Fractile is building chips to run large language models two orders of magnitude faster. Existing hardware is good for training LLMs, but very poorly suited to subsequent inference of the trained model, which is increasingly the dominant workload. A network’s weights need to be moved onto a chip once per word generated, and this movement takes a few hundred times longer than the subsequent computations themselves. Fractile’s revolutionary approach to fusing computation with memory eliminates this bottleneck, and can scale to allow running the world’s largest models at a global scale.