Google looks to electronic waste as cost-effective AI server solution

Every year, billions of phones are discarded globally, many of them with perfectly usable processors. At the same time, the tech industry is preparing to spend billions on new AI computing hardware, at high environmental costs. Google, in collaboration with researchers at the University of California, San Diego, is developing a way to bridge those two realities by building a server from recovered phone processors, tackling a waste problem while providing low-carbon computing.

The idea is called phone cluster computing. Instead of treating old smartphones as dead consumer gadgets, the researchers strip them down to the parts that still matter for computing: the motherboard, which contains the processor, memory, and storage. The display, battery, chassis, cameras, and other phone-specific parts are removed. The boards are then collected into clusters and redeployed as a general-purpose computing platform. Google and UC San Diego are putting the theory to the test with a data center built from 2,000 Pixel smartphones, expected to come online in fall 2026.

The AI boom is driving an unprecedented exponential surge in the demand for computing power. Processing and memory chips are hardware providing this power, with the AI industry projected to spend over $1 trillion on AI infrastructure this year. On the flip side, semiconductor manufacturing is one of the most complex and energy-intensive industrial processes in the world, with emissions projected to reach 277 million metric tons of CO₂ equivalent by 2030.

On the other side of the pond, billions of phones are ending up in landfills with their processing components still intact. According to the WEEE Forum, over 5 billion phones were discarded in 2022 alone. If we take an estimated average sustained SoC compute draw of 3.2 W per phone (we did the math), using flagship phone chips produced mostly before and around 2022, and we apply the 2022 5 billion figure, we get an estimated 16 GW of compute capacity in the trash. Even if only half of the phones had functional processors, that's still 8 GW. For context, one of the world's largest planned multi-facility data center initiatives, Meta Hyperion, is aiming for around 5 GW of capacity.

In summary, the tech industry is going to spend a fortune on new chips to create processing power, emitting millions of tons of carbon in the process, while simultaneously throwing away available processing power. Of course, the discarded compute figures are all theoretical estimates, and the realities are far more nuanced. Still, if only there were a way to capture a fraction of that processing power, the carbon footprint of computing could meaningfully shrink.

This is exactly what Google and UC San Diego are trying to do, at least a tiny fraction of it for a start.

Their solution targets embodied carbon, the emissions baked into the hardware before it is ever switched on, encompassing all the energy and process gases consumed in manufacturing it. By redeploying a processor that has already paid its embodied-carbon debt, the project aims to deliver a server's worth of compute without a server's worth of manufacturing environmental harm.

So, how do we go from old phones to server-level computing? After all, you cannot just bolt a stack of intact phones into a rack and call it a server. First, each phone is processed down to the motherboard alone, the part that holds the core computing functionality. Interestingly, the motherboard alone accounts for roughly 40% of a phone’s embodied carbon. While removing the screen and battery means some embodied carbon is still wasted, rescuing the board still represents a massive win for circularity.

Next are software changes. Android is already based on Linux, but its mobile-oriented userspace is built for consumer devices, not cloud workloads. To fix this, the researchers replace it with a general-purpose Linux distribution, giving them a more programmable environment and allowing the cluster to behave more like conventional compute infrastructure. Updating the operating system also removes some protections that are essential on a personal phone but unnecessary in the cloud. The result is a motherboard that is basically a small, efficient Linux server. But how about performance? Can a phone be in the same sentence as cloud-level computing?

Well, the technical case is stronger than you may think. Google says the single-threaded performance of modern smartphone cores can be on par with, or better than, the per-core performance of modern multicore servers. In the company’s comparison, a 2023 Pixel Fold was tested against an ASUS RS720A-E11 server using SPEC benchmarks, and the Pixel’s large cores beat the baseline data-center server core in several cases.

Now, this does not mean one phone is equivalent to one server. A proper server has many more cores, much more memory, higher bandwidth, better I/O, enterprise-grade management, and hardware designed for continuous data-center operation. A smartphone, by contrast, has a handful of heterogeneous cores and typically something like 8 to 12 GB of memory. The trick is to find workloads that fit within those constraints or can be cleanly split across many small nodes.

UC San Diego’s early target is educational and research computing. According to Google, a moderately sized cluster of 20 phones can support peak submission rates for a class of more than 75 students, with grading latency below that of the default AWS backend. Scale that to 2,000 phones, Google's fall 2026 target, and the university expects to support about 100 such classes simultaneously. Google describes the resulting deployment as about 50 server-equivalents worth of compute at a fraction of the usual cost.

Now, the project is still in its early stages, with many kinks, known and unknown, to iron out. Reliability is one of the big unknowns. Consumer phones were never built to run flat out, around the clock, for years on end. The project is explicitly meant to act as a testbed for smartphone-based computing at scale, investigating how consumer-grade hardware holds up under sustained use. Nobody yet knows the failure rates of a rack of repurposed phone motherboards grinding away continuously, but finding out is part of the experiment.

Then there is everything around the silicon: the labor of safely tearing down phones at volume, removing batteries and other components not rated for the environment, and the question of whether the whole pipeline can be made economical enough to scale beyond a research showcase. These are the unglamorous frictions that decide whether the idea ever becomes infrastructure. Hopefully, we will have our answers when the deployment goes live this fall.

Source: Google Research