The Carbon Removal Paradox

The very companies pouring billions into servers and supercomputers like OpenAI’s Stargate or xAI’s Colossus - each with expected build costs north of $100 billion - are also the biggest buyers of carbon removal (CDR). Microsoft has pledged to purchase 30 million tonnes of removal, while Google and Stripe together add 2.1 million tonnes.

Here lies the paradox: their efforts barely scratch the surface, covering less than 0.4% of the annual removals needed by 2050 to stay on a net-zero path.

The Climate Impact: AI's hidden footprint

Meanwhile, the emissions burden is already mounting. Data centres consume about 2% of global electricity - that's roughly the same as Spain’s entire demand. Training a single advanced AI model can emit 250,000 kilograms of CO₂, equivalent to 125 transatlantic flights. Training GPT-3 alone consumed 1,300 MWh of electricity - enough to power 120 U.S. homes for a year - and 700,000 litres of water, equivalent to the annual drinking supply for 3,700 people.

The Human Impact: AI's effect on communities

And for communities hosting this infrastructure, the impacts are already being felt. In Ireland - another global hub for hyperscalers - data centres now account for 21% of national electricity demand, more than all urban households combined. Fearing grid instability, Dublin has frozen new data centre connections until 2028.

In Boxtown, Memphis, Elon Musk’s xAI data centre is powered by 35 unpermitted gas turbines, the facility has triggered surges in nitrogen dioxide levels - up 79% at peak times - and raised serious health concerns. In India, the rapid expansion of AI infrastructure also risks locking in coal dependency (due to their grid) unless renewable investment accelerates at equal pace.

The Carbon Crux: AI's growth outstrips CDR

AI’s energy and water demands are scaling exponentially - Google reported a 27% increase in year-on-year electricity consumption as it struggles to decarbonise as quickly as its energy needs increase, driven largely by its scope 3 emissions, which are in large part influenced by a growth in data centre capacity.

And carbon removal capacity is only crawling forward. Analysis by the Boston Consulting Group warned that AI-related carbon removal demand could reach 200 million tonnes annually by 2030, but supply today is only about 33 million tonnes - a mismatch that underlines how AI’s emissions are already outstripping the growth of carbon removal.

So, while AI promises to bring real efficiency gains across industries, its infrastructure risks creating a new wave of climate and community burdens, and a new wave of hard-to-abate emissions.