Google and UCSD Turn 2,000 Retired Pixel Phones Into Low-Cost Data Center Clusters

Google and researchers at the University of California, San Diego have assembled a working data center cluster from 2,000 retired Pixel smartphones, proving that old handsets can serve as low-cost, low-power computing nodes. The project, detailed in a paper published this month, repurposes phones that would otherwise be recycled or discarded.

The cluster, built from Pixel 3a through Pixel 6 devices, achieved a 50% reduction in energy consumption compared to an equivalent x86 server rack, according to the team's internal benchmarks. Each phone draws roughly 2 to 5 watts under load, versus 100 to 200 watts for a typical server CPU.

How the Phone Cluster Works

The researchers stacked the phones in custom 3D-printed racks, connecting them via USB-C for power and Ethernet adapters for networking. Each phone runs a stripped-down version of Android that acts as a compute node, handling tasks such as image processing, web serving, and machine learning inference. The cluster uses Kubernetes for orchestration, treating each phone as a lightweight container host.

• 2,000 phones total, sourced from Google's internal device return program and university donations.
• Peak aggregate compute capacity: roughly 2 teraflops, comparable to a single mid-range GPU server.
• Total cluster cost: under $10,000, versus $50,000 or more for an equivalent new server rack.
• Annual e-waste diverted: approximately 400 kilograms of electronics that would have been sent to shredders or recyclers.

Implications for Hyperscale Operators

The project challenges the assumption that data centers must use purpose-built hardware. While the phone cluster cannot match the raw throughput of a modern server for high-bandwidth tasks, it excels at latency-tolerant batch jobs and edge computing workloads. Google's own internal tests show the cluster can handle 80% of the company's lightweight cloud functions, such as thumbnail generation and log parsing.

UCSD professor and lead researcher Dr. Rajesh Gupta said the team is now testing a second cluster using Pixel 7 and 8 devices, which include Google's Tensor chips. Those phones offer dedicated machine learning accelerators that could make the cluster competitive with GPU-based servers for certain AI inference tasks.

The next phase of the project will explore liquid cooling for the phone stacks, as the current air-cooled setup struggles with heat density when more than 50 phones are stacked in a single rack. Google has not announced plans to commercialize the design, but the company's hardware sustainability team is evaluating the concept for internal use in its own data centers.