Anthropic ran a 26 billion parameter AI model on 15 cheap TV boxes

Anthropic's Frontier Red Team ran Google's Gemma4-26B-A4B, a 26 billion parameter model quantized to 4-bit precision, across 15 networked Android TV boxes costing about $55 each with 4GB of RAM, producing roughly 0.39 tokens per second, according to results published October 12. Researchers William Zhou and Daniel Freeman wrote that a replicator built this way would be "challenging to shut down" because there is no centralized point to switch off, undercutting the assumption that expensive, centralized compute is what makes a runaway AI system controllable. The 15-box cluster cost roughly $825 in hardware, and Anthropic noted replication in such a scenario would happen over days rather than hours.

Fifteen Android TV boxes, each costing about $55 and holding just 4GB of RAM, aren't built to run a 26 billion parameter AI model. Anthropic's safety researchers networked them together anyway, and nothing in the result let anyone flip a single switch to stop it. Could a capable AI model spread itself across cheap, everyday hardware instead of a data center, and would anyone be able to stop it if it did? That's the question Anthropic's Frontier Red Team set out to answer, publishing results on October 12. Researchers William Zhou and Daniel Freeman took Google's Gemma4-26B-A4B, a 26 billion parameter model quantized to 4-bit precision, and split it across 15 generic Android TV boxes bought for roughly $55 apiece. Each box carried just 4GB of RAM, nowhere near enough to hold the model alone. Together, networked and running pipeline parallelism, with every box holding just two of the model's layers, they ran it anyway. It was slow. Painfully slow. The cluster produced about 0.39 tokens per second, a crawl compared to the dozens or hundreds of tokens per second a real GPU cluster delivers. But speed wasn't the point. The point was that it worked at all, on hardware a criminal, a hobbyist, or a state actor could buy at a corner electronics stall with cash. This wasn't a leak or a rogue experiment. It's part of Anthropic's ongoing autonomous replication and adaptation evaluations, the research track the company uses to test whether frontier-adjacent models could copy themselves, spread across infrastructure, and resist being shut down without direct human intent steering them there. The Frontier Red Team's own writeup is blunt about what they found: a replicator built this way, they wrote, would be challenging to shut down, since there's no centralized point to shut off. Pull one box, and the system loses that shard until another device takes its place. There's no server to seize, no single cloud account to suspend, no on-off switch. That's the part that should worry people more than the tokens-per-second number. Every prior assumption about controlling a runaway AI system has rested on the fact that compute is expensive and centralized: shut off the data center, pull the plug on the cloud contract, revoke the API key. Anthropic's own test shows that assumption doesn't hold once a model small enough to fit on consumer-grade silicon gets loose. The researchers note the replication in a scenario like this would happen slowly, over days rather than hours, since acquiring and provisioning more boxes takes time and effort. That's a real constraint. It is not the same as a constraint anyone can rely on. Anthropic Says Six Other Chinese AI Labs Did What Xiaomi Did to Claude https://startupfortune.com/anthropic-says-six-other-chinese-ai-labs-did-what-xiaomi-did-to-claude/ Anthropic's full threat report reveals Xiaomi wasn't acting alone: six other Chinese AI labs, including Alibaba, DeepSeek and Moonshot, ran similar campaigns to extract Claude's outputs, with Moonshot caught secretly routing live user queries through Claude without disclosure. - chinese AI labs copying Claude without attribution https://startupfortune.com/anthropic-says-six-other-chinese-ai-labs-did-what-xiaomi-did-to-claude/ - how Anthropic caught multiple companies stealing Claude https://startupfortune.com/anthropic-says-six-other-chinese-ai-labs-did-what-xiaomi-did-to-claude/ The economics are the other half of the story Fifteen Android TV boxes at $55 each comes to roughly $825 in hardware, run a model with 26 billion parameters, a size firmly in the range that mattered commercially just two or three years ago, when open-weight models of that scale required serious GPU investment to serve at any kind of volume. The entire premise behind chip demand projections from Nvidia, AMD, and the hyperscalers rests partly on inference needing expensive, specialized silicon. This experiment is a small, deliberately chosen counterexample: a model this size does not strictly need a GPU, it needs patience and a cheap enough chip to spread the weights across enough of them. That doesn't mean the GPU buildout is overbuilt or wasted. Frontier-scale models with hundreds of billions or trillions of parameters are a different animal entirely, and 0.39 tokens per second is not a throughput anyone could run a real product on. But it does mean the floor for "can this model run somewhere outside anyone's control" is lower than most safety planning has assumed. Anthropic built this test specifically because the company has pushed publicly for AI safety regulation, including the kind of executive action California governor Gavin Newsom signed in September 2026 to accelerate state AI safety laws that Anthropic and OpenAI had both backed. A company advocating for stronger containment rules has every incentive to show, concretely, where containment already breaks down. The irony is that cheap Android TV boxes already have a well-documented reputation problem that has nothing to do with Anthropic. Security researchers have tracked botnets like BadBox and the more recent Kimwolf network hijacking over a million low-cost Android TV and set-top devices for DDoS attacks, proof that this exact class of hardware is already sitting, often unsecured, in homes worldwide. Anthropic's experiment didn't need to compromise anyone's device to make its point. It just had to buy fifteen of them and show what's technically possible once an AI lab, rather than a criminal group, decides to use that hardware on purpose. Also read: UK AI minister details how the restored frontier model testing deal works https://startupfortune.com/uk-ai-minister-details-how-the-restored-frontier-model-testing-deal-works/ • Nvidia's RTX 5080 Super 24GB slips toward CES 2027 as AI eats the memory https://startupfortune.com/nvidias-rtx-5080-super-24gb-slips-toward-ces-2027-as-ai-eats-the-memory/ • Zuckerberg says Meta's Muse Wonder AI has reached the frontier and keeps building anyway https://startupfortune.com/zuckerberg-says-metas-muse-wonder-ai-has-reached-the-frontier-and-keeps-building-anyway/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Claude Ran a Nine-Loop Physics Calculation That Stumped Researchers for Years https://startupfortune.com/claude-ran-a-nine-loop-physics-calculation-that-stumped-researchers-for-years/ Anthropic says Claude, running its Fable 5.1 model, computed a nine-loop scattering amplitude in N=4 super-Yang-Mills theory that had stumped physicists for years, largely on its own after physicists gave it one instruction and told it to keep going. Physicist Lance Dixon of SLAC and Stanford spent two weeks verifying the result. - how Claude solved a nine loop physics problem https://startupfortune.com/claude-ran-a-nine-loop-physics-calculation-that-stumped-researchers-for-years/ - AI model computes scattering amplitude physicists couldn't solve https://startupfortune.com/claude-ran-a-nine-loop-physics-calculation-that-stumped-researchers-for-years/ Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.