Callosum, a London-based AI infrastructure startup, raised $100 million in seed funding led by Atomico, Plural, DCVC and the UK Sovereign AI Fund.
The company is building a software layer that sits between AI applications and diverse compute hardware, decomposing each workload into component tasks and routing them to the optimal processor for the job. The insight is straightforward but powerful: most AI deployments waste compute dollars by pushing all tasks through the same general-purpose hardware, regardless of whether a specialized chip could do the job faster or chea. The problem is structural. As the AI economy matures, spending is shifting from training to inference—the marginal cost of running models at scale. AI-native companies allocate half or more of revenue to inference. Meanwhile, the chip market is fragmenting rapidly. Hyperscalers are building custom silicon. Specialized processors—wafer-scale, optical, neuromorphic—are entering production. Callosum treats this fragmentation as an opportunity, not a constraint.
In early deployments on agentic workloads in financial services, tested through a partnership with chipmaker Cerebras, the platform achieved four times faster execution, cut compute costs by 70 percent and improved task success rates by 10 percent compared with routing everything through a single frontier model on conventional infrastructure.
The unit economics shift materially. Rather than optimizing for cheaper tokens, Callosum's model optimizes for useful intelligence per unit of compute. Each new chip entering the market expands the platform's decision tree rather than adding complexity—a compounding moat in an ecosystem where no single architecture will dominate.
Callosum is expanding its silicon partner network beyond Cerebras to include Rebellions, Axelera, d-Matrix, Lumai and Tendrils. Infrastructure partners include Supermicro and HPE. The company launched its first software layer for heterogeneous inference as a family of APIs that let developers integrate frontier hardware without rewriting existing applications.
Founders Danyal Akarca and Jascha Achterberg developed the core technology during PhD research at Cambridge, studying how biological brains achieve intelligence by integrating specialized circuits rather than scaling a single one. They have published more than 70 papers and assembled a team in London. Early backers also included the UK's Advanced Research and Invention Agency (ARIA).
