Callosum Raises $100M Seed to Route AI Workloads to Cheaper Chips

Founded by Cambridge neuroscientists Danyal Akarca and Jascha Achterberg, Callosum is building software infrastructure that directs AI workloads to the least expensive chip capable of handling a given task, rather than defaulting every job to general-purpose Nvidia GPUs. According to Ventureburn, inference, the process of running deployed models, accounts for more than 50% of revenue at many AI-native companies, making it a decisive factor in profitability.

The company positions its approach as a counterpoint to the view that greater intelligence will emerge from scaling a single large model on identical hardware, an approach it argues is energy- and capital-intensive. Callosum contends that real-world problems are varied and require different capabilities, and says its method is faster and cheaper at scale.

Callosum has announced a partnership with US chip company Cerebras. Tech.eu reported that UK Minister for AI Kanishka Narayan said that by integrating Cerebras into Callosum's platform, the companies are making ultra-low-latency inference available where it creates the most value.