Positron, an AI chip startup, has surpassed a $1 billion valuation by developing inference accelerators with a unique memory-centric architecture that dramatically increases on-chip memory and efficiency compared to competitors like NVIDIA. Backed by $230 million in new funding and early customers such as Jump Trading, Positron aims to scale production and become a major player in AI hardware by addressing memory bottlenecks in large-scale inference tasks.
Positron, a new AI chip startup, has recently achieved a valuation exceeding $1 billion, drawing attention for its innovative approach to inference accelerators. The company’s core differentiator lies in its focus on memory architecture, specifically addressing the memory bottlenecks that limit inference performance. Unlike many competitors, Positron is building chips with memory placed much closer to the processing arrays, resulting in significantly faster data access and decode speeds. Their upcoming second-generation chip will feature an unprecedented 2.3 terabytes of directly attached memory, far surpassing NVIDIA’s forthcoming Rubin chip, which will offer 384 gigabytes.
The company’s architecture is designed to meet the demands of large-scale AI inference tasks, such as video generation, code generation, and advanced reasoning models, all of which require both high memory capacity and low latency. Positron’s approach reduces the need for scaling out across multiple chips, which is the strategy NVIDIA is pursuing with optical interconnects at the data center level. By integrating massive memory directly onto the chip, Positron enables more efficient processing of extremely large models, consuming less power and requiring fewer chips for the same workload.
Positron recently raised $230 million in funding, which will be used to accelerate development and production. While the company acknowledges that it cannot compete with NVIDIA’s vast financial resources, it believes its fundamentally different technology gives it a unique position in the market. The focus is on optimizing for power consumption, cost per token or video generation, and output speed, with each silicon architecture finding its own niche. Positron aims to carve out a space where its memory-centric design delivers superior efficiency for specific inference applications.
The company’s first-generation chip has already found customers, notably Jump Trading, which became both a user and an investor after testing the product and reviewing the roadmap. This validation from a major financial player encouraged Positron to pursue its latest funding round. The customer base is diverse, including not only financial firms but also ecosystem partners like ARM and sovereign data center operators, all interested in deploying heterogeneous silicon for inference workloads.
Looking ahead, Positron’s primary goal is to scale production and deploy millions of chips worldwide, aspiring to reach the scale of industry leaders like NVIDIA. While acquisition or going public are possible future outcomes, the company’s immediate focus is on technical execution and customer adoption, particularly among hyperscalers who drive the majority of global inference workloads. By delivering chips that address the critical bottlenecks in AI inference, Positron aims to become a key player in the rapidly evolving AI hardware landscape.