What Positron does
Positron (positron.ai) is an AI infrastructure company focused on accelerating Transformer model inference using purpose-built hardware and an accompanying software stack. The company’s positioning centers on improving cost and energy efficiency versus GPU-based inference, particularly for memory- and bandwidth-sensitive workloads, and on making inference hardware “GPU optional” for customers serving generative AI models.
Positron’s product line spans (1) Atlas, described as a Transformer inference server/appliance shipping for production workloads, and (2) Titan, described as a next-generation inference system intended to scale higher memory capacity per accelerator/system by using its planned custom silicon (“Asimov”). The company also describes a “memory-first” architectural approach in Asimov, including streaming vector acceleration and using commodity LPDDR5x instead of HBM in order to target cost/power and capacity advantages.
From a software integration perspective, Positron describes mapping HuggingFace Transformers models onto its hardware and using an OpenAI API–compatible endpoint (api.positron.ai). Positron also publishes developer portal release notes for Atlas software versions, including Atlas v1.1 with additional supported Llama 3.1 and Mixtral 8x7B variants.
Strategically, Positron has progressed from an initial prototype and public demo (stated as April 2023 founding; NeurIPS 2023 demo; Atlas shipped in Aug 2024) to deploying large numbers of Atlas racks (including a stated “50+ racks deployed at Oracle” milestone in Aug 2026) and raising substantial late-stage capital (Series A in 2025, Series B in Feb 2026, and Series C in Sep 2026) to fund continued scaling of Asimov silicon and Titan systems.