AI might be “saving” us time and money, but at what air quality cost?

Artificial Intelligence (AI) is driving a lot of innovation, especially for healthcare, and it’s a great thing.  It discovers new ways to use current drugs, and new drugs and treatments for as-yet uncurable diseases.  Here’s the thing though: AI seems to be “free” and accessible to most people, but most people don’t know how much AI really costs. 

I once thought that the internet was a low-cost tool.  To the average smartphone or computer user, it is… you pay an (mostly affordable) subscription every month, or you just go to a coffeeshop and use their subscription (but for the average minimum price of a coffee in a coffeeshop, $3, you would be better off buying internet access for your home for use anytime).  However, AI is now built into search engines, and it’s not easy to switch off.   For every random question we ask, it jumps in and gives an “AI overview”.  (Just writing this article prompted over a dozen queries!)  Behind each AI overview, there’s a data center somewhere that’s consuming more electricity that what would be used for a non-AI answer.  AI answers actually cost more in terms of air quality to make electricity and keep the data centers cool. 

The power plants and backup generators needed to keep data centers working generate harmful air pollutants, such as fine particulate matter and nitrogen oxides (NOx). Sometimes they are built in areas that are already burdened with pollution.  For example, a new xAI data center exacerbated air pollution in a Memphis, Tennessee, neighborhood that was already struggling with poor air quality. (Elon Musk sparks backlash after alarming trend surfaces over data center: 'Glad they're bringing attention to the issue')  These pollutants take an immediate toll on human health, triggering asthma symptoms, heart attacks, and even cognitive decline.  (We Need to Talk About AI’s Impact on Public Health)  These particulates and NOx are part of “ambient air pollution”, which is responsible for approximately 4 million premature deaths worldwide each year.  Air pollutants that are created to generate electricity don’t stay near their emission sources (power plants): they can travel hundreds of miles, and PM 2.5 is considered a “nonthreshold” pollutant, meaning that there’s no safe level of exposure.  Data centers also have “back up” power sources such as diesel generators to run their servers in the event of a power outage, and these contribute more PM2.5 than the normal source. 

Then, there is the cost of building the data centers and the equipment that goes inside them.  The AI server market is growing phenomenally. ABI Research forecasts that the global market size will reach US$245 billion in 2025, up 25% from 2024. By 2030, AI server sales will grow even further, pushing the market to US$524 billion, representing an 18% Compound Annual Growth Rate (CAGR).  

You may have heard of Rare Earth Elements (REEs).  REEs consist of 17 metallic elements with similar chemical traits. This group includes the 15 lanthanides, plus scandium and yttrium. These elements aren’t truly "rare" regarding their presence in the Earth’s crust. However, they are typically scattered rather than gathered in deposits that are easy to mine profitably. This spread-out nature complicates their extraction and purification.  There are two primary methods for REE mining, both of which release toxic chemicals into the environment. For every ton of rare earth produced, the mining process yields 13kg of dust, 9,600-12,000 cubic meters of waste gas, 75 cubic meters of wastewater, and one ton of radioactive residue.  (Not So “Green” Technology: The Complicated Legacy of Rare Earth Mining)  Here are some ways REEs are used in the data centers: (Rare Earths in the AI Era: How Data Centers Are Driving Demand for Forgotten Metals)

• Superconductors: Yttrium aids superconductors that enhance computing speed and efficiency.

• Hard Drives: Neodymium magnets allow compact designs with high storage space.

• Cooling Systems: Yttrium-based superconductors boost energy savings.

• Power Supplies: Lanthanum enhances batteries for steady power during outages.

Then, there are the specialized “racks” that hold the servers.  These racks are designed to supply high power and cooling loads for the servers. A fully loaded inference rack costs $400,000, while cutting-edge systems can be $3M.  Based on current forecasts, the number of fully-loaded server racks deployed in the U.S. will need to triple in 2025 compared to 2024, and continue to scale in subsequent years. (Beyond GPUs: Fixing the Server Manufacturing Crisis That Could Derail America’s AI Future)

Governments can hardly say “no” to investment propositions, whether in the US or abroad.  Whether it’s news of a 250,000 square foot manufacturing facility and thousands of jobs coming to Texas with the Apple facility in Houston, or China’s offer to build national roads, highways, and hospitals in the Democratic Republic of Congo in exchange for lithium mining rights, these expansions have costs in clean air and water.  

All sorts of industries and manufacturers have jumped to embrace AI, but personally, upon learning that all the pollution that it’s producing, I’ve started disabling it from my searches at least part of the time.  For example, here’s an easy way to stop Google from producing an “AI overview” on a search: just add “-AI” to the end of your query.  I don’t feel like I’m missing out, because these overviews generate sources that I have to double check anyway.   Doing things “the old-fashioned way” (like searching without AI) seems to be more environmentally responsible for non-”life and death” questions.

Photo by Ilya Pavlov on Unsplash