Tech giants are no stranger to making big claims, and their climate goals are no different. Amazon, Microsoft, and Google have all committed to net zero carbon goals, but the same thing that is keeping them relevant today, is pulling them back from their goal: AI.

It’s no secret that modern state-of-the-art algorithms take a vast amount of compute resources to train and deploy, but there is a hidden carbon cost that many neglect to mention — electricity. The server farms that train AI models will only grow larger in the future with the AI boom, but will companies be able to balance it with their net zero goals?

AI makes carbon footprints bigger

Tech giants rarely release information about the cost of training AI algorithms, and it seems there is a reason for withholding this data. Hugging Face, an open-source AI community, has shed some light on the approximate carbon footprint of training BLOOM, their LLM. According to its researchers, the training emitted approximately 24.7 tonnes of carbon for power consumption alone. This number ballooned to 50 tonnes when accounting for an end-to-end carbon footprint, including equipment manufacturing. 

A paper published in 2021 estimated GPT-3’s training cost to be around 1.287 gigawatt hours, which is enough energy to power 618 Indian homes for a year. This comes up to 502 tonnes of carbon emissions, and these are just estimates for GPT-3. GPT-4 could have cost the planet more, but we have no way of knowing, and that’s not even considering the operating cost of ChatGPT and its 100 million users. Thanks to the Microsoft-OpenAI partnership, this was done on the Microsoft Azure cloud supercomputer. Coincidentally, in 2021, Microsoft paid for the removal of 1.3 million tonnes of CO2 from the atmosphere. 

Buying carbon credits to cover up the environmental cost of training AI models seems to be the trend of the hour. To this end, Google also recently purchased 300,000 carbon offsets, and Meta bought 7 million carbon credits earlier this week. While the complete data of carbon offset purchases are not available in the public domain, the timing of these purchases seem to reflect a renewed enthusiasm for sustainability from companies who already seem well on their way there.

One possible explanation is that training and deploying AI models have thrown a spanner in the works for companies, offsetting their carbon deployment goals. While the narrative around AI training by the media focuses on the negative environmental impact of the process, tech giants usually provide positive comments. 

A paper published by Google predicts that the carbon footprint of AI training will plateau, then go down owing to eventual adoption of best practices. These practices include using efficient models, optimised compute, and cloud computing. According to the researchers, this can bring down energy costs by up to 100x and emissions by up to 1000x. 

Indeed, there is research to prove that optimising ML models results in a lower carbon footprint. Systems have emerged that intelligently cut down the power draw of GPUs, resulting in a cost-cutting of up to 75%. While the AI world is moving closer to Google’s best practices by creating specialised chips for AI compute, it seems that AI models are getting bigger and more difficult to train. 

Sustainability for CSR’s sake?

Sustainability has become more than just a buzzword for tech giants, with many companies aiming to appease their shareholders and audience with their green goals. This has resulted in top tech giants like Amazon, Microsoft, and Google pledging to go net zero carbon by a certain time frame. While Microsoft and Google have set net zero goals at 2030, Amazon has given itself a longer timeframe, instead claiming its goal as 2040. Meta, on the other hand, has already reached carbon neutrality, and is now pursuing a negative carbon footprint. 

Whatever the case may be, the companies that are moving towards a zero carbon goal are doing so quickly. There is a race going on behind the curtain, especially when it comes to pursuing new technology for carbon-free energy generation and carbon storage technologies.

Google, for example, has even initiated a movement called the 24/7 carbon-free energy compact that aims to decarbonise electricity systems worldwide. Microsoft has invested $1 billion into the Climate Innovation Fund to accelerate the creation of technologies to reduce climate impact. Amazon has a bigger fish to fry, as it needs to cut down on carbon for its logistics chain, and has multiple corporate social responsibility (CSR) plans to cut down its climate impact. 

However, away from the surface-level, sustainability misrepresentation runs rampant. Amazon has been shown to drastically undercount its carbon footprint, while Microsoft has slowly fallen behind its 2030 goal despite pressure from regulators. To add to this, AI seems to be a big pain point for tech giants. 

It seems that big techs don’t have a sustainability solution to deal with the current upsurge in AI demand, instead relying on carbon credits to stay on track for their sustainability goals. As we inch closer to 2030, big techs can only hope that optimisations and specialised compute will push AI’s carbon footprint down to stay in line with their climate pledges.