What Positron does
Positron is a U.S.-fabricated AI inference hardware company focused on making transformer inference more cost- and energy-efficient than GPU-based approaches, and on enabling large-context, memory-bound inference workloads. The company ships “Atlas” as a transformer inference server/inference accelerator system today, and is building next-generation custom silicon and systems—“Asimov” (custom AI accelerator chips) and “Titan” (a next-generation inference system designed for very large memory and context)—as the roadmap to replace more of the traditional GPU stack with proprietary components over time. Positron positions its differentiation around memory-first architecture and performance-per-watt/per-dollar improvements, targeting data centers and inference operators that are constrained by power availability, heat/power densities, and memory/capacity limits rather than only raw compute throughput. The company also emphasizes compatibility and deployability via an inference software stack and the use of standard model ecosystems, with Atlas exposed via Positron’s inference engine and supported by a developer portal for OpenAI-compatible inference endpoints.