Turning retired smartphones into AI servers? Google proposes building low-carbon computing clusters using old Pixel devices

Google Research proposes a plan to repurpose decommissioned smartphones into data center compute nodes: removing the screen and battery, leaving only the motherboard. 25 to 50 phones' combined processing power is equivalent to a modern server, with significantly lower carbon emissions than new hardware. UC San Diego has planned a cluster of 2,000 phones, expected to go online in fall 2026.
(Background: Amazon first disclosed data center water usage: 2.5 billion gallons in 2025)
(Additional context: The Information: Google plans to commission Samsung to produce the 10th generation AI chip "Icefish," diversifying TSMC supply risks)

Table of Contents

Toggle

• A single core of a foldable phone outperforms data center servers
• Modification process: keep the motherboard, dismantle everything else
• UC San Diego’s 2,000-phone experiment and unresolved reliability issues

As global data center electricity and carbon emissions become real bottlenecks for AI expansion, Google Research’s solution runs counter: instead of building new data centers or purchasing new servers, it turns your discarded old phone into a cloud node.

Google Research published this study on their official blog this week, focusing on reducing the "embedded carbon" of electronic products. Simply put, the carbon footprint from raw material extraction to factory release of a phone becomes a sunk cost once you discard it; if it can be reused for five more years, that carbon is effectively diluted.

A single core of a foldable phone outperforms data center servers

The team compared the processor in the 2023 Pixel Fold against an ASUS RS720A-E11 data center server, using industry-standard SPEC CPU2017 benchmarks. The result: the Pixel Fold’s performance core’s "single-core performance" outperformed the server’s single-core performance in most tests.

At first glance, this seems counterintuitive, but the logic is clear. Smartphone chips are optimized for power efficiency within mobile device constraints, having been refined over years; data center servers prioritize multi-core parallelism, large memory, and I/O throughput, with single-core performance not being the main focus. When running on the same benchmark, the performance advantage of mobile chips becomes evident.

In conversion, 25 to 50 old phones’ combined processing power is equivalent to a modern server. The proposed approach is to form self-managed clusters of 25 to 50 phones, containerized and orchestrated with Kubernetes. To the upper-layer workloads, the cluster behaves similarly to a cloud machine.

Modification process: keep the motherboard, dismantle everything else

Installing phones into racks isn’t just plugging the whole device in. Before deployment, researchers must remove the screen, battery, casing, camera modules, and all peripherals, leaving only the motherboard.

Here’s the key figure: according to Google’s internal carbon footprint assessment, the motherboard accounts for about 50% of the embedded carbon in a phone. Retaining only the motherboard and discarding other materials maximizes carbon reduction, as the most carbon-intensive component is still in use. Batteries are unsuitable for data center standards, and casings and screens have no computational value.

Software modifications are equally essential. Although Android is based on Linux, its user space for mobile devices includes mechanisms needed for consumer devices, such as the "low memory killer," a memory management program that kills background processes when memory is tight.

In simple terms, it’s designed to keep the phone interface smooth, but in a cloud server environment, it can interfere with normal memory allocation. The team replaced the mobile Android user space with a standard Linux distribution and disabled these consumer-oriented protection mechanisms.

UC San Diego’s 2,000-phone experiment and unresolved reliability issues

The research has empirical data. A cluster of 20 phones, during peak submission times for courses with over 75 students, showed grading delays lower than the default AWS backend. Each phone’s computing power roughly matches an AWS t3.micro instance, a small cloud instance with 2 vCPUs and 1GB RAM, sufficient to host most tasks for university EdTech, grading, and research workloads.

UC San Diego is planning a 2,000-phone compute cluster to support courses in "parallel computing" and "system programming." Once fully deployed, it can support hundreds of such courses simultaneously. The full system is scheduled to go online in Fall 2026, providing roughly the same compute power as 50 servers, at a fraction of the cost of conventional procurement.

However, the team acknowledges an unresolved issue: the reliability of consumer-grade hardware under long-term, high-load operation. Phone motherboards were never designed for continuous server workloads; data on component lifespan and failure rates over time are still needed at scale. The UC San Diego cluster of 2,000 phones is, to some extent, also a long-term data collection effort.

More pragmatically, the real significance of this experiment may not be "cost-saving" or "environmental impact," but rather challenging a fundamental assumption: that compute power must come from hardware explicitly designed for computation. If consumer-grade hardware’s single-core performance can outperform data center devices in certain scenarios, then the definition of "computing power" itself becomes more flexible.