Positron

Positron is a San Jose-based semiconductor company developing specialized AI inference chips designed to run large language models more cost-effectively and energy-efficiently than general-purpose GPUs.

Positron was founded to address the rapidly growing cost and energy consumption associated with running, or "inferencing," already-trained large language models in production to serve the enormous and continuously growing volume of user queries that popular AI applications and chatbots now handle, an operational expense that has grown to represent one of the largest ongoing costs for companies deploying AI at scale, arguing that general-purpose graphics processing units (GPUs), while highly effective and currently dominant for AI model training, are not necessarily the most cost- or energy-efficient hardware choice specifically for the different computational profile of running already-trained models to serve inference requests in production. The company's chip architecture is designed specifically around the computational patterns involved in large language model inference, aiming to offer meaningfully better cost efficiency and energy consumption per inference request than using general-purpose GPUs originally designed with a broader range of computational workloads, including both AI training and traditional graphics rendering tasks, in mind, an increasingly significant argument as AI inference costs and the electrical power consumption of AI data centers have grown into major operational and even broader energy infrastructure concerns for companies operating at the scale of serving popular consumer AI applications and enterprise AI services to millions of users. Positron has positioned itself within a growing category of specialized AI inference chip companies emerging specifically to address this large language model inference cost and efficiency challenge, competing both against incumbent GPU maker Nvidia's continued architectural improvements aimed at inference efficiency, and against other emerging inference-focused chip startups pursuing broadly similar strategic bets that a chip architecture purpose-built for inference, rather than a more general-purpose architecture adapted from training-focused designs, can offer meaningful cost and efficiency advantages for this increasingly significant category of AI computing workload. The company has attracted venture investor interest reflecting the broader intensifying competitive landscape around specialized AI inference hardware as inference costs have become an increasingly significant and closely scrutinized expense for companies operating large-scale AI applications and services.

News

No news available for this company.