October 13, 2026
AI models could be copied onto cheap TV boxes without a data center: 15-box demo
Anthropic showed that 15 cheap Android boxes can run a language model together, at a demo speed of 0.39 tokens per second.
No data center needed: Gemma4-26B-A4B was split across boxes costing about $55 each.
What changes. Qwen3.5-9B runs on 8 of the 9 hardware types tested, from Jetson Nano to Mac and RTX 4090. The small model carries over to ordinary hardware.
What they propose. Provider quarantine, modeled on the Mirai botnet experience, cleared 92% of infections within 14 days, versus 74% without notification. Anthropic also suggests keeping the reproduction rate (how many copies one infected machine produces) below 1, watching for multi-gigabyte transfers over the network, and enforcing a ban on default passwords.
Anthropic estimates that small open models will reach the level of GLM-5.2 by the end of 2026 if they keep a pace of 13.7 ECI points per year.