AI has a massive carbon footprint, how do we solve this? As models grow, they require more power, more water for cooling and a supply of new hardware. To make AI truly sustainable, the industry is focusing on 3 aspects.

Hardware Loop:

Chips such as NVDIA’s H100s have a short life cycle and will eventually be replaced by more advanced chips. Designing servers for proper disassembly is crucial for recycling companies to extract rare earth metals (neodymium) and reuse them in the production of new chips. Reducing the demand for new rare earth metals reduces energy intensive mining demand, thereby creating a smaller energy footprint.

Secondary markets can also be established, allowing old chips to be reused for less intensive tasks such as as edge computing or basic data processing, instead of being shredded.

Energy and Resource Loop:

Training large models requires significant amount of energy and can consume as much energy as homes use in a year.

Data centers are now experimenting with heat reuse, capturing waste heat and pumping it into local district grids. Water circularity is another technique that data centers employ, creating closed loop cooling systems that recycle the same water, instead of evaporating millions of gallons of water.

Getting energy from renewable sources such as solar, wind and nuclear energy is essential in reducing AI’s carbon footprint. Apple’s data centers run on 100% renewable energy since 2018. Companies such as Radiant, have created modular nuclear micro-reactors, which may prove to be a solution to the AI energy waste problem. Companies such as Boom Supersonic, have also created cleaner natural gas turbine jet engines that power data centers.

Data and Model Loop:

This involves the use of “circular software” techniques. Models are not trained from scratch, which requires an intensive amount of energy. Model pruning and distillation allows companies to strip away unnecessary parameters from huge models, creating smaller, faster and more efficient models with the intended intelligence it was supposed to have. An example is Ecosia, the environmentally friendly web browser.

Transfer learning involves reusing pre-trained knowledge from one model to another, reducing the amount of compute hours required.

For AI to be sustainable, we have to focus on these 3 aspects. The costs of pursuing agentic intelligence should not outweigh the cost it demands from the environment. With newer technology and leaner models, we can ensure that AI remains a responsible industry that serves the common good.