00:00One-fifteenth of a teaspoon. That's how much water the average single interaction with
00:06ChatGPT uses, according to Sam Altman, the boss of OpenAI.
00:11So if you type, can you help me solve this maths problem? That's a drop. Or, can I put
00:17lime instead of lemon in this recipe? That's a drop. Or, why is the sky blue? Help me write
00:23this email. Help me improve my website code.
00:27Mr Altman claims there are one billion messages sent to ChatGPT every day. And ChatGPT is
00:34just one AI bot. Chuck in Gemini, DeepSeek, Claude and others, it's clear that the AI revolution
00:41is a thirsty one. Striking though it is, some experts are more than a little sceptical of
00:48Sam Altman's estimate on water usage.
00:50Sam Altman, At this point, there was just not enough information for me to agree with
00:57or trust the number. Their number was perhaps referring to some tiny models. We are considering
01:05a medium-sized larger language model. That's the size of GPT-3. Basically, if you write an
01:12email or ask someone questions, if you have 10 to 50 queries, you're going to be consuming
01:17roughly 500 millilitres water. This calculation includes water used in cooling and electricity
01:23generation. The BBC asked OpenAI for more details about Sam Altman's estimate, but the
01:28company declined. Either way, it's clear, AI uses a lot of water. But why? Every time you
01:36send a prompt to an AI, it has to run complex calculations to understand and respond. This work
01:44is done by the most powerful and specialised computer chips in the world, housed inside
01:49enormous data centres. Even before users can send prompts, the training process for the
01:55models uses the chips to carry out intense work. And all that extra power means the hardware
02:02can overheat and become damaged if not cooled properly. Most data centres use air cooling systems,
02:09which was fine until AI came along. But now, because these data centres and the infrastructure
02:16that's going in is so much more energy intensive, there are liquid cooling approaches that are
02:23now being implemented. For liquid cooling, the water must be clean to prevent bacteria growing
02:29or clogs and corrosion in the system, which means using mostly drinking water. Here's how the
02:34most common liquid cooling process works. It begins by piping coolant over the processing
02:40chips within the servers. This cooling liquid absorbs the heat and takes it away from the
02:45electrics to a heat exchange unit. Water is used to reduce the temperature of the coolant.
02:51The coolant then recirculates back to cool the servers. Meanwhile, the now hot water is piped
02:56to cooling towers, where a combination of fans and water vapour dissipate the heat, cooling
03:01the water. Some of the water evaporates in that process, while the rest is recirculated
03:07through the cooling process several times before being discharged back into the nearby
03:12water source. Overall, up to 80% of the water evaporates.
03:16What it means is that this type of water is gone and that we are extracting water from a water
03:23circuit that is necessary for irrigation, for human consumption and hygiene.
03:31Communities around the world concerned about data centres putting stress on water sources
03:35and electricity grids are pushing back. Protests have been held in Spain, India, Chile, Uruguay
03:41and parts of the US. And it's not just the operations within the data centre that need water. Generating
03:48the electricity to run them requires a lot of water too, because power plants like coal, gas
03:53and nuclear heat water to create steam, which drives a turbine.
03:58The International Energy Agency has said electricity demand for AI optimised data centres is expected
04:04to increase by 400% by 2030 to 300 terawatt hours. That's roughly the electricity consumption
04:13of the whole of the UK for a year. And aside from electricity, water is also needed when manufacturing
04:20the semiconductor chips used to run AI.
04:24So water is both used directly and indirectly in the whole supply and creation chain of AI
04:31technologies. It is used for the refination of the critical raw materials that are needed
04:37to create the hardware of AI.
04:39Getting accurate figures on how much water it takes to build AI systems and run them is difficult.
04:45Google, Meta and Microsoft release annual figures showing that their data centres use billions of litres
04:52of water every year from local sources, but none of them indicate how much of it is due to AI.
04:58Most tech giants recognise the impact it's having. Many, including Google, Microsoft and Meta, have pledged to be water neutral by 2030.
05:11We hope that can happen, that there is a long way to go to get to those kind of numbers.
05:17Part of what we hope to see is, across the industry, a range of innovations that allow us to maybe minimise the use of water.
05:25Companies are trialling, for example, ways to cool data centres without evaporating any water at all,
05:31and to use the heat that's generated to warm homes.
05:34There are also experiments to move data centres away from communities entirely under the sea, to the Arctic, or even off the planet.
05:43Can we actually put capacity out in space?
05:47It's very, very early stage, so what we at NTT are looking at is,
05:52can we launch satellites that can at least do some more backup-oriented or other-oriented tasks?
05:59Though sceptics point to the many hurdles that need to be overcome,
06:03there is optimism, too, about a more sustainable future.
06:07Let's remember that that Gen.AI capability is still very, very young.
06:11It's moved exponentially fast, but as an industry and as a youth, it is still young.
06:17Ideally, we can learn together, as a society and as a world society,
06:22how do we minimise, again, the use of water and energy?
06:25Because this is all, you know, a world resource when we talk about water.
Comments