AI's Energy Impact: Challenges and Opportunities
Concerns have arisen regarding AI’s potential environmental toll. Recently, Microsoft shared that the company’s CO2 emissions are up 29.1% from the 2020 baseline, largely due to indirect emissions (Scope 3) from more data centers. It is not just Microsoft. Goldman Sachs predicts that AI demand will lead to a 160% growth in power demand at data centers. In 2023, Nvidia shipped 100,000 AI servers that consume an average of 7.3 TWh of electricity annually, according to the IEA and Nikkei Asia.
These figures have sparked discussions at the highest levels, with board members grappling with the difficult decision of whether to prioritize AI over their companies' sustainability commitments.
However, the long-term net impact of AI on sustainability is far more nuanced than the alarming headlines might suggest. While the growth of AI undoubtedly presents energy challenges, there are promising developments in data centers, AI model efficiency, energy innovation, and AI's ability to drive energy efficiency across various industries. By leveraging these advancements, we can mitigate AI's energy impact and potentially achieve a net positive effect on sustainability goals.
Data Center Developments: Specialized for AI Workloads
NVIDIA prioritizes power efficiency in its AI GPU chips. Jensen Huang, NVIDIA's CEO, stated in an interview ahead of the COMPUTEX 2024, "“Accelerated computing is sustainable computing.” Through the combination of GPUs and CPUs, NVIDIA can deliver up to a 100x speedup while only increasing power consumption by a factor of three, achieving 25x more performance per Watt over CPUs alone." (Source: NVIDIA).
AI data centers can be more energy efficient than traditional ones. AI has different use patterns compared to other hyperscale data center use cases. For instance, unlike e-commerce or financial trading applications, response latency is not as critical in a data center primarily used for training AI models. Therefore, the design of data centers can prioritize utilization and manage power load more efficiently, optimized for the use case of AI model training. These data centers may utilize advanced cooling technology and be located in areas with ample renewable energy sources.
AI Model Efficiency: Small Can Be Beautiful
AI models do not have to be large. Small language models, on-device and locally available models for specific use cases and tasks consume significantly less energy than general-purpose large language models (LLMs). For example, Aidan Gomez, CEO of Cohere, mentioned in a recent Decoder podcast with Nilay Patel that their AI models built with enterprise use cases in mind can take only 20% of the resources to build compared to their competitors' models.
New architectures for model building, such as sparse transformers and mixture-of-experts, might further improve power efficiency.
Currently, there is some hype in AI with companies investing and building similar models and products with limited differentiation. Such repetitive AI investments with unclear return on investments should revert to a more reasonable level over time, as some of these startups run out of venture capital funding.
Innovation in Energy Steps up to the Challenge
Given how strategically important and energy-intensive AI model training and inference are, billions of dollars will be invested in procuring energy needed for their success. As a result, innovative energy solutions that previously seen too expensive are now becoming commercially viable.
Bill Gates unveiled an advanced design of nuclear plants in Wyoming with the company TerraPower, in which he has invested. These next-generation nuclear plants use a traveling wave reactor design, which can operate for decades without refueling and has inherent safety features that prevent meltdowns (Source: TerraPower).
According to a report by the International Renewable Energy Agency (IRENA), the cost of renewable energy has fallen by 81% for solar PV and 62% for onshore wind between 2010 and 2020, making them increasingly competitive with fossil fuels. A modernization of power grids has been ongoing to better match the production and consumption of the clean energy. Now AI demand is set to accelerate the process.
As AI data centers specifically built for model training can be located in remote areas, we can use previously wasted and stranded energy sources to power them and reduce carbon emissions simultaneously. An innovator in this space, Crusoe Energy does just that – their solution captures wasted natural gas from oil and gas operations that would otherwise be flared, and uses it to power modular data centers (Source: Crusoe Energy).
AI Helps Optimize Energy Consumption
An optimistic analysis by PwC suggests that the proliferation of AI has the potential to reduce global greenhouse gas emissions by around 4.0% by 2030, an amount equivalent to 2.4 Gt CO2e – equivalent to the 2030 annual emissions of Australia, Canada, and Japan combined. This change is mainly driven by the reduction of carbon intensity in the economy and the utilization of AI levers in environmental applications (Source: PwC).
Across industries, people have already started using AI to optimize energy usage, and the potential is significant. Here are a few examples:
Global shipping industry applies AI in predictive maintenance, route optimization, cargo planning to save fuel consumption significantly.
Doosan Heavy Industries used Azure Digital Twins and AI to optimize wind farm operations, boosting efficiency and reducing maintenance costs by 15% (Source: Microsoft).
A Santa Maria de Lamas pilot using Optimise AI's building optimization software achieved 30% electrical energy savings, 42% thermal energy savings, 24% electrical CO2 reduction, and 34% thermal CO2 reduction (Source: UKGBC)
While the growth of AI undoubtedly presents energy challenges, there are promising developments in data centers, AI model efficiency, energy innovation, and AI's ability to drive energy efficiency across various industries. By leveraging these advancements, we can mitigate AI's energy impact and potentially achieve a net positive effect on sustainability goals.
Commenti