Dunamu & Partners said on the 25th that it has made a $5 million (about 7 billion won) early-stage investment in Callosum, a UK-based AI inference optimization startup.
The investment was led by Atomico, a European venture capital firm, with participation from the UK government's sovereign AI fund. The startup raised $100 million (about 140 billion won) in total.
Callosum was founded in 2024 by CEO Danyal Akarca, a former brain science professor at the University of Cambridge, and CTO Jascha Achterberg. The two founders drew on the observation that the brain builds intelligence not by multiplying a single type of neuron endlessly, but by combining circuits with different functions. They concluded that AI would likewise evolve toward combining multiple models and chips that each excel at different tasks, rather than scaling up a single massive model.
AI inference is currently carried out by running one large model on a single type of chip. Callosum, by contrast, breaks a task into smaller units and distributes them across different models and chips according to cost, power and speed requirements. In an actual AI agent task in the financial sector, processing was four times faster and computing costs fell 70% compared with the single-model approach on conventional graphics processing unit (GPU) infrastructure. At the same time, the task success rate improved 10%.
Callosum works with global semiconductor and AI infrastructure companies including Rebellions, Cerebras, Axelera AI, Tendrils and Supermicro. Among them, Rebellions is a company that Dunamu & Partners invested in at an early stage, and the partnership between the two firms was arranged through Dunamu & Partners' introduction.
"When multiple models and chips operate as a single system, AI can work faster and more accurately at a reasonable cost," CEO Akarca said. "Korea is a market with a world-class AI chip ecosystem, and we will expand our opportunities in Asia, including the partnership with Rebellions that began through Dunamu & Partners' introduction."
