Recently, I came across an article about ChatGPT’s electricity consumption. It prompted me to reflect on the interesting topics I have been investigating as a conference producer working in the AI and electricity sectors over the past few years – specifically, the innovations that may dictate the future of AI’s electricity consumption and climate impact:
1. Obtaining large quantities of renewable power is becoming easier.
Recently, buyers and sellers of renewable power are looking to reduce uncertainty when conducting transactions and not be caught out by costly shifts in wholesale prices. There are now more creative ways of buying and selling power, including innovative power purchase agreements (PPAs).
2. Hardware innovations may reduce AI’s vast energy consumption.
As AI is incredibly computationally complex and energy intensive, technologists are using special hardware, like GPUs or FPGAs, to handle the workloads – and that hardware is getting better and better each year. A plethora of AI hardware start-ups are emerging that offer neuromorphic, photonic and other forms of hardware to reduce AI's vast energy consumption.
3. Model optimization may also help.
To reduce the burden on hardware, developers are shrinking and compressing AI models. There are optimization tools in the software layer coming out of academia that are becoming more user-friendly, such as quantization, pruning and knowledge distillation that can decrease AI's carbon footprint.
4. As the AI stack changes, more electricity may be required in the future.
One niche area I’ve investigated is how regulation and IP secrecy are increasingly influencing AI development. Model and data privacy will become a priority, meaning technologists are proposing tools, such as the ‘holy grail of cryptography’ – homomorphic encryption – to keep models and data private during AI computation. This is to ensure that IP, privacy rights and the law are all respected. However, a downside to this is that, when encrypted, AI will consume more electricity. Nevertheless, researchers are working hard to make such tools more efficient, so hopefully this problem will diminish in the future as encryption and other privacy-enhancing technologies become leaner.
5. 24/7 green AI is possible!
This may seem like an amalgamation of marketing buzzwords, but #247GreenAI is real. Energy storage innovation means that when the wind doesn’t blow and the sun doesn’t shine, renewable electricity can be stored when there is excess and saved for later consumption when there is a shortage of clean energy to power AI. IoT and blockchain solutions are also enabling AI developers to match their energy consumption data with clean energy production every hour of the day – when linked to the official energy certificates, you’ll be able to verify how green AI is!
There is plenty to be optimistic about: with enough time and resources, these innovations could result in a future of carbon-neutral AI.
