UCSD builds private cloud from 2,000 retired Pixel phones

A team of computer scientists at the University of California, San Diego (UCSD) is preparing to launch a private cloud built from 2,000 decommissioned Google Pixel Fold smartphones. The project, developed in collaboration with Google, aims to demonstrate how retired consumer devices can serve as a cost-effective and environmentally friendly computing platform for academic and research workloads.

The initiative originated with Jennifer Switzer, a former UCSD PhD student now working as a postdoctoral researcher at Google. Switzer initially tested the concept with small-scale clusters before scaling up to the current deployment. Google estimates that the average smartphone is replaced every four years, leaving behind devices with significant remaining computational capacity but limited reuse options beyond recycling or resale.

How the cluster works

The Pixel Fold smartphones powering the cluster feature Google’s Tensor G2 processors, which include two 2.85 GHz Cortex-X1 cores, two 2.35 GHz Cortex-A78 cores, four 1.80 GHz Cortex-A55 cores, a Mali-G710 MP7 GPU, and 12 GB of RAM. Early benchmarking using the SPEC suite indicates that 25-50 phones can deliver performance comparable to a conventional server. However, the primary challenge lies in efficiently distributing workloads across thousands of devices, each with limited memory and heterogeneous cores.

To address this, the team is pursuing two approaches. The first targets applications that can run within the constraints of a single device, such as grading systems or lightweight cloud functions. The second leverages Kubernetes to orchestrate containerized workloads across clusters of 25-50 phones. The devices are being reflashed with a Linux operating system optimized for server use, stripping out Android’s battery-saving features that would otherwise limit performance in a datacenter environment.

Physical deployment required significant modifications to the smartphones. Early testing revealed that intact devices posed fire risks due to their lithium-ion batteries, prompting Google to partner with a third party to extract the motherboards. The team is also developing custom PCBs to provide power and wired Ethernet connectivity, replacing the phones’ original wireless networking capabilities. This setup ensures both safety and scalability for the cluster.

Why it matters

The project highlights the potential for repurposing consumer electronics to reduce e-waste and lower the carbon footprint of computing. Google estimates that the motherboard accounts for roughly half of a smartphone’s embodied carbon, meaning reusing these components could significantly cut emissions compared to manufacturing new hardware. The cluster will initially support UCSD’s San Diego Supercomputing Center, with plans to expand access to other research teams if the pilot succeeds.

Beyond sustainability, the initiative offers a practical testbed for parallel computing and systems programming. Researchers compare the setup to the Beowulf clusters of the 1990s, which repurposed consumer PCs into supercomputers. While the Pixel cluster won’t rival traditional HPC systems, it could provide a viable alternative for workloads that don’t require high-performance interconnects or massive memory pools.

What to watch

The cluster is scheduled to go live in fall 2026, with initial workloads focused on EdTech, grading, and function-as-a-service applications. If successful, the project could expand to include more devices or additional use cases, such as edge computing or distributed AI inference. The team also plans to explore whether other retired consumer devices, like tablets or laptops, could be integrated into similar clusters.

For now, the project remains an academic proof of concept, but its outcomes could influence how organizations approach hardware reuse in datacenters. The broader industry has already experimented with unconventional compute clusters, such as UC Santa Barbara’s Raspberry Pi cluster and Gigabyte’s compact Lunar Lake-based system, suggesting growing interest in alternative architectures for specific workloads.