The Environmental Cost of AI: Balancing Innovation with Sustainability

For regular internet users, generative AI is just another way to find answers to 'how-tos' and more, but behind the scenes, things are far more complicated. Of course, we have to be grateful to companies like OpenAI and Google for offering their powerful language models for free. That's why many of us don't even realise the exponential investments these AI models require.

As more and more people turn to ChatGPT and Gemini, we take a closer look at the responsible practices of similar companies, because what they offer costs nothing to our pockets, but a lot to the environment, unless certain measures are taken.

Using Rotating Residential Proxies for Greener AI Workflows

One often overlooked problem in AI is how data is collected. Training AI models means gathering huge amounts of data from the web, and if this is done poorly, it can waste a lot of time and energy, so integrating third-party solutions are a must here. Using proxy servers can help, especially during web scraping, which is when large amounts of online data are pulled in. Proxies switch IP addresses automatically, so websites don't block the scraper for sending too many requests from one place. This lets data be collected smoothly in one go, instead of starting and stopping. In short, proxies help AI teams gather data faster and more efficiently without getting blocked.

On top of this, rotating residential proxy servers also improve the quality and breadth of training data, which can reduce the need for retraining (another energy-heavy process). Because these proxies can simulate user requests from diverse regions and devices, they enable more representative global data gathering. An AI model trained on a rich, diverse dataset is less likely to have blind spots or biases that require fixing later.

In other words, proxies help 'get it right' the first time by avoiding narrow or biased training data, so companies don't have to redo data collection or retrain models from scratch (saving substantial computation and electricity).

Smarter Models and Sustainable AI Practices

Even with efficient data gathering, the overall environmental footprint of AI remains a concern. Consider some eye-opening statistics: Training a single large model (like the 175-billion-parameter GPT-3) was estimated to emit about 626,000 pounds of CO₂ (around 284 metric tons) – roughly equivalent to 300 round-trip flights between New York and San Francisco. It also consumed enormous amounts of water for cooling; one study found that training GPT-3 in Microsoft's U.S. data centers directly evaporated about 700,000 liters of clean water. And the impact doesn't end with training – deploying generative AI at scale draws ongoing power.

Answering user queries with AI can be energy-intensive: an OpenAI ChatGPT response uses about 10 times more electricity than a typical Google search. If ChatGPT had to handle 9 billion queries a day (comparable to Google's worldwide daily searches), it would demand nearly 10 TWh of electricity per year – roughly as much power as an entire country like Ireland uses. This spiraling energy and carbon cost, along with associated water use and even e-waste from rapidly upgrading hardware, poses a serious sustainability challenge as AI adoption grows.

MIT scientists have created an AI training approach that drastically cuts energy use by training one large neural network containing many smaller pre-trained 'sub-networks', which can be tailored to different devices without retraining separate models each time. By avoiding redundant training cycles, such techniques dramatically reduce the electricity (and emissions) required to develop and deploy AI across various platforms – in one experiment, the 'once-for-all' network approach slashed carbon emissions to about 1/1300 of the usual amount for a computer vision model, compared to conventional methods.

Another vital approach is powering AI with clean, renewable energy. Leading tech companies are increasingly running their data centers on renewable sources or purchasing clean energy offsets, which directly cuts the carbon emissions from training and inference workloads. For instance, firms like Google have committed to 24/7 carbon-free energy for their data centers by 2030. Using greener power, AI providers can ensure that even if electricity consumption is high, its climate impact is much lower. Additionally, optimising when and where AI computations are done can reduce environmental strain.