Controlled Phone Clusters Handle Grading and Image Recognition in Tests

Tests of controlled phone clusters have handled specific computing jobs, but the results do not show that smartphones spread across the internet can replace centralized infrastructure for large-scale artificial intelligence.

At the University of California, San Diego, an organized set of 20 retired phones processed peak grading requests for a class with more than 75 students. Grading latency was lower than with the default Amazon Web Services backend. The test used devices arranged for a local computing job; it did not evaluate a pool of independent phones communicating across the internet.

The project planned a 2,000-phone deployment for fall 2026, with capacity for 100 similar classes. The results described cover the 20-phone setup.

Acurast announced Sept. 29, 2026, that its network included more than 318,000 phones in more than 175 countries. Its provider instructions call for a stable internet connection and a power source. The Lite option runs on an everyday phone during idle periods, such as when it is charging; the Core option uses a dedicated Android phone reset to factory settings. The network count does not establish that Acurast phones performed the workloads in the university test or research study.

In controlled tests published in May 2025, smartphone clusters performed competitively on some image-recognition workloads. Its tests used strong 5 GHz Wi-Fi, devices with screens off and limits on network activity from other apps. Those controlled settings make the results specific to the tested conditions.

Servers have more processor cores and memory than phones. Phone-based systems need workloads that fit their smaller capacity, and the tests do not establish performance across a widely distributed network.

“Servers are better at multi-threading, but smartphones offer unmatched compute efficiency,” said Alessandro De Carli, Acurast’s co-founder. The 2,000-phone deployment was planned for fall 2026.