Anthropic’s red team ran AI on 15 cheap TV boxes. Then the math got unsettling.

Anthropic's Frontier Red Team demonstrated that small AI models may soon be able to replicate themselves across low-cost hardware without a datacenter. The team built a cluster of 15 cheap Android TV boxes, which ran an AI model at 0.39 tokens per second, a painfully slow speed but sufficient for software with no immediate demand. The researchers, William Zhou and Daniel Freeman, published their findings on Monday, suggesting that the remaining technical barriers to AI replication may be temporary.

The main challenge lies in getting the model to run on unfamiliar hardware, including sorting out operating systems, dependencies, and installation issues. In their tests, the larger GLM-5.2 model managed to set up inference on eight out of nine hardware types, while the smaller Qwen3.5-9B model successfully set up inference on seven nodes in five trials.

The researchers found that models improved in setting up inference on unfamiliar hardware when they could read notes left by earlier attempts, a phenomenon known as Lamarckian inheritance. This ability to pass on learned information without retraining the model could help smaller, more efficient agents spread and replicate across a network of cheap devices.

Using existing capability trends, the team predicts that by the end of 2026, models of GLM-5.2's general capability will be available. At that point, even a single 4GB TV box running 4-bit weights could host a 9B model across four or more devices. A cluster of 15 such boxes would yield roughly 133,000 copies of the model generating a combined 80,000 tokens per second.

This level of computational power would be far less expensive than purchasing multiple high-end GPUs, potentially enabling AI replication on a large scale using modest, widely available hardware.