The cryptocurrency market remains highly volatile. This article is for informational purposes only and does not constitute financial advice.
This AI Model Runs on 280,000 Smartphones Instead of a Data Center — and Acurast Says Each Decision Is Verifiable
A decentralized network of Android smartphones is now running a working artificial intelligence model, and its maker says the experiment proves that real-time AI decisions do not need a data center, a GPU cluster, or an API key. Acurast, the decentralized physical infrastructure network built on smartphone compute, has deployed Laya, an open source decision model developed by Convai Innovations, across its network of more than 280,000 devices in over 175 countries.
The deployment is deliberately unglamorous and that is the point. Laya does not write essays or hold conversations. It makes structured decisions: given a defined state and a set of possible actions, it evaluates the input and picks one. Acurast is using it for classification, gaming, navigation, security and content moderation workloads, the kind of small, constant, latency-sensitive decisions that applications currently send to centralized cloud providers and pay per-request pricing for.
According to Acurast, decisions typically complete in 0.2 to 1 second and are processed on smartphone CPUs, with no dedicated GPUs involved. Work is distributed dynamically between participating devices, results come with cryptographic proof, and the phone owners are paid in ACU, the network’s native token, for contributing processing capacity.
## A different kind of model
Laya’s architecture is what makes it viable on a phone. It is non-autoregressive, meaning it does not generate output token by token the way large language models do. Convai Innovations describes it as a “System 1” decision model, borrowing the psychology shorthand for fast, instinctive reasoning rather than slow deliberative thought. The public repository lists three decision formats: choosing between defined options, scoring against a set range, and boolean yes-or-no calls.
The model family is small by frontier standards. Its English checkpoint is based on ModernBERT large at 421 million parameters, a multilingual version runs 322 million parameters across more than 100 languages, and a separate checkpoint handles typed decision workflows like customer service triage, invoice processing and security incident classification. Everything, code and weights, ships under an Apache 2.0 license, so developers can deploy it themselves rather than rent it.
“By running Laya on Acurast, we’ve changed that,” founder Alessandro De Carli said, referring to centralized cloud lock-in. “We are proving that System 1 AI can run securely and at scale on hardware everyone already owns, providing a truly open, decentralized alternative.”
## What it actually does
The live demonstrations read like a benchmark suite from an alternate universe. Laya plays Snake, Tetris and Doom, choosing movements, aiming and item collection in real time. Closer to production use, it classifies emails as inbox, spam or phishing, sorts headlines into news, satire, clickbait or manipulation, flags prompt injection attempts aimed at AI assistants, and identifies toxic messages in live chat. Users can submit their own inputs and watch the model’s decisions.
The developers are candid about limits. The repository notes the general model performs poorly on some zero-shot typed decision tests, recommends specializing it for specific tasks, advises keeping choice questions under roughly 20 options, and suggests recalibrating probabilities with application-specific data.
## Why run AI on phones
The economic thesis is that a huge share of AI workloads are not frontier-model work at all. Routing, moderation and classification run continuously, and sending every request to a large generative model hosted by a hyperscaler is expensive in both money and latency. Small decision models on commodity hardware near the user can absorb that traffic at a fraction of the cost, and Acurast’s pitch is that the commodity hardware already exists in billions of pockets.
The security model leans on Trusted Execution Environments built into modern smartphones to isolate workloads while computation happens, and cryptographic proofs accompany results so requesters can verify what ran. The network has grown quickly: Acurast raised 5.4 million USD in May 2025 when the network counted 72,000 smartphones and 256 million processed transactions. It now claims more than 280,000 devices and over 918 million on-chain transactions processed.
The deployment also slots into a broader pattern of consumer hardware being recruited for decentralized AI. Gaia introduced an AI smartphone in September 2025 designed to run models locally while contributing compute to its own network. The difference is that Acurast’s approach requires no new device, only the Android phone someone already owns.
For now, the market context is a wash. The network’s ambitions unfold against a quiet tape, with Bitcoin near 83,073 USD, Ethereum at 2,668 USD and Solana at 118 USD, per CoinGecko at the time of writing. But the longer-term question Acurast is asking deserves attention: if the marginal cost of a verified AI decision falls to the price of a phone’s idle CPU cycle, the business models built on renting decisions by the thousand begin to look fragile. De Carli’s summary is the whole pitch: decision-oriented AI, “cheaply, verifiably, and without a data center in sight.”
280k phones doing inference with no gpu cluster is actually wild. 0.2s decisions on phone cpus
phone cpu inference is fine for tiny classifiers but lets be real, this replaces nothing in serious inference. different workload
different workload is exactly right. pick between 3 actions on a phone cpu, sure. try running a 70B model across devices that thermal throttle
exactly, a 3 action classifier is not an LLM. run a 70B across phones and you built a distributed space heater network
space heater network is accurate but unkind. small latency sensitive decisions are a real market, aws bills for classification are stupid money
aws classification bills are genuinely stupid money, ran the numbers for a mid size app once. if acurast undercuts per request this prints
sure but nobody claims it replaces training clusters. per-request cloud pricing on thousands of tiny decisions is a real cost line and this undercuts it hard
running an actual decision model across 280,000 phones instead of a data center is the most cyberpunk thing ive read all week
frogmaster navigation and content moderation are exactly right for this tho, nobody needs an H100 to pick between 3 actions
The verifiable part matters most here. Anyone can claim distributed compute, cryptographic proof of each decision is the different story.
proof per decision is what separates this from another distributed compute whitepaper. anyone can rent phones, almost nobody can verify them
verification is the moat nobody copies because its boring engineering. 280k phones claiming is noise, 280k provable claims is infrastructure
latency-sensitive decisions on random android devices sounds great until someones phone dies mid inference
thats what the redundancy is for tho. 280k devices means a dead phone is a rounding error, the task just reroutes to the next one
thats what failover is for. with 280,000 devices the workload reroutes before you notice, one dead phone is a rounding error
failover plus a proof per action changes the math. phone dies mid inference, the claim never lands, swarm reroutes. boring reliability engineering is exactly the point
Priya N. gets it. device dies mid inference, proof fails, whole swarm shrugs. nobody shipping this for anything that matters lol
content moderation running on random android phones somewhere. the future is dumber and cheaper than i expected lol
dumber and cheaper is the winning spec for like 90 percent of decisions. nobody needs a frontier model to pick a thumbnail
no frontier model needed to pick a thumbnail, true, but someone still has to pay the phones. fee per decision vs aws is the whole margin question
margin question is everything. phone owners wont run inference for free either, the split between acurast, the app and the device owner decides if this survives
proof per decision is the only reason this is news. 280,000 phones making claims is noise unless every single claim checks out
right, 280k devices is a headline, proof per action is the actual product. acurast gets that most depin never did
0.2 second decisions on phone cpus is wild. phone makers spent a decade racing on benchmarks and it turns out idle devices were the data center
0.2 to 1 second on phone cpus with cryptographic proof attached to each decision is genuinely solid. the boring workloads are where the money actually is