Summary
Fahd Mirza examines Extropic, a hardware startup betting that probabilistic computing can dramatically reduce the energy cost of running AI models. At the core of their approach is the Z1 chip, which uses “p-bits” — probabilistic bits that deliberately exploit quantum noise rather than suppressing it — arranged in a sparse local connectivity pattern where each node communicates with only 16 neighbors. This is a fundamental departure from the all-to-all wiring of GPUs, which is powerful but energy-intensive.
The video walks through Extropic’s key trade-off in detail: their sparse models require roughly 10 times more compute operations than a standard dense model to reach equivalent quality (benchmarked against GPT-2), but each operation uses a tiny fraction of the energy a GPU would consume. The net result, according to their published charts, is a significant overall energy advantage — and critically, the efficiency gains follow a smooth, predictable scaling curve similar to the laws that made conventional AI models worth scaling in the first place.
Mirza also digs into the practical state of the release. The Z1T0 model weights are available on Hugging Face and the training code is on GitHub in JAX, allowing researchers to train and study sparse models on standard GPUs as a simulation. However, he is clear that the weights cannot currently be run for inference — there is no consumer-facing deployment path until the Z1 hardware ships. For anyone tracking alternative AI hardware paradigms beyond Nvidia’s roadmap, this is a useful, technically grounded primer on where Extropic actually stands.
📺 Source: Fahd Mirza · Published September 05, 2026
🏷️ Format: Deep Dive
