Thomas Sohmers
CEO
Positron AI
Positron AI provides a transformer inference server and custom hardware designed to run large language models more efficiently than GPU-based systems. The Atlas platform supports direct deployment of Hugging Face models with an OpenAI-compliant API, while the Asimov hardware claims 5.7M tokens per dollar versus 2.4M for leading GPUs. The company also offers Testflight, a managed inference service for remote evaluation.
- Artificial Intelligence
- Developer Tools
- Hardware
- Enterprise Software
- Software Only
Updated today
Funding
Raised to date
$1.2BRaised to date based on public sources. This may differ from the amount the company actually raised and is based only on what is publicly available on the internet.
across 4 rounds
+18$875M · Series C Tranche
Announced Sep 2026
Disclosed cumulative: $1.2B
4/4
RoundDateAmountValuationInvestors
Founders
Product
Problem
The increasing demand for large language model (LLM) inference necessitates high-performance computing solutions that can handle complex models efficiently, especially in power-constrained environments. Existing GPU-based systems often suffer from bottlenecks related to memory, processing, and high power consumption, leading to increased operational costs and limited scalability.
Solution
Positron AI offers a transformer inference server, Atlas, designed to deliver optimized performance and cost-efficiency for AI model deployment. By utilizing a hardware and software co-design approach, Atlas achieves superior performance per watt and per dollar compared to traditional GPU-based systems. The platform supports seamless integration with Hugging Face models, allowing users to deploy any trained transformer model directly onto the hardware. Positron also provides a managed inference service, Testflight, for remote evaluation, enabling efficient scaling and reduced operational expenses. The company's latest hardware, Asimov, is claimed to deliver 5.7M tokens per dollar versus 2.4M for leading GPUs, and 4.1× faster tokens at equal cost per token.
Target Audience
The primary target audience includes enterprises, research teams, and service providers seeking high-performance, cost-effective, and energy-efficient solutions for deploying and scaling large language models.
Features
- High performance per watt, exceeding that of GPU-based systems by up to 4x
- Lower cost per token compared to Nvidia DGX-H100 systems, up to 2.5x
- Seamless integration with Hugging Face Transformers Library models
- OpenAI API-compliant endpoint for easy integration with client applications
- Model Manager for uploading and linking trained model files (.pt or .safetensors)
- Managed Transformer Inference (Testflight) for remote access and evaluation
- Increased model density for power-constrained racks
- Hardware and software explicitly designed for generative and large language models (LLMs)
- Asimov hardware delivers 24.8× revenue per total cost of ownership dollar and 26× tokens per dollar compared to Blackwell-class GPUs