London-based startup Callosum raised $10.25 million to build software that distributes AI workloads across a mix of chips, directly challenging Nvidia's GPU dominance. Founded by Cambridge neuroscientists, the company also secured U.K. government research funding for sovereign cloud infrastructure. Its platform manages tasks across Nvidia, AMD, AWS custom silicon, and other processors.

The founders' approach stems from neuroscience: the brain combines specialized neurons, not copies of one type. This coincides with an industry pivot toward inference workloads, which Deloitte forecasts will account for two-thirds of AI compute by 2026, creating a $50 billion market for optimized chips. Inference workloads favor diverse hardware over uniform GPU farms.

Callosum asserts its system can achieve twice the accuracy, seven times faster performance, and four times lower cost on complex enterprise tasks by matching subtasks to specific hardware. It targets firms building multi-agent AI systems and新兴 chipmakers needing scale validation. The software integrates with existing multi-cloud setups without re-architecture.

Backed by Plural's Ian Hogarth and others, Callosum will use funds to expand in the U.S. and develop proprietary hardware. Its model exploits chip diversity as an advantage, not a hurdle, positioning it to reshape data center design and dismantle the single-vendor dependency that has defined AI infrastructure.