Participated · Seed · Feb 2026

Callosum’s platform sits between applications and hardware, decomposing AI inference workloads, matching sub-tasks to specialised models and processors, and routing execution to the most suitable hardware to balance speed, cost and precision. The company claims up to 4x speed improvements, 70% lower compute costs and a 10% higher task success rate on complex agentic financial workloads compared with running a single frontier model on conventional GPUs. A family of tailored inference APIs is being launched, with some APIs already in production for complex workloads in financial services. Callosum works with a range of silicon and infrastructure partners — including Cerebras, Rebellions, Axelera, Tendrils and Supermicro — and is a founding member of ARIA’s Scaling Inference Lab. The startup was founded by Danyal Akarca and Jascha Achterberg, who met while doing PhDs at Cambridge, and plans to use the new funding to scale the technology, deepen ecosystem partnerships and grow the team.

Led · Grant · Oct 2024

Founded in 2022, Fractile develops chips and systems aimed at dramatically speeding up inference for frontier AI models. The company’s thesis is that time-to-output and inference unit economics—particularly memory bandwidth limits—are the next major bottlenecks as models generate increasingly long outputs. Fractile is working across AI research, chip microarchitecture, and foundry process innovation to reduce inference latency and cost. Its technology targets workloads that can require tens of millions of tokens and seeks to enable new, longer-running AI tasks by making faster inference economically viable. The company is expanding its team and hiring across London, Bristol, San Francisco, and Taipei as it scales development. Financial metrics, customers, and prior funding amounts were not disclosed in the articles.