Callosum

Market position

Callosum operates as a software orchestration layer in a hardware market shifting from training toward inference. Fortune cited Deloitte estimates that inference workloads would account for roughly two-thirds of AI compute in 2026, up from a third in 2023, with the inference-optimized chip market exceeding $50 billion that year. Coverage positions the company as challenging dependence on Nvidia GPUs, and contrasts it with chip builders such as Fractile and Etched, which are raising capital to build compute rather than to allocate it.

Callosum positions itself against the prevailing approach of running ever-larger models on banks of identical Nvidia GPUs, instead matching subtasks to specific models and silicon. The company states its system can deliver twice the accuracy, sevenfold faster performance, and fourfold lower cost versus running the same complex workloads on identical hardware, and that its platform does not require customers to re-architect existing cloud setups. It also cites a founding team spanning both intelligence and compute, with backgrounds across Cambridge, Oxford, MIT, Imperial College London, Microsoft, ETH Zurich, and Intel.

Technology

Software that decomposes complex AI workloads into discrete components and dynamically routes each to the model and chip configuration best suited to it, across multi-cloud environments (AWS, Google Cloud, Microsoft Azure) and heterogeneous silicon (Nvidia GPUs, AMD, AWS Trainium/Inferentia, Cerebras, SambaNova, Rebellions, Axelera AI). The company describes co-evolving chips and algorithms, publishing research alongside production engineering, and collaborating on data center interconnect technologies including photonics-based networking. Its Tailored Inference API family provides automated task decomposition, persistent multi-agent coordination, and sovereign deployment.

Go-to-market

The company targets two customer groups directly: enterprises and teams building multi-agent AI systems, and chip manufacturers seeking to demonstrate their hardware at scale. It pursues partnerships with silicon vendors — including an announced partnership with Cerebras Systems and deals involving Rebellions and Axelera AI — to broaden the hardware its software can address, and publishes research and technical results publicly.

Companies building multi-agent AI systems requiring performance across complex, multi-step workflows (cited examples include automating computer use and payments processing), and emerging chip manufacturers seeking to show their hardware's capabilities at scale. The Tailored Inference launch also references sovereign deployment for enterprise buyers.

Geography

Headquartered in London, United Kingdom. The founders stated plans to expand the London team and begin scaling into the U.S.; the company also said it would expand its engineering team across international hubs.

History

The cofounders, Danyal Akarca and Jascha Achterberg, met during PhD studies at Cambridge around 2019. Fortune reported in February 2026 that the company disclosed a $10.25 million round led by Plural together with ARIA research grant funding, and published an introductory blog post the same day describing 22 investors. In August 2026 the company announced a $100 million seed round led by Atomico and launched Tailored Inference. Reporting on the founding year is inconsistent: TechStartups states the company was founded in 2025.

Risks & controversies

Sources report conflicting founding dates (internal record of 2024 versus TechStartups' statement of 2025). Performance claims (2x accuracy, 7x speed, 4x lower cost) are company-stated and not independently verified in the sources. The company's approach depends on continued fragmentation of AI silicon and on dissolving software lock-in to single vendors.