Overview
- Callosum announced on Thursday, August 20, 2026, that it closed a $100 million seed round led by Atomico, bringing its disclosed funding to roughly $110 million with participation from Plural, DCVC and the UK’s Sovereign AI Fund.
- The startup builds orchestration software that breaks inference jobs into smaller subtasks, analyzes each subtask’s compute needs, and routes them to the chip or model best suited for that work to try to improve speed, accuracy and cost.
- Callosum has formal integrations with Cerebras and a partnership with Korean chipmaker Rebellions to put wafer-scale and alternative accelerators behind its routing layer and to offer low-latency inference through Callosum APIs.
- The company claims up to 2x accuracy, 7x speed and 4x cost savings versus uniform GPU setups, but those performance figures have not been independently benchmarked and Nvidia’s CUDA-led GPU ecosystem remains the market incumbent with about an 85% share.
- Callosum also received a £2.9 million ARIA research grant and is pitching this stack as a way to change inference economics by easing lock-in to single vendors, a shift that could speed wider use of specialised accelerators in production AI systems.