Positron AI

What Positron AI does

Positron AI delivers vendor freedom and faster inference for enterprises and research teams by providing hardware and software designed for generative and large language models (LLMs). Their solutions offer lower power usage and drastically reduced total cost of ownership (TCO), enabling the running of popular open-source LLMs at high token rates and long context lengths. Positron is also developing its own ASIC to support training and other parallel compute workloads, in addition to inference and fine-tuning. The company aims to make advanced machine learning accessible and efficient, with leading performance per dollar and energy efficiency, and their products are designed, fabricated, and assembled in the United States.

Products

• AtlasA production-ready inference appliance supporting up to 500B Parameter Models. It is the world's first LLM-inference-first accelerator, built and manufactured in the U.S., designed for efficiency to address cost and energy constraints in the AI sector.
• Asimov chipsPurpose-built AI Inference accelerator silicon with 2TB+ memory per chip, currently in development and expected to be part of future products.
• Titan systemsA second-generation system currently under development, aiming to be even faster and more efficient than Atlas, leveraging insights from Atlas to deliver a near limitless inference solution.
• Superintelligence-in-a-BoxA future product coming in 2027, featuring 8TB+ memory, powered by 4x Asimov chips.

Key capabilities

• Highest performance for Transformer model inference
• Lowest power consumption for Transformer model inference
• Best TCO for Transformer model inference
• Supports all HuggingFace Transformers Library models seamlessly
• OpenAI API-compliant endpoint for client applications
• Lower power usage for LLMs
• Drastically lower total cost of ownership (TCO)
• Ability to run popular open source LLMs to serve multiple users at high token rates and long context lengths
• Designed, fabricated, and assembled in the United States
• 3x lower end-to-end latency for trading inference workloads versus comparable H100 systems
• Consumes 1/3rd of the power compared to comparable H100 systems

Use cases

• Generative AI acceleration
• Large language model (LLM) inference
• Fine-tuning LLMs
• Training LLMs
• Parallel compute workloads
• Trading inference workloads

AI approach

Positron AI develops purpose-built hardware, including an ASIC, for accelerating generative AI and large language models (LLMs) inference. Their products aim to provide higher performance, lower power consumption, and better total cost of ownership compared to traditional GPUs. They support all Transformer models and allow users to upload HuggingFace Transformers Library models directly onto their hardware.

Tech named: generative AI, large language models (LLMs), Transformer models, ASIC, HuggingFace Transformers Library, OpenAI API-compliant endpoint, LLama-2 7B

Industries served

• Networking
• Gaming
• Content Moderation
• CDN
• Token-as-a-Service

What it says sets it apart

• Purpose-built hardware and software designed from the ground up for generative and large language models (LLMs)
• Vendor freedom
• Faster inference
• Lower power usage
• Drastically lower total cost of ownership (TCO)
• Enables running popular open source LLMs for multiple users at high token rates and long context lengths
• Proprietary ASIC development for expanded capabilities (training, parallel compute)
• Focus on making GPUs optional for inference
• Designed, fabricated, and assembled in the United States
• Significant performance and efficiency gains over traditional GPU systems (e.g., 3x lower latency and 1/3rd power consumption vs. H100 for trading inference)

Funding rounds we track

From the AI funding tracker — rounds as reported by the linked publications.