Category: Environmental

Potential Source of Harm: Energy Consumption

Updated October 19, 2024

 

Nature of Harm

AI systems require a huge amount of computational power ("compute"), both for model training and for execution at scale. This compute is provided by servers and other specialized hardware, in particular graphics processing units (GPUs). Operation of this hardware and associated data center infrastructure requires a large amount of energy in the form of electricity, recently estimated at between 85 to 134 terawatt hours per year by 2027 (approximately 0.5% of global electricity consumption and 10% of total data center consumption). This in turn risks accelerating climate change, because global energy production continues to rely heavily on carbon-intensive power sources.

 

There is recent evidence of fast-increasing global demand for electricity, with significant growing uses of electricity including data centers (for AI and otherwise), electric vehicles and heat pumps -- see, e.g.

 

However, a 2024 paper by Google and UC Berkeley explores the complex uncertainties of energy usage by AI models, and argues that energy consumption may be lower than expected if efficiencies are realized. 

 

Regulatory and Governance Solutions

Electricity markets are generally highly regulated around the world, but this regulation may struggle to directly address electricity demand that is growing faster than expected, including because electricity is mostly fungible and not usually restricted or designated to particular applications. If electricity grid demand begins to exceed capacity in particular countries, affecting grid availability and stability, it may be necessary for utilities and regulators to increase restrictions on customers and usage.

 

The European Climate Law (June 2021, part of the European Green Deal) sets out a target of EU "climate neutrality" by 2050, with intermediate targets including a 55% reduction of greenhouse gases by 2030 (compared to 1990). Various AI systems will be subject to these mandates. The EU AI Act contains provisions on development of standards on AI energy efficiency (Art. 40(2)) reporting by providers of general-purpose AI models (GPAIs) of the known or estimated energy consumption of their models (Annex XI, Section 1(2)(e)).

 

New data centers typically require various regulatory approvals, including zoning / planning permission, and may benefit from tax incentives (which governments can adjust).

 

Technical Solutions

The most likely solutions for the harm of energy consumption are technical (and industrial) in nature, including:

  • increasing the compute-efficiency of machine learning models
  • increasing the efficiency of hardware used for compute
  • using renewable energy to generate electricity.

We intend to add further information on such technical initiatives.

 

Government and Private Entities

We will later add information on the large number government and private entities associated with energy generation and efficiency.