Can 335,000 Smartphones Rival Data Centers for AI Compute?

Decentralized compute network Acurast says it has onboarded more than 335,000 smartphones, with around 40,000 fully active on an average day — a pool its co-founder claims matches the compute capacity of roughly 2,340 high-end data-center servers.

The claim, benchmarked against the ASUS RS720A-E11, a high-density server built for AI and high-performance computing, comes from Acurast co-founder Alessandro De Carli. He says the phones handle workloads ranging from web scraping and cybersecurity to AI inference and agent hosting. The push comes as markets weigh the revenue needed to repay the estimated $38 billion cost of building a single, one-gigawatt AI data center, while campaigners target the energy and water such facilities consume.

What are the phones actually good at?

Smartphones are not general-purpose server replacements, and De Carli doesn't pretend they are. He says the devices excel at confidential computing and isolated environments that install and run executables, and at CPU-intensive workloads requiring wide distribution and unfiltered internet access.

"Servers are better at multi-threading, but smartphones offer unmatched compute efficiency," De Carli says. Ordinary phones have limited memory, and large servers keep the edge when handling multiple tasks in parallel, but he argues many workloads still fit within available phone memory.

The built-in hardware security features of modern phones are the draw: secure enclaves and similar modules give distributed tasks a trust layer that commodity cloud instances don't always provide.

Who else is building this?

Another project, Phase, launched Phase Proxy in October. It lets smartphone owners provide internet connectivity and blockchain validation for AI agents. Founder Kirill Eves explains the reasoning: "websites routinely block or rate-limit automated traffic originating from data centers." Phase instead routes agents through users' mobile and residential connections.

Phase plans to charge 0.5 USDT per 1,000 requests plus 1 USDT per 1GB of traffic. A follow-up product, Phase Router, will direct simpler requests to models running on phones in the network, or route them to external providers like OpenAI and Anthropic based on "task complexity, model capabilities, performance and cost." Eves says the network could eventually expand to desktops, servers and GPUs.

The backdrop is scale: as of October 2026, there are an estimated 7.68 billion mobile connections worldwide, and billions of smartphones sit unused in drawers and on desks.

Does the research back it up?

Partially, and only in controlled settings. A 2025 study tested small clusters of smartphones on image recognition tasks and found competitive performance under certain conditions. A Google-backed project at the University of California, San Diego found that 20 retired smartphones could handle the peak computing workload of a class of more than 75 students, with grading times faster than the university's standard cloud setup.

Real-world networks face harder problems. Phones disconnect, overheat, drain batteries, switch networks and carry wildly different processing power and memory. Blockchain networks show the same constraint: Ethereum's throughput is limited by cross-network communication latency and the capability of its lowest-spec nodes. Splitting one task across several phones fails when a single connection lags; distributing independent tasks is far more straightforward.

Can the economics work?

Phone networks avoid data-center capital costs, but they still pay for connectivity, coordination and user incentives — and the reward models differ sharply.

De Carli says there is "no fixed figure" for Acurast participants: rewards vary by device uptime, hardware benchmark performance, total ACU staked and developer demand. Operators compete for a pool of inflation-funded rewards rather than receiving direct customer payments. Phase ties rewards to uptime, participation and demand. "Phase will use a portion of its service revenue to purchase PHASE on the open market and distribute it to node operators based on their contributions to the network," Eves says.

Not everyone buys the model as AI infrastructure. Markus Levin, co-founder of decentralized infrastructure provider XXYO, argues decentralized networks may fit data collection and verification better than model execution. "A useful decentralized AI network should be judged by whether it solves a problem customers will pay to fix, not by its token price or node count," he says.

Running an AI model on a phone is no longer the hard part. Coordinating tens of thousands of devices into something that behaves like a reliable cloud service — while the token economics stay solvent — is the test these networks still have to pass.