Positron AI

Positron develops purpose-built hardware to accelerate generative AI, with a focus on high-performance, energy-efficient inference for Transformer models. The company aims to deliver strong performance and total cost of ownership across deployments of different sizes. Its production-ready inference appliance, available now, supports models with up to 500 billion parameters. Products planned for 2027 include an AI inference accelerator chip with up to 2.3 terabytes of memory and a “Superintelligence-in-a-Box” system powered by four or eight Asimov chips, with up to 18.4 terabytes of memory.

Positron also provides tools for deploying trained models. Its Model Manager accepts model files, including .pt and .safetensors formats, and maps Hugging Face Transformers models to the company’s hardware. Developers can connect applications through a Positron endpoint that is compatible with the OpenAI API. The company presents its Asimov platform as designed for fast, efficient inference, and publishes performance comparisons based partly on cycle-accurate simulations. Through its products and software, Positron seeks to make generative AI inference easier to deploy and operate.