Fractile Raises $220 Million to Scale 1,200‑Token‑Per‑Second AI Chips

Fractile was founded in 2022 by Walter Goodwin and is headquartered in London. The company designs hardware specifically to relieve memory bandwidth bottlenecks that limit the speed of large language models. By focusing on inference rather than training, Fractile aims to deliver faster responses for AI applications that require massive sequential reasoning. The company’s architecture targets the memory wall, a common limitation where data movement between processors and memory becomes the bottleneck for AI workloads. By redesigning data pathways, Fractile aims to keep the compute units fed with information at high rates.

The Series B round totals $220 million, led by Accel, Factorial Funds, and Founders Fund, with additional participation from Conviction, Gigascale, O1A, Felicis, Buckley Ventures, and 8VC, alongside existing investors. Fractile’s platform is engineered to achieve up to 1,200 tokens per second, a rate that can shrink multi‑million‑token workloads from months to days. To support full‑stack semiconductor development, the startup has opened engineering hubs in the United Kingdom, the United States, and Taiwan. Investors cited the potential to reduce latency in AI‑driven applications such as chatbots, translation services, and code generation tools as a key reason for their participation.

With fresh capital, Fractile can move from design to volume production, putting its chips directly into the hands of enterprises that need rapid AI inference. The speed target positions the company to compete with established GPU vendors by offering a specialized alternative for workloads that are limited by memory bandwidth. Successful deployment could shorten iteration cycles for research teams and lower operating costs for large‑scale AI services. Enterprise customers looking to deploy large models at scale often face trade‑offs between cost, power consumption, and response time. Fractile’s approach attempts to shift the balance toward lower latency without a proportional increase in energy use.