The massive and exponentially increasing computing requirements of AI models may make the industry discard the e-waste equivalent of more than 10 billion iPhones annually by 2030, researchers estimate.

In a Nature paper, researchers from Cambridge University and the Chinese Academy of Sciences take a swing at trying to predict just how much e-waste this growing industry could produce. Their goal is not to limit adoption of the technology, which they stress is promising and likely inevitable but to prepare the world more fully for the real impact of its rapid growth.

Energy costs, they say, have been studied because they are in the game.

The physical materials that make up their lifecycle, and the waste stream of old electronic products … haven't.

Our study does not attempt to precisely forecast the quantity of AI servers and their associated e-waste but provides initial gross estimates that indicate the potential scales of the challenge ahead and explores circular economy solutions.

It's necessarily a hand-wavy business, projecting the secondary consequences of a notoriously fast-moving and unpredictable industry. But someone has to at least try, right? The point is not to get it right within a percentage, but within an order of magnitude. Are we talking about tens of thousands of tons of e-waste, hundreds of thousands, or millions? According to the researchers, it's probably toward the high end of that range.

A series of few scenarios low to high growth, modeling as well the kinds of computing resources they would demand support that long; how long each would last. In total, basic finding of such a thing is the potential increase by as much as a thousandfold for wastes over 2023.

Our results indicate potential for rapid growth of e-waste from 2.6 thousand tons (kt) [per year] in 2023 to around 0.4–2.5 million tons (Mt) [per year] in 2030," they write.

Now admittedly, using 2023 as a starting metric is maybe a little misleading: Because so much of the computing infrastructure was deployed over the last two years, the 2.6 kiloton figure doesn't include them as waste. That lowers the starting figure considerably.

But in another sense, the metric is pretty real and accurate: These are, after all, the approximate e-waste amounts before and after the generative AI boom. We will see a sharp uptick in the waste figures when this first large infrastructure reaches end of life over the next couple years.

There are various ways this could be mitigated, which the researchers outline (again, only in broad strokes). For instance, end-of-lifetime servers could be downcycled instead of chucked on the scrapheap, and elements such as communication and power could be also repurposed. In addition, the quality of software as well as efficiency could also be upgraded to extend the effective lifespan of a certain generation or type of a given GPU chip. Interestingly, they favor updating to the latest chips as soon as possible, because otherwise a company may have to, say, buy two slower GPUs to do the job of one high-end one — doubling (and perhaps accelerating) the resultant waste.

These mitigations might reduce the waste load anywhere from 16 to 86% — obviously quite a range. But it's not so much a question of uncertainty on effectiveness as uncertainty on whether these measures will be adopted and how much. If every H100 gets a second life in a low-cost inference server at a university somewhere, that spreads out the reckoning a lot; if only one in 10 gets that treatment, not so much.

So not only is the low end a option, so it seems, but the high one an inevitability too-that could be construed by some as awkward.