Industrial AI—Increasing or decreasing carbon emissions?

Posted: May 20, 2024

“Industrial AI has the potential to both reduce and increase CO2 emissions depending on how it is implemented and used.”

That’s what OpenAI’s large language model (LLM) chatbot, ChatGPT–3.5, had to say when we asked it whether industrial AI—artificial intelligence—would increase or decrease CO2 emissions.

As it turns out, ChatGPT is right. AI increases the efficiency of industrial processes and makes it easier to integrate distributed energy resources (DERs), like wind and solar, into the power grid.

But, the computing infrastructure powering AI produces about .01% of global greenhouse gas emissions. That might not sound like much, but as more people use AI for more applications, it’s likely to grow to consume as much power as the Netherlands by 2027.

Whether industrial AI can reduce energy demands enough to offset its own carbon footprint is an open question.

Reducing CO2 with industrial AI

LLMs like ChatGPT work by analyzing what words tend to appear together in huge bodies of text to predict which words will most likely follow each other in a new text. In much the same way, industrial AI analyzes data from machinery sensors, supervisory control and data acquisition (SCADA) systems, distributed control systems (DCS), simulations, supply chain contingencies, weather sensors, financial records, etc. to predict the optimal operating parameters for industrial processes.

AI assimilates large quantities of data from many sources to make predictions that can help with:

Schedule optimization
Root-cause analysis and predictive analytics
Predictive maintenance
Asset efficiency and reliability
Waste management
Materials and engineering innovation
Power generation and distribution optimization

By helping industrial processes run more efficiently, AI can decrease the carbon emissions those processes produce and partially offset AI’s own emissions.

AI’s carbon cost

Whether using industrial AI produces more emissions than it saves depends on several factors.

Processor efficiency

While AI hardware does consume a lot of energy, it continues to become more efficient. For example, NVIDIA’s latest processor for AI computing consumes up to 25x less energy than its predecessor. Koomey’s law proposes that the energy efficiency of computing doubles about every 1.5 years—though that rate may have slowed since the year 2000.

AI architecture

AI that’s tailored to specific industrial purposes could use drastically less power than general-purpose AI like ChatGPT—perhaps 30 times less energy on a given task. Even large AI models can also significantly reduce their power requirements by implementing more efficient computational architectures.

Data center management

Fortunately, most cloud computing data centers where AI calculations take place, are either already running on carbon-free energy or have plans in place to do so by 2030.

But running data centers on carbon-free energy is not enough to zero out AI’s carbon footprint. Estimates of the energy requirements of AI generally don’t take into account the energy required to cool the computers as they produce heat from all their calculations. To minimize the energy needed to cool servers, some companies are developing data centers that are under water or submerging servers in special cooling fluid. Other organizations are even looking at hosting servers in space.

Indirect impacts

Researchers caution that assessing the carbon footprint of AI is not as simple as just weighing up the energy it saves against the carbon it emits. If AI reduces the carbon footprint of making widgets by 10%, but also causes people to use 20% more widgets, the net effect is increased carbon emissions.

The way forward

Whether AI emits more carbon than it saves, it might at least create fewer emissions than viable alternatives. According to one study, AI may use less carbon per task than people—at least when it comes to some writing and illustrating tasks. That is, assuming that human writers don’t end up expending more CO2 fact-checking the output of general-purpose AI text generators.

So, as to whether industrial AI has net positive or negative effects on CO2 emissions, ChatGPT has it right: it’s complicated. What’s clear is that AI can truly make industrial processes more efficient—so much so that many decarbonization efforts will be hard to achieve without it. The best way forward is to encourage transparency about the carbon cost of AI so we can work together to reduce it.

Back to Our Industrial Life

Rebecca Ahrens

Senior Editor, Our Industrial Life, AVEVA

Rebecca Ahrens is the Senior Editor of Our Industrial Life and the host of the Our Industrial Life podcast. She has been writing about industrial software and critical industry for nearly a decade, covering topics like AI, the water-energy nexus, process simulation, digital twins, and industrial decarbonization. She's written stories about using data to save oysters in Tasmania, create sustainably-powered microgrids in California, build digital twins of hydrogen and nitrogen gas pipeline networks in Singapore and Malaysia, train shipping crews using 3D simulations, and create green hydrogen to feed half the global population. She holds a degree in philosophy and the history of math and science from St. John's College in Santa Fe, NM, and attended ArtCenter College of Design's graduate program in Pasadena, CA where she taught and studied writing.