NTT DATA: Can We Achieve Sustainable AI?

Immediate and fundamental changes are needed in the development and delivery of artificial intelligence to prevent climate goals from becoming unattainable. Industry-wide collaboration is essential to address the escalating environmental impact of AI and data centres, with a focus on lifecycle thinking and circularity.

The considerable environmental costs associated with the development, deployment and maintenance of AI systems are widely acknowledged. The technology’s demand for energy presents a substantial challenge to global commitments for a more sustainable future. This issue is prompting both the public and private sectors to devise strategies that balance the growth of AI with decarbonisation efforts.

NTT DATA, a technology services provider, is among the companies addressing the challenge. In its latest report "Sustainable AI for a Greener Tomorrow", NTT DATA calls for substantial changes in how AI systems are designed and operated.

The whitepaper projects that AI workloads could represent more than half of all data centre power consumption by 2028. At that stage, the technology's annual electricity consumption could be equivalent to that of 22% of all US households.

AI growth and resource consumption

The environmental consequences of AI extend beyond just electricity usage. Training a single large AI model can require millions of litres of fresh water for cooling data centre systems. Furthermore, executing just 10 to 50 AI queries can consume up to half a litre of water.

The report also highlights that data centre carbon footprints are projected to more than double by 2030, reaching an estimated 860 million tonnes of carbon dioxide equivalent.

"The resource consequences of AI's rapid growth and adoption are daunting but the technology also can empower innovative solutions to the environmental problems it creates," says David Costa, Chief Sustainability Business Officer at NTT DATA. The report outlines four primary metrics for assessing AI's ecological footprint: energy demand, global warming potential, water consumption and abiotic resource depletion.

Overlooked hardware lifecycles

Another critical factor is the lifecycle of hardware. Digital user devices are responsible for 9.4% of global cobalt production and 8.9% of palladium output, a situation exacerbated by short product lifecycles and frequent replacements.

Data centres contribute to this issue by consuming large amounts of copper, aluminium and rare earth elements, with servers often being replaced every few years to keep pace with performance demands.

The NTT DATA report notes that the AI industry has historically prioritised performance metrics such as speed and accuracy over efficiency. This has led to some modern AI models consuming over 300,000 times more computational power than their predecessors, creating what the report terms an "increasingly exclusive domain accessible only to those able to sustain the energy demands".

An ecosystem approach to sustainable AI

NTT DATA asserts that mitigating AI's environmental impact requires a coordinated effort across the entire technology ecosystem. This includes hardware manufacturers, data centre operators, software developers, cloud providers and policymakers.

The report suggests several interventions, such as running AI workloads in locations and at times that align with renewable energy availability. It also recommends applying green software engineering patterns and prioritising modular, upgradeable hardware to minimise electronic waste.

"AI's capabilities can help manage energy grids more efficiently, reduce overall emissions, model environmental risks and improve water conservation," explains David. He adds, "It's vital for the industry to recognise the challenge and build sustainability into AI systems from the start."

In response, NTT DATA has launched its own initiatives, including remote GPU services that move AI workloads to energy-optimised locations. NTT DATA has also developed the tsuzumi LLM, which it claims requires 250 to 300 times less energy for training compared to conventional models.

Despite these advancements, the report acknowledges that significant barriers persist, including fragmented assessments, inconsistent metrics and a lack of standardised reporting frameworks.

The analysis suggests many focus too narrowly on energy or emissions, failing to consider water usage, rare material depletion and electronic waste.