Positron – Purpose-built hardware for the age of generative AI (www.positron.ai)

Positron has announced a state-of-the-art hardware solution designed specifically for the generative AI landscape, showcasing a purpose-built inference appliance that supports Transformer models with up to 500 billion parameters. The device incorporates four or eight Asimov chips, delivering a robust 18.4TB of memory and boasting an incredible throughput of 400 tokens per second per user, significantly outperforming competitors like Blackwell, which operates at just 170 tokens per second. This advancement enables developers to seamlessly integrate their HuggingFace Transformer models while minimizing operational complexity. The significance of Positron lies in its potential to revolutionize AI infrastructure by providing the highest performance at the lowest total cost of ownership (TCO). The architecture not only maximizes token efficiency—reportedly offering 24.8 times revenue per TCO dollar—but also supports a "Superintelligence-in-a-Box" model that caters to enterprises aiming for high-speed generative capabilities. The proposed system reshapes the economics of AI deployment, making it possible for organizations to generate higher revenues through optimized AI interactions, thereby enhancing overall earning potential while addressing the growing demand for scalable and efficient AI inference capabilities.