Artificial intelligence is a vampire, draining the Earth of its water and energy resources. Powering generative AI systems requires an enormous amount of physical infrastructure and data centres now account for 4.4% of all energy used in the US. Little surprise that Silicon Valley firms have largely dropped their commitments to achieving net zero.
Given that the planet is hurtling towards a climate catastrophe, some in the industry have recognised that the AI boom is unsustainable, in multiple senses of the word. Companies, universities and research labs are investigating ways to reduce the technology’s carbon footprint and breakthroughs can’t come soon enough. Here are six projects seeking to lessen AI’s disastrous environmental impact.
MIT and Clover
As the old cliché goes, if all you have is a hammer, everything looks like a nail. One problem with GenAI is that, even when it’s fed simple prompts, it still must run compute-intensive workloads to respond.
To alleviate these energy demands, researchers at MIT and Northeastern University developed a machine-learning platform called Clover, which includes ‘carbon intensity’ among its parameters. Instead of running every prompt through the latest, most energy-intensive model, the tool selects an appropriate model based on the complexity of the user’s input.
This seemingly small intervention helped to reduce the carbon intensity of some operations by as much as 90%.
The IEEE’s P7100 standard
It’s difficult to solve a problem if you don’t understand the extent of its effects. That’s why a working group from the Institute of Electrical and Electronics Engineers (IEEE), the largest technical organisation for STEM professionals, is working to develop a new standard to measure the environmental impact of AI.
The ‘IEEE P7100 standard for measurement of environmental impacts of artificial intelligence systems’, as it’s officially called, aims to provide a framework to quantify and report on AI’s carbon footprint, with input from academia, industry and government.
Crucially, the standard differentiates AI from general-purpose computing. Among the metrics it tracks are compute intensity, environmental impact and the supply-chain impact of the whole manufacturing life cycle.
GREEN.DAT.AI
An experimental project funded by the EU, GREEN.DAT.AI seeks to develop and apply energy-efficient algorithms to keep AI’s power consumption low. The idea is to create shared ‘data spaces’, where companies can pool their resources without sharing confidential information. All of the data can then be used to train new AI models in a decentralised way.
The project also sets out to produce open-source AI services that can be used across industries. Its platforms are deployed at the ‘edge’ of networks – that is, where the information is being gathered from, rather than a centralised data centre – to further reduce energy consumed in data transfers.
Aria’s scaling compute project
In early 2024, Aria, the investment vehicle founded by Matthew Clifford, a former AI adviser to the UK government, kicked off the ‘scaling compute’ project, an R&D programme that aims to make AI hardware 1,000 times cheaper.
One of its initiatives is a collaboration with Oxford University to create systems to drastically improve the speed and efficiency of AI hardware. Because GPUs consume significant resources to communicate and synchronise information, the training method used for most AI systems contains a so-called communication bottleneck.
The Aria-Oxford project proposes a new ‘on-computer’ communication system to boost the rate at which data can travel. By attaching hundreds or even thousands of AI chips to glass fibres and installing them on a single serve, the system reduces the information required for computers to communicate effectively, as well as the distance it must travel.
The Green Data Foundation and climate KPIs for software
Much has been written about making hardware greener. But all software runs on hardware, so if the software is inefficient or wasteful the problem scales up and you’re left with lots of unnecessary carbon waste.
A recently certified ‘software carbon intensity’ specification from the Green Software Foundation enables tech professionals to estimate the carbon emissions of software using openly available data. The organisation’s founder hopes the spec will lead to the creation of energy ratings and sustainability KPIs for software.
Vegetable oil for data centre cooling
The world was in the throes of a water crisis even before the GenAI boom caused data centre construction to skyrocket. Now, with the rise of AI factories – which could use up to 1.7 trillion gallons of water globally by 2027 – the water problem is even more severe.
Few might have guessed that the US food giant Cargill would play a role in solving the problem. But, with its NatureCool 2000 immersion cooling fluid, it may have pioneered a unique solution to the existential issue of water waste.
The fluid, designed specifically for data centres, is made mostly from soy oil, replacing petroleum-based cooling. The firm claims it’s got a 10% higher heat capacity than other synthetic cooling fuels and, crucially, it’s biodegradable; it breaks down in less than 10 days if there’s a spill. Cargill says the direct-immersion method means no water-based heat exchangers or cooling towers are needed.
Other businesses are experimenting with hydro-treated vegetable oil (HVO) as a replacement to diesel fuel for backup generators. That could lead to a 90% reduction in carbon emissions compared with plain old diesel.
Artificial intelligence is a vampire, draining the Earth of its water and energy resources. Powering generative AI systems requires an enormous amount of physical infrastructure and data centres now account for 4.4% of all energy used in the US. Little surprise that Silicon Valley firms have largely dropped their commitments to achieving net zero.
Given that the planet is hurtling towards a climate catastrophe, some in the industry have recognised that the AI boom is unsustainable, in multiple senses of the word. Companies, universities and research labs are investigating ways to reduce the technology's carbon footprint and breakthroughs can’t come soon enough. Here are six projects seeking to lessen AI's disastrous environmental impact.
