Retired phones build a low-carbon cloud: Google plans to use 2,000 old Pixel devices to provide the computing power of 50 servers

According to Beating Monitoring, Google researchers and a team from the University of California, San Diego (UCSD) are exploring phone cluster computing, planning to reassemble the mainboards of 2,000 decommissioned Pixel smartphones into a low-carbon cloud data center to provide students and faculty with low-cost, low-carbon cloud computing resources.

Smartphone upgrade cycles are typically four years, but the computing cores of decommissioned devices remain intact, with CPU core performance approaching that of modern server cores, and performing comparably in some single-threaded tests of SPEC CPU 2017. To adapt to data center environments, the research team removed non-essential components such as the screens, batteries, casings, and cameras from the old phones, keeping only the mainboards, because the mainboard manufacturing process accounts for most of the embodied carbon emissions of the entire device, and also avoids battery safety hazards in data centers.

At the software level, the team replaced the original Android user space with a general Linux distribution, thereby bypassing memory restrictions and protection mechanisms such as the “Low Memory Killer” designed for mobile devices. For hardware management, every 25 to 50 phone mainboards are divided into a self-managed cluster, and containerized application scheduling is handled via Kubernetes. Based on SPEC CPU 2017 test results, clusters made up of 25 to 50 phone mainboards have overall CPU throughput approaching that of a traditional server.

Early experiments show that a micro-cluster of 20 phones can smoothly handle submission peaks of more than 75 students’ assignments in parallel computing courses. Evaluation task completion time is about 50 seconds, and grading delays are even lower than those of AWS cloud instances. According to the team’s estimates, a system built from 2,000 phones has total compute power roughly equivalent to 50 traditional servers, and can support the teaching and research of over a hundred system programming and parallel computing courses at the same time. The entire system is expected to officially go live in the autumn of 2026, serving as a long-term running test platform to evaluate the hardware reliability of consumer-grade equipment under sustained high load.