Callosum has raised $100 million in funding to develop software that directs artificial intelligence tasks to the most cost-effective chips capable of performing the required work. This initiative aims to address the problem of high operational costs in the AI sector.
The investment round was led by Atomico, with participation from Plural and DCVC. Furthermore, it marked the first investment from the new UK Sovereign AI Fund. At a total of £73.5 million, this round is among the largest seed rounds in UK history.
The funding comes at a critical time as AI transitions from training large models to their widespread deployment. During this period, companies are finding that the majority of costs are incurred during the execution of these models.
Most companies today run every AI task on expensive general-purpose Nvidia GPUs. Regardless of whether the task requires massive computational power or just basic processing, the same costly chip is used.
This brute-force approach quickly depletes budgets. For many AI-focused companies, the inference stage consumes over 50% of revenue. Although training attracts more attention, inference determines business profitability.
Callosum offers a different approach by rejecting a one-size-fits-all attitude towards all computational processes. The company was founded by neurobiologists from Cambridge, Daniel Akarcay and Yashkha Akhterberg, who drew inspiration from biology. They created software that mimics the human brain's function, which uses small, specialized neural circuits to perform specific actions.
Callosum's software functions as an intelligent routing layer between hardware and applications. It breaks down complex AI tasks and automatically distributes each part to the most efficient and inexpensive available chip and model.
Instead of defaulting to Nvidia, Callosum collaborates with specialized chip manufacturers such as Axelera, Cerebras, and Rebellions. A simple task receives a simple chip, while a complex one receives heavy-duty hardware.
Initial results are already being demonstrated in early deployments. In partnership with chip manufacturer Cerebras for agentic financial workloads, Callosum reports a 70 percent reduction in costs. By matching tasks with specialized silicon instead of general-purpose GPUs, the company also achieves a fourfold increase in speed.
The main thesis the company presents to investors is that the next phase of AI development is not about buying more chips, but about maximizing the efficient use of existing hardware. The company notes that 'running generative AI models is incredibly expensive' and seeks to change how software interacts with physical microchips so that companies stop overpaying.
The AI market is undergoing changes: the race to train foundational models is slowing down, but the race for their cheap deployment is accelerating. This makes workload orchestration the next major battleground in hardware. With support from Atomico and the UK government, Callosum occupies a central position in this trend.
The founders assert that the future of AI computing lies not in running every task on the largest available GPU, but in choosing the right tool for the specific job, much like the brain does.