London-based firm Fractile, which builds AI inference hardware designed to speed up large language model execution and reduce compute costs, has raised a $220 million Series B round from Accel, Factorial Funds, Founders Fund, Conviction, Gigascale Capital, O1A Ventures, Felicis, Buckley Ventures, and 8VC.
- Founded in 2022 by Walter Goodwin, Fractile builds specialized processors for AI inference that aim to remove the latency and cost bottlenecks of running large language models at scale.
- Its hardware redesign integrates memory and compute to improve data movement efficiency, targeting much higher throughput for model execution. The company states its systems are designed to reach up to 25× faster inference performance while reducing operating costs to about 10% of current levels, addressing limitations in conventional chip architectures.
- Fractile focuses on workloads where AI models generate extremely long outputs, sometimes reaching tens of millions to around 100 million tokens, which on existing infrastructure running near ~40 tokens per second can stretch into multi-week runtimes. The goal is to push this toward ~1,200 tokens per second, significantly shortening execution time for complex reasoning and long-context tasks.
- The company positions this capability as essential for emerging compute-heavy applications, including autonomous coding, scientific research, and other domains requiring extended chains of reasoning.
Details of the deal
- The funding will help Fractile accelerate the development and deployment of its next-generation AI inference chips and systems, aimed at dramatically increasing the speed and efficiency of running advanced AI models.
