Callosum has raised $100 million in a Seed round led by Atomico, with significant participation from Plural, DCVC and the UK Sovereign AI Fund alongside other investors and angels. The company is building infrastructure intended to coordinate artificial intelligence workloads across different models, chips and compute environments rather than binding them to a single stack.
The size of the round at Seed stage is itself the headline for allocators. Capital of this scale before a Series A is now largely confined to AI infrastructure, where the cost of building and validating systems software against real silicon is high and where the competitive window is judged to be short.
Why heterogeneous compute matters
Callosum uses the term "heterogeneous intelligence" for an approach in which AI workloads are distributed across different models and computing architectures according to performance, cost, energy and latency requirements. The underlying premise is that no single model or accelerator is optimal for every task, and that the gap between the best and the merely available option is widening as chip designs proliferate.
What the company calls a "programmable heterogeneity" layer is designed to break complex workloads into component tasks and assign those tasks to different models and chips based on operational constraints. In investment terms, that positions Callosum as a coordination layer above the hardware — closer to a scheduler or compiler than to a model developer or a chip designer.
Callosum argues that this routing produces better outcomes on cost, energy and speed than running an entire workload on one architecture. Those are the company's claims rather than independently verified results, and the approach should not be assumed to hold uniformly across all AI workloads; some are simple enough that the overhead of splitting them outweighs any gain.
Callosum refers to its production offering as "Tailored Inference", delivered through APIs that coordinate inference across heterogeneous compute resources. The company says the technology is already being used in fields including cybersecurity and finance — both sectors where latency and unit cost per query are commercially material rather than incidental.
Partnerships and production deployment
Announced alongside the funding is a flagship partnership with Cerebras, the wafer-scale AI chip company. Callosum says the integration is intended to support low-latency inference within heterogeneous AI systems. Andy Hock, Chief Strategy Officer at Cerebras, commented on integrating Cerebras into Callosum's platform.
The company also announced partnerships with Rebellions and other infrastructure and OEM providers. Sunghyun Park, Chief Executive of Rebellions, framed the arrangement around combining different hardware architectures within a single platform — the practical test of whether a routing layer can hold together silicon that was never designed to interoperate.
No contract values or deployment volumes have been disclosed for any of the partnerships.
The UK sovereign AI angle
For UK-focused investors, the notable participant is the UK Sovereign AI Fund. Callosum states that it is the first investment made by the fund, and the company has also been named in the UK's £1.1 billion AI hardware plan. That figure describes the scale of the wider national programme; it is not money received by Callosum, whose disclosed funding in this announcement is the $100 million Seed round.
Kanishka Narayan, UK Minister for Artificial Intelligence, commented on the importance of using AI chips efficiently — a framing that aligns state interest less with owning models than with extracting more output from hardware that is expensive, power-hungry and, in the UK's case, largely imported.
This intersection is becoming a recurring feature of the market. Sovereign AI policy is increasingly co-investing alongside private capital rather than operating separately from it, which changes the risk profile of the companies involved: it can validate a technology and shorten procurement cycles, while also tying part of the commercial thesis to political timetables.
What investors will watch next
Three themes run through this round for family offices and institutional allocators. AI infrastructure is fragmenting rather than consolidating into one vertically uniform stack, which creates room for software that mediates between layers. Chip diversity gives that coordination software something to coordinate — its value depends on heterogeneity persisting. And inference economics, rather than model training alone, is becoming a distinct investment theme, driven by the fact that inference cost recurs with every query while training cost is incurred once.
The open questions are the usual ones for infrastructure software at this stage. Whether a routing layer can maintain integrations as chip vendors iterate; whether hardware makers and hyperscalers build comparable capability themselves; and whether measurable efficiency gains hold outside the workloads the company has chosen to showcase.
Callosum has not disclosed a valuation, revenue figures or customer numbers alongside the round.
UKFOS editorial · published 23 August 2026