Energy-saving AI, or Norway-style green hypocrisy?

A recent entry on the list of pointless inventions is a cleaning robot for solar panels that glides like a hovercraft over the smooth black surface. It's not only autonomous but also fully self-powered, which means it guzzles most if not all the energy the solar panel generates. But at least it's not in deficit. The same can't be said about various other applications labeled as artificial intelligence (AI).

If a graphics processing unit (GPU) – the best-known category of AI chips – were a person, he (and this is one example where no woman is likely to call for an "or she" in the sentence) would eat a cow a day, drive an SUV flat out and need the energy resources of a coal-fired power plant to run his home. Rene Haas, CEO of UK chip designer Arm, is not pretending otherwise, either. During an interview with The Wall Street Journal last week, he reckoned that AI data centers may consume between 20% and 25% of all US power by 2030, up from about 4% today.

Far more nonsensical than our solar panel cleaning robot, then, would be an AI with a green purpose that consumes more energy than it saves. This could smack of hypocrisy, too. So, it almost certainly won't happen, right? Not necessarily, as anyone familiar with Norway will know.

With its Tolkienesque landscape of emerald fjords and mountain forests, Norway projects environmentalism. It is a glossy brochure for the green movement. And yet Norway is to its critics one of the great climate-offending hypocrites, exporting its dirty fuels to other countries so its own people can frolic in a smog-free renewables paradise. It's the clean-living drug pusher of northern Europe.

Data explosion

The telecom sector had a problem with its energy efficiency story long before AI was given new impetus by Sam Altman and crew. The narrative telcos want told is that modern-day technologies like 5G are super efficient compared with their sputtering sires. Spain's Telefónica even has data in its annual reports showing that 1 petabyte of data consumed 268 megawatt hours back in 2016 but just 54 five years later. Despite this, annual energy consumption by Telefónica fell as little as 11% during this period over a smaller footprint (due to asset sales). At Vodafone, consumption is up 2% in the last few years. The overall picture is of not much change.

This could be the result of a data explosion. Telefónica managed 25,662 petabytes in 2016. By 2021, the figure had grown to 113,547. Or it could simply mean that data consumption by customers has little impact on energy use by telcos. Whatever the explanation, the telecom industry does not look dramatically more energy efficient to the average outsider, unconcerned (perplexed, even) by references to petabytes, than it did years ago.

The question now is what impact AI will have as it steadily infiltrates the telco world, often pitched as an energy-saving or climate-friendly tool. In one "use case" shared at FutureNet World, an industry conference held last week in London, France's Orange helped deploy a connected underwater station to take photos of marine life around coral reefs, which apparently play a more important role than rainforests do in carbon removal. Computer vision models supported by Microsoft Azure were used to analyze those images.

A more inward-looking application is radio access network (RAN) energy efficiency. Deutsche Telekom already uses predictive AI to power down basestations when the Oktoberfest revelers have stopped drunkenly sending texts and finally passed out. Does RAN efficiency require GPUs? Nvidia, the GPU powerhouse, seems to think so. "Spectral efficiency," determining how much data can be fired down the same bandwidth channel, would be improved with Nvidia's latest products, the company said in early 2024. Tarana, a telco vendor, wrote in a blog several years ago that boosting spectral efficiency is like tuning the engine of a car to make it drive faster.

Miniature Norways

Telco executives aren't morons, and none will deliberately invest in energy-saving AI that does the opposite of what's intended. The risk in this specific case is that a network becomes a bit like a miniature Norway – appearing climate friendly at a quick glance because the icecap melting activities happen far from home. Telcos are already big customers of the GPU-wielding hyperscalers. An emerging model could see those hyperscalers put their energy-hungry technologies in "edge" facilities the telcos can use for RAN processing and other stuff.

Energy consumption at these companies has rocketed in the last two years. Although Microsoft hasn't published an update for 2023, in 2022 the company guzzled 18,645 gigawatt hours, about one-and-a-half times as much as Deutsche Telekom, Europe's biggest telco and the owner of T-Mobile US. Yet while annual energy consumption barely changed at Deutsche Telekom between 2020 and 2022, at Microsoft it was up roughly two thirds.

That doesn't mean Microsoft is dirty. Those 18,645 gigawatt hours could have been generated by renewables, and Microsoft's data shows that fewer than 500 originated with fossil fuels. As few as 236, however, came from on-site renewable energy. The difference is almost entirely attributed to power purchase agreements (PPAs) and renewable energy credits (RECs), the twenty-first-century successor to the medieval indulgence. A REC similarly absolves its holder of sin but doesn't necessarily stop the sinning.

Heavy reliance on RECs and PPAs explains why there is such a vast delta between market-based and location-based carbon emissions. The former is based on the paperwork a company signs. The latter assesses the energy composition of the electricity grid where that company shops. In 2022, Microsoft's Scope 1 and 2 emissions, those it directly generates, came to 422,442 metric tons under the market-based system. Under the location-based one, they totaled 6,520,663.

Microsoft is this voracious only because of external demand for data center services, AI included. This should, then, be reflected at telcos (among others) in Scope 3, the emissions blamed on suppliers and customers. At Deutsche Telekom, these came to nearly 10.5 million metric tons last year, down from almost 12.3 million in 2022.

As good as a smaller number sounds, Scope 3 emissions are notoriously hard to calculate, involving what Deutsche Telekom calls "assumptions" and what less charitable onlookers would refer to as guesswork. The sharp drop last year "resulted from methodology adjustments for upstream activities in particular," the telco acknowledged in its annual report. Regardless, Scope 3 accounts for a monstrous 98% of Deutsche Telekom's emissions total, based on its preferred, market-based system. For anyone who lives close to sea level, it may be time to start pumping the inflatables.