📰 Key Takeaways

Can decentralized phone networks power AI compute? That’s the question Cointelegraph Magazine explores. De Carli says smartphones are especially efficient for certain workloads: built-in hardware security suits confidential computing, meaning isolated environments that can install and run executables, as well as CPU-heavy tasks and applications that need wide distribution and unfiltered internet access. He notes servers excel at multithreading, but phones are unmatched in compute efficiency. Phones have limited memory and large servers still win at parallel processing, yet many workloads fit in a phone’s existing memory. He claims the combined compute of active phones on the network equals roughly 2,340 high-end data center servers, benchmarked against the ASUS RS720A-E11, built for AI and HPC. Another project, Phase, launched Phase Proxy in October, letting phone users provide network connectivity and blockchain validation for AI agents. Founder Kirill Eves says websites often block or limit automated traffic from data centers, so it uses users’ mobile and residential networks as proxies, priced at 0.5 USDT per 1,000 requests and per 1GB of traffic. The next step, Phase Router, will route simple requests to models on phones in the network, or hand them to external providers like OpenAI and Anthropic, based on task complexity, model capability, performance and cost, and may expand to desktops, servers and GPUs. But XYO co-founder Markus Levin argues decentralized networks may be better at collecting and verifying real-world data than running AI models, and that data must be accurate and hard to manipulate. He stresses the test is whether a network solves problems customers will pay for, not token price or node count. The main challenge is coordinating devices so hardware, power and connectivity are as predictable as in a data center. A 2025 study showed small phone clusters can be competitive under certain conditions, but the results came from controlled test environments. See the original link for the rest.

💬 JudyAI Lab Take

This story is worth watching because it breaks the question of whether phones can serve as AI compute nodes into three angles: efficiency, memory limits, and business value, and the people interviewed don’t agree with each other.

For AI builders, this case shows a design mindset: compute doesn’t have to live only in data centers. Phase Router’s approach is to route simple requests to models on phones and send the complex ones to external providers like OpenAI and Anthropic, which is really just tiered routing. Phase Proxy, on the other hand, is betting on the mobile and residential networks of phones to get around websites blocking data center traffic, which suggests the value of decentralization may not be the raw compute at all. XYO’s Markus Levin also reminds us that the test is whether it solves a problem customers will pay for, not token price or node count. And since the 2025 study’s results came from controlled test environments, take those numbers with a grain of salt.

My suggestion: look at your own AI app and figure out which requests are actually simple, whether they can be shifted to a cheaper model, and start with a small-scale test of the cost and quality difference.

📅 Source Info

🔗 Further Reading