Fractile

Inference silicon that puts compute and memory on the same die, skipping the DRAM bottleneck; claims LLMs run far faster at a fraction of the running cost. London HQ plus a Bristol hardware engineering site, with first datacentre chips due 2027.