AI’s growing popularity raises a hidden cost: its environmental impact. AI models require significant energy from data storage to processing power, leading to CO2 emissions.
AI can do more and more. Think of any topic; an AI tool can effortlessly generate an image or text. Yet the environmental impact of, say, generating an image by AI is often forgotten. For example, generating one image by AI consumes about the same amount of power as charging your mobile phone. This relevant fact is that more and more organisations are betting on AI.
After all, training AI models requires huge amounts of data, and data centres are needed to store all this data. There are estimates that AI servers (in an average scenario) could consume 85 to 134 terawatt hours (TWh) of power annually by 2027. The message is clear: AI consumes a lot of energy and will have a clear environmental impact.
Does AI have a sustainability problem?
Several things are needed to create a useful AI model. These include training data, a stable internet connection, sufficient storage space, and GPUs. Each component consumes energy to some extent, but the computing power required by GPUs consumes the most. According to researchers at OpenAI, computing power has doubled every 3.4 months since 2012. This huge increase is likely to continue shortly, given the popularity of various AI applications. This increase in computing power is having an increasing impact on the environment. To illustrate, a study by the University of Massachusetts found that training popular AI models could lead to 284,000 kilograms of CO2 emission — as much as an average car driving 31 times worldwide.
Organisations wishing to create an AI model should, therefore, carefully weigh the added value of the AI model against its environmental impact. In addition, the underlying infrastructure and the GPUs themselves need to become more (energy-efficient.
Reducing the impact of AI on the environment
A number of industries are important while making an AI model: the data centre industry, the power sector, the semiconductor industry, telecom operators, and the storage industry. To reduce AI’s environmental impact, steps must be taken in each of these sectors to improve sustainability.
The storage industry and the role of flash storage
In the storage industry, concrete steps can be taken to reduce the environmental impact of AI. An example is all-flash storage solutions, which are significantly more energy-efficient than traditional disk-based storage (HDD). In some cases, all-flash solutions can deliver an 85% reduction in energy consumption compared to HDD. Some vendors even go beyond off-the-shelf SSDs and develop their flash modules, allowing all-flash arrays to communicate directly with flash storage. This makes it possible to maximise Flash’s capabilities and achieve even better performance, energy usage, and efficiency i.e., data centres require less power, space, and cooling.
An additional advantage of all-flash solutions is that they are better suited to running AI projects than HDD solutions. This is because linking AI models with data requires a storage solution that provides reliable and easy access to data across silos and applications at all times — this is often not possible with an HDD storage solution.
Data centres backup power
Data centres can take a sustainability leap with better, more efficient cooling techniques. However, backup generators, such as fuel cells based on green hydrogen instead of diesel, can also be more sustainable. Of course, this hydrogen has to be generated ‘green’; this plays a big role in the energy sector.
Semiconductor industry
More energy from renewable sources is needed because semiconductor manufacturers — especially of the GPUs that form the basis of many AI systems — are making their chips increasingly powerful, requiring more power to run. For instance, 25 years ago, a GPU contained one million transistors, was around 100 mm², and did not use that much power. Today, GPUs contain 14 billion transistors, are around 500mm², and consume 200W of power. So, there are more powerful GPUs, and as a result, they consume more energy. So, the semiconductor industry must bet a lot on energy efficiency — something already happening.
Telecom operators
Telecom operators are essential for fast and reliable data exchange. This is relevant because the effectiveness of AI largely depends on data generated somewhere and then having to be transported to the data centre where the AI is running. For example, consider an AI application that needs input from sensors in a factory. Telecom operators can become more sustainable in various ways, such as by focusing on innovation or reducing emissions in the supply chain.
Conclusion
AI will impact the environment, but initiatives like switching to flash storage or improving data centre sustainability can reduce this impact. Every sector can take concrete steps towards a more sustainable course. This is already being done on a large scale but can always be done faster. It is important to keep investing in combating climate change!
