total electricity consumption will grow nearly 10-fold in under a decade. There is no precedent for energy demand of this scale growing this quickly.

Models that power much of today’s AI known as frontier models, these are large-scale machine learning processes that produce models like GPT, Gemini, Claude, and many others demand powering thousands of specialized accelerator chips for weeks or months on end. Training one state-of-the-art model can cost hundreds of megawatt-hours of energy.

But energy to train models is just the tip of the iceberg. A far larger, and faster-growing, source of AI electricity demand is inference. Inference is the computation that goes on every time an AI model provides a user with a response, generates an image, makes a recommendation, and so on. As more AI-powered services are called on to power the billions of consumer interactions taking place every day, inference energy use proliferates.

Chip makers are working to design ever-more-efficient architectures, and companies like NVIDIA, AMD, and Intel are pouring resources into decreasing the energy intensity of these computations. But advancements in energy efficiency are being dwarfed by growth in demand a phenomenon known as Jevons Paradox.

Beyond Energy: Supply Chains, Resources, and Environmental Justice

AI's environmental impact is vast, and energy use is just one aspect. AI technologies and the data centers that support them rely on scarce and geopolitically concentrated resources, can have severe human rights impacts where material sourcing takes place, and are vulnerable to disruption from extreme weather events.

One area which isn't discussed often enough is AI hardware supply chain sustainability.

Training datasets may be energy-intensive to build, but training AI requires an incredible amount of silicon delivered in the form of application-specific integrated circuits, or ASICs.

Building chips and assembling them into AI accelerators is resource-intensive.

Semiconductor fabrication plants require incredible volumes of ultrapure water, rare earths, and energy. Extracting the cobalt, lithium, and rare earths necessary to build AI accelerators and batteries can result in massive environmental damage and human rights abuses in mining communities in Central Africa and Southeast Asia.

Another area which isn't discussed enough is where the environmental impact of AI is located. Much of the wealth generated by AI will go to high-income countries and wealthy corporations, but the environmental costs won't be shouldered equally. Data centers will be built wherever land, water, and power are cheap which usually means communities without much political power to oppose industrial-scale water withdrawals and heat pollution.

Finally, we should be thinking about what to do with AI hardware at the end of its lifecycle.

As billions of dollars are spent on chip hardware each year, developers will need to think about what to do with all of this hardware once it's considered e-waste.

Data Centers and the Emerging Grid Infrastructure Crisis

This datacenter power demand reality is already being felt today as a crisis in electric grid infrastructure throughout many of the most wired regions on Earth.

As electric grid operators have come to expect demand to grow 1–2% per year, they are suddenly fielding interconnection requests for new grid capacity that would add hundreds of megawatts of load at a time. Utilities across key data center clusters including Northern Virginia, Phoenix, and the greater Chicago area have begun issuing warnings that they will have to push back retirements of existing coal and natural gas generation in order to keep the lights on over the near-term.

This not only burdens utilities and grid operators, but creates a negative feedback loop for climate policy. While AI is being touted as a solution for improving clean energy grid management and speeding up decarbonization efforts, it is simultaneously driving the need to keep fossil fuel generation online.

Grid constraints from data center development are also beginning to spur policy responses globally. Ireland, Singapore, and the Netherlands have all raised concerns or enacted restrictions related to data center growth and sustainability.

Policy and Technology Solutions for Sustainable AI

I don’t think there’s a silver bullet. But there are policies and technologies which align around some big tent principles that I think could significantly improve matters if there’s serious will to do so.

The first step is demanding and institutionalizing transparency. We have to know how much energy AI and data centers use and begin accounting for water use, hardware impacts, supply chain emissions, and grid carbon intensity impacts of data center electricity use before we can hope to improve it.

That information has to be publicly disclosed and independently audited. Ideally these requirements would be legislatively mandated, but corporations can and should take it upon themselves to publish this information if governments won’t require it.

In terms of technology solutions, improving cooling technology, building better chips, and improving algorithmic efficiency at the software level offer the best near-term opportunities for reduction. Model distillation also shows promise as a potential solution to improve the energy efficiency of AI itself.

We could also pull some levers at the level of data center siting and planning. Requiring new data centers to be powered by on-site renewables and implementing strategic moratoriums in strained regions can help mitigate the issue.

A Choice About the Future of AI and the Planet

AI’s energy challenge, then, really is a choice problem. It’s a question of choices by tech leaders about where they locate and how they power their data centers; choices by governments about how to regulate and support an industry that has largely outpaced existing policy levers; and choices by all of us about what we expect from the technology systems we’ve become reliant upon.

Estimates that global data center electricity demand will reach 2,200 TWh by 2030 are not forecasts of an unavoidable future, but rather warnings about where the industry is heading if we take no collective action to change course.

But changing course will necessitate facing some hard truths about AI’s environmental impact; shedding the comforting myth that corporate pledges will be enough to solve industry-wide problems; and dedicating the political will and resources required to realize a vision of AI that’s truly sustainable on a finite planet.

Intelligence without wisdom won’t mean much if we’re not wise enough to prevent building it from ruining the only planet we’ve got. We’re at AI’s climate reckoning moment, and our actions now will speak volumes about what we value.