New TM Forum survey reveals impact of AI on CSPs’ sustainability strategies

The International Energy Agency estimates that AI usage will cause worldwide data center demand for electricity to double to 945 terawatt-hours (TWh) by 2030. That level of consumption is slightly higher than the entire current energy usage of Japan, which has a population of 124.5 million.

So, with the telecoms industry ramping up deployment of AI, TM Forum set out to ascertain whether the environmental impact of AI, for good or ill, will affect the adoption choices of communications service providers (CSPs). We carried out a survey of nearly 100 leaders and experts from CSPs worldwide, with 20% working for global telecoms operator groups, 25% for operators in Asia / Asia-Pacific, 23% in Europe and 8% in the Americas. Job roles were spread equally across AI and data, networks and sustainability.

Calculating AI energy usage is tricky (see section ‘Reading behind the survey figures’), so it came as no surprise that only 2% of respondents to our survey said they can very accurately measure the impact that AI has on their company’s overall greenhouse gas (GHG) emissions. Indeed, 31% stated they are unable to report it with any accuracy or with little accuracy, as the pie chart below illustrates. A further 43% sat in the middle, which suggests they either have visibility for some but not all use cases or are simply unsure.

Yet a supplier’s ability to demonstrate the environmental impact of AI products could be a selling point, if they have a strong story to tell. In total, 43% of CSP respondents consider sustainability to be a very important or important factor when deciding which AI use cases to address first.

Dialing down GenAI?

Despite the relative newness of GenAI usage in telecoms-related services and operations, answers to one of our questions could set alarm bells ringing for AI-focused product and service providers. Almost one third (31%) of our respondents say the energy requirements of GenAI applications have already caused them to curtail their use of them. Yet some factors might explain why that figure isn’t even higher.

If telcos partner with hyperscalers for GenAI, then associated Scope 3 emissions – indirect emissions including those from the production, use and disposal of purchased goods and services – could rise (although this is not a given: Hyperscalers are busy investing in renewable and carbon neutral energy sources, including nuclear power.)

And even if telcos are unsure of the impact of new AI tools on Scope 3 emissions, they may decide the promise of gains in productivity and innovation, or lower Scope 2 emissions, is enough of a reward.

Scope 2 emissions are the indirect GHG emissions that result from the energy a company consumes. Decreasing these by increasing the energy efficiency of a network, for example, can deliver a direct and rapid benefit to a telcos's financials in the form of lower energy costs.

In contrast, although Scope 3 emissions typically account for the large majority of a telco's GHG emissions, reducing them brings few direct financial benefits. In addition, cutting Scope 3 emissions is a long-term goal. Even telcos that are serious about sustainability tend to set net zero targets of 2040. And with some reason. Scope 3 emissions lie outside the telcos’ direct control, making them very difficult to manage: Only 18% of respondents said that collecting the right data for Scope 3 is either easy or very easy to do.

AI as a force for good

AI deployment, of course, can also help CSPs reduce GHG emissions. Verizon, MTC and Openserve, for example, are part of a TM Forum Catalyst that uses GenAI to reduce Scope 3 emissions by suggesting lower-emission products, plans and services to customers, based on user preferences.

And today many telcos are relying on AI-driven automation to better understand and curtail energy usage, thereby reducing Scope 2 emissions. A common example of this is the use of predictive AI to shut off mobile network cells when they are not in operation, something more than half our respondents said their companies are doing as the graphic below illustrates.

What’s more, 63% of survey participants, and in particular those working for global operator groups and large national telcos, said they are using AI to build more sustainable products for customers.

AI is helping in other ways, too. For example, a director within a large Asian telco who took part in the survey highlighted the use of AI to facilitate a switch to solar energy.

How important is sustainability?

In recent years, many telcos have made sustainability part of their procurement criteria and have added AI products and services to their evaluation mix. Ideally, CSPs would be able to use standardized, audited data from AI suppliers that are all following the same reporting methodologies to gauge which vendors to work with. But that isn’t possible today.

That inconsistency, and even lack of transparency, in the data, however, hasn’t prevented some companies from trying to drive supplier compliance with sustainability requirements, as the graphic below demonstrates.

Those CSPs that are trying to give sustainability a weighting of above 30% are spread across both global telco groups and national players in Asia, Europe, the Middle East and Africa. Yet even CSPs that set the most ambitious procurement goals cite the same frustrations as their more pragmatic peers – namely, a lack of clean data with which to track progress on targets, a lack of industry-specific reporting standards and a dearth of AI-based sustainability assessment tools (see bar chart).

Despite the many issues telcos face in accessing data that helps them determine the impact of AI on their sustainability scores, our survey participants are cautiously optimistic about AI’s ability to help their companies control GHG emissions. Only 14%, for example, believe the impact will be net negative, while 41% say the impact of AI will be net positive (see pie chart). As we have seen, there are several reasons for CSPs to hope AI will help them decrease overall GHG emissions, from the use of AI to improve network energy efficiency through to the development of more sustainable products and helping customers consume services more sustainably. The real challenge will be ensuring they have the data and tools to measure their success in doing so.

Reading behind some of the survey figures

Today, the US is home to much of the predicted data center expansion. The 2024 US Data Center Energy Usage Report found that US data center electricity consumption rose from 58 TWh in 2014 to 176 TWh in 2023. It estimates usage will continue to increase to reach between 325 TWh and 580 TWh by 2028. That would mean data centers consume between 6.7% and 12% of total US electricity by 2028, up from 4.4% in 2023.

Hyperscalers are investing heavily in data centers as AI usage grows, as research from our recent report, Monetizing the AI connectivity opportunity, illustrates.

Generative AI (GenAI) is at the center of their new AI product launches, ranging from Microsoft’s CoPilot, Google’s Vertex or AI Mode search, to tools to build GenAI applications from AWS.

GenAI is just one form of AI, of course, but it is particularly hungry for energy. Training GenAI models can entail millions or billions of parameters, making it heavy on computational resources. Once the training model is built, it can be set to work on solving problems with new data – the so-called inference phase.

Scientists have tended to view the training phase as more energy-intensive than inference. But new AI models are altering this perspective, according to Phd student Alex de Vries, writing in a Institut Polytechnique de Paris review: “With the massive adoption of AI models like ChatGPT, everything has been reversed, and the inference phase has become predominant.” De Vries goes on to explain that “recent data provided by Meta and Google indicate that it accounts for 60%–70% of energy consumption, compared with 20%–40% for training.”

The AI emissions obstacle course

Currently, however, it is difficult to calculate the total carbon footprint associated with AI usage. This is partly because AI companies are not releasing data about the energy use of closed-source AI models, which is needed to build an accurate picture, as an article in The MIT Technology Review points out.

Whether AI companies are held accountable could ultimately depend largely on users of the technology. Consumers, for example, typically do not question how much energy goes into running online video services such as YouTube or TikTok. Given that GenAI services are simple to use and offer helpful shortcuts for work, study and entertainment, it is probably safe to assume they will follow a similar trajectory.

But AI is also an enterprise play, and large companies such as telcos that are serious about sustainability reporting need to know how the growing use of AI will impact their greenhouse gas (GHG) emissions.

Read our previous in-depth reports on sustainability: The Sustainable Telco: Navigating the maze of scope 3 emissions and The sustainable telco: Engineering networks for net zero.