As cool as artificial intelligence may seem, it’s actually quite hot. Like really, physically hot.

Whether it’s being used for content creation, big data management or auto-completing the email you just sent to a work colleague, in the past few years, AI has made a big impact on the day-to-day lives of your average Joe and Jill. 

But it’s also making a big impact on the environment, and not a positive one.

Reports have come out fingering AI as a major resource-guzzler. According to a recent estimate by the World Economic Forum, the power consumption linked to AI is increasing annually at a rate ranging from 26% to 36%. By 2028, this consumption is expected to equal that of a small country such as Iceland.

And this problem is only a microcosm of a much larger issue posed by the Internet at large.

Before we get to that, though, let’s take a closer look at AI’s environmental cost.

A quick note on greenhouse gases

This article is not about greenhouse gas emissions — it’s mostly about the electricity consumption of AI chips and the data centers that house them. While the two metrics are inherently linked, it’s still an important distinction to make.

The reason that we aren’t about to go into statistics and data on the CO2 output of AI is because reports on that number vary pretty extensively. 

It’s much simpler to look at the hard electricity input — how many watts are going into AI — rather than guess which of those electricity sources are renewable, which aren’t, which emit more carbon dioxide, etc.

There have been claims that the data center industry accounts for 2.5% – 3.7% of global carbon emissions, thereby making it a worse polluter than the aviation industry (which clocks in at about 2.5%).

This is probably true, but calculating how much of that “data center CO2 pie” is being baked by AI specifically is difficult, and statistics published by different sources often conflict with one another.

In short, AI’s electricity usage is directly linked to its carbon emissions, but the impact varies significantly based on the energy sources used and the efficiency of the systems involved. 

This means there’s a question mark hanging over the final climate impact vis-a-vis carbon output, but by looking at the required electricity, we can get a decent idea about the ultimate emissions.

Still with me? Great.

Watt’s the big idea?

In case you don’t know what a data center is, imagine a little beehive filled with bees (in this metaphor playing the role of data servers). 

The bees bumble around doing hard work, but they need a place to hang out and keep their honey and dance, so the hive acts as a facility that provides the necessary infrastructure to do that work. 

Now imagine the beehive is 100,000 square meters. That’s a data center.

While there are dedicated AI data centers designed and optimized for AI workloads, a lot of companies are also putting AI chips into their traditional data centers as well. 

Either way, the increased emphasis on AI within the data center industry ultimately requires increased resource usage.

As such, modern data centers and the AI chips they house have something in common with George Clooney after a 72-hour juice fast: they’re really hot and they’re really hungry.

Each chip — and by extension, each server and each data center — consumes a hefty amount of power. That power usage can be broken down into two main categories: 1) fueling the mechanical operation of the equipment; 2) keeping said equipment cool so that it doesn’t melt a hole in the floor.

Overall, those numbers add up into a whopping power bill, so much so that some countries have even resorted to reopening nuclear power plants to keep a handle on things.

Several companies are developing solutions to the former category: notably Nvidia, which creates the GPUs that most AI models run on these days.

While AI models definitely suck up a lot of resources during the training process (though it’s unclear exactly how much), in a Q&A earlier this month, Nvidia CEO Jensen Huang noted that, once an AI model is trained, “the amount of energy used versus the alternative way of doing computing is much, much lower.”

In a similar vein, there’s an idea going around that it’s possible to lighten the load (and by extension, the power bill) of data servers by installing “AI accelerator” chips on individual computers.

AI accelerators allow AI applications to run efficiently and effectively on local devices, improving performance and reducing latency, while saving both bandwidth and overall power.

Tel Aviv-based Hailo, a fine example of a company operating in this field, recently signed a deal to supply AI accelerators to Raspberry Pi, one of the most popular products for computer nerds who like building their own machines for niche purposes.

Besides just building them better and making them local, there’s another way that companies are trying to make AI computation less energy-hungry: NeuReality, for instance, is an Israeli startup working on an AI system add-on that effectively boosts the efficiency of preexisting GPUs and accelerators.

Recent tests show that when the NeuReality device is paired with certain Qualcomm accelerators, it can achieve up to 90% cost savings and 15 times better energy efficiency than AI setups running on Nvidia GPUs.

Whether or not Nvidia’s claim that AI computation is cheaper in the long run turns out to be apocryphal has yet to be seen. But considering that it’s simply financially smarter to develop chips that use as little power as possible, good ol’ capitalism will likely ensure that Big AI will try its best to reduce the electricity requirement on the computation end.

Cool it, man

That leaves cooling.

Traditional data center cooling relies heavily on air- and water-based methods, mainly using air conditioning systems, large fans and liquid cooling to dissipate heat generated by servers.

While that’s pretty effective in preventing overheating, these systems are notoriously energy-intensive and inefficient: The energy required to power these cooling systems often matches or even exceeds the amount of energy consumed by the servers themselves, effectively doubling the overall energy usage of data centers.

At this point, I’d forgive you for thinking something along the lines of: “Okay, so we all melt and the world ends before SkyNet can even take over, the end?”

I forgive you, but no.

Lucky for us, the smarty pants of the tech world aren’t only working on solutions for AI’s computational power requirements; they’re also figuring out better ways to keep it cool.

One such company is Israel-based ZutaCore, whose cooling technology promises to significantly reduce the environmental footprint of AI-driven data centers.

The company’s liquid cooling system efficiently removes heat from processors and servers using a two-phase boiling and condensation process. It directly cools chips without using water, which protects equipment from damage.

The system utilizes warm water for condensation, which after use in the cooling process can be recycled for various purposes, such as heating buildings, contributing to a more environmentally friendly and cost-effective approach to data center cooling.

Zutacore’s VP Product Shahar Belkin boils down the solution’s benefits to a simple equation. “Just to compare it to traditional cooling system efficiency: if before it was one watt of cooling for every watt of computing, now we’re talking about one watt of cooling for every 10 watts of computing.”

I’m not a math guy, but that seems like a pretty big improvement.

Looking at ourselves in the mirror

We can summarize the main takeaways of this article in two sentences:

• AI uses a lot of electricity.
• Companies are trying to make it use less, both on the computation side and the cooling side.

However, in the intro I alluded to a larger conundrum that warrants further attention: AI only makes up a portion of the overall energy being hastily gulped up by data centers each second of the day.

Every online interaction, whether a search query or video stream, triggers multiple servers in data centers worldwide, translating into tangible energy consumption. 

This demand for connectivity drives data centers to consume a significant portion of global electricity, with projections indicating a potential rise to 7% of the global total by 2030.

I’m not about to get preachy. The Internet is an incredible tool (as is AI, for that matter), and there’s obviously no chance that modern society will say, “Okay, got it, so less content and less putting stuff online then.”

Still, it’s interesting to consider that people have pretty universally decided that it’s fine to sacrifice environmental resources for the sake of reading articles or watching cat videos on TikTok, but they aren’t quite so enthusiastic about people jet setting around the world on airplanes.

With that said, data companies need to use more renewable energy sources and report on how much power they’re using. By extension, governments must take responsibility for putting guidelines in place to make sure that actually happens.

Lastly, it’s critical to promote the innovative solutions being developed by companies like ZutaCore, which enable us to push forward the advancement of the digital age without turning the planet into an uninhabitable superheated hellscape.

With any luck, we’ll be able to make it long enough to turn the planet into an uninhabitable AI overlord-driven hellscape instead. But at least the servers will stay cool!