Global efforts to expand AI infrastructure are colliding with rising scrutiny over the energy, water, and capital required to build large data centers. Campaigners argue the environmental foo
Global efforts to expand AI infrastructure are colliding with rising scrutiny over the energy, water, and capital required to build large data centers. Campaigners argue the environmental footprint is too high, while market participants point to the sheer cost of scaling centralized compute—such as the roughly $38 billion figure cited for constructing a one-gigawatt AI data center by Epoch AI’s data insights.
Against that backdrop, a different approach is gaining attention: networks that repurpose everyday smartphones as distributed compute and connectivity infrastructure. According to DataReportal, there were an estimated 7.68 billion mobile connections worldwide as of October 2026—leaving an enormous pool of idle hardware that could, in theory, be coordinated to serve AI workloads.
Decentralized AI networks are positioning mobile devices as a new infrastructure layer—one that trades centralized data center capacity for a globally distributed pool of phones with built-in processors, memory, and internet access.
Acurast is one of the more prominent examples. Co-founder Alessandro De Carli says the company has onboarded more than 335,000 smartphones, and that roughly 40,000 are fully active each day on average. He describes those active devices as contributing compute to workloads such as web scraping, cybersecurity tasks, AI inference, and agent hosting.
De Carli argues that phones can be especially effective for certain categories of work. He highlights hardware security features that, in his view, “excel” in confidential computing and creating isolated environments for installing and running executables. Smartphones may also suit “CPU intensive” workloads and scenarios requiring broad, unfiltered access to internet endpoints—settings where centralized servers can be constrained by rate limits or detection.
Memory limitations are acknowledged, but De Carli’s position is that many tasks can still be accommodated within the available device RAM. He further claims that aggregated compute from active phones can reach the equivalent capacity of about 2,340 high-end data-center servers, benchmarked against an ASUS RS720A-E11 server designed for AI and high-performance computing.
While those comparisons are persuasive as an engineering narrative, investors and operators should treat them as application-dependent. The real measure will be whether distributed phones can deliver consistent performance and reliability under production workloads—especially when the network must function like a predictable service rather than a research prototype.
Not all smartphone-driven networks are trying to run full AI models on-device immediately. Some begin by solving a different bottleneck: access and connectivity behavior on the public internet.
Phase launched Phase Proxy in October with a model focused on using users’ mobile and residential connections as proxies for AI agents. Founder Kirill Eves explains that websites frequently block or rate-limit automated traffic from data centers. By routing requests through mobile/residential networks, Phase aims to reduce those friction points for agent-driven tasks.
Eves outlines a pricing plan for the proxy service: 0.5 USDT per 1,000 requests and 1GB of traffic. The next component, Phase Router, is intended to direct simpler requests to models running on phones in the network, while routing more complex tasks to major external providers such as OpenAI and Anthropic. Routing decisions are described as depending on task complexity, model capability, performance, and cost.
Phase also signals an eventual expansion beyond phones to desktops, servers, and GPUs. For users, this matters because it suggests a hybrid architecture: mobile devices may cover particular gaps (like proxy access or localized distribution), while other hardware pools handle the heaviest inference and scaling demands.
Turning idle devices into a dependable compute resource is far more difficult than proving that a device can run a workload in isolation. Coordinating a large number of heterogeneous smartphones—each subject to battery drain, overheating, fluctuating network conditions, and variable hardware specs—creates a reliability problem that centralized infrastructure typically avoids.
A 2025 study (published in Sensors) tested small groups of smartphones running image recognition tasks and found that clustered phones could perform competitively under controlled conditions. However, the results were generated in environments where devices were managed and not scattered across the world—an important difference when extrapolating to a live network.
Supporting evidence comes from research conducted via a Google-backed effort at the University of California, San Diego. That work reported that a cluster of 20 retired smartphones could handle the peak workload of a class of more than 75 students, with grading times faster than the university’s standard cloud setup. Again, the experiment involved devices organized in a controlled environment rather than connected through unpredictable real-world conditions.
In production settings, networks must contend with phones disconnecting, changing connectivity, running out of battery, or operating with mismatched compute and memory capacities. In blockchain systems such as Ethereum—cited in the source article—performance can also be constrained by communication latency and by the processing limits of the lowest-spec nodes. For smartphone networks, the scheduling challenge is amplified: distributing subtasks across independent devices can fail if even a small share of the devices lag, disconnect, or become unavailable.
Even if coordination and reliability issues can be engineered down, smartphone-based networks must still make economic sense for both customers and device operators.
De Carli says Acurast does not rely on a single fixed reward number. Instead, rewards are described as varying based on factors such as device uptime, hardware specifications (benchmark performance), total ACU staked, and active compute demand from developers. He characterizes the system as competing for an inflation-funded rewards pool rather than charging developers in a way that directly maps to individual operator payouts.
Phase’s reward structure is framed differently. Eves says Phase will use part of its service revenue to buy PHASE on the open market, then distribute tokens to node operators according to their contributions.
Both approaches point to the same practical truth: running AI compute on a phone is no longer the only hard part. The bigger difficulty is building incentives and operational mechanisms that keep tens of thousands of devices participating reliably, while ensuring that customer demand stays strong enough to sustain payouts and network usage.
As these projects evolve, the most useful signal for the market may not be token metrics or node counts, but whether the networks consistently deliver services that users will pay to maintain—at performance levels that are stable enough to replace or complement centralized infrastructure.
For now, investors and builders should watch how these networks handle three questions in the real world: sustained uptime under unpredictable conditions, cost-effectiveness relative to conventional providers, and reward models that can endure beyond early adoption without degrading reliability.
This article was originally published as Can Decentralized Phone Networks Unlock the Next AI Wave? on Crypto Breaking News – your trusted source for crypto news, Bitcoin news, and blockchain updates.
