$100M · Callosum + 2 more

What they're building

Callosum builds the orchestration layer for heterogeneous compute. The bet is that an AI workload is not one thing: it is a workflow problem (how you define the job), an agent problem (which models you call), and a hardware problem (which silicon each of those calls should land on). Today almost everyone answers all three with the same answer, a frontier model on a rack of identical GPUs. Callosum's software decomposes the request and routes each piece to the model and chip actually suited to it. The technical surface is wide because the problem is vertical. The company works from the AI systems layer down to the compile level, and the open roles map the stack exactly: inference engine, accelerator systems software, cloud resource orchestration, networking and interconnect, evolutionary optimisation, agent runtime, benchmarking and evals. Partnerships with Cerebras, Rebellions, Axelera, Tendrils and Supermicro are the other half of the work, because routing to a chip means having tested real workflows on that chip first. The intellectual root is the founders' actual research. Danyal Akarca and Jascha Achterberg spent their Cambridge PhDs on how brains achieve flexible cognition by combining many specialised circuits rather than scaling one uniform one, publishing in Nature journals with Google DeepMind and Intel Labs. Callosum is that finding turned into infrastructure. Akarca's framing in the round announcement: "The industry has been betting that one model or one chip will rule them all. But nature shows us the opposite. Intelligence is collective and will emerge from many specialised systems working together." Achterberg puts the sovereignty argument more bluntly: "AI sovereignty isn't about owning a chip fab. It's about never being hostage to one." That sentence is why the UK's Sovereign AI Fund made Callosum its first disclosed investment since launching in April 2026.

Why this matters

The numbers Callosum publishes are the argument. On agentic financial-services workloads, its production platform reports 4x faster performance, a 70% cost reduction, and a 10 percentage-point improvement in task success rate against a single frontier model on GPU. Plural's investment note cites 2x accuracy, 7x speed and 4x lower cost on autonomous computer-use tasks. Those are the company's own figures, not third-party benchmarks, but they are specific and falsifiable, which is more than most infrastructure pitches manage. The structural case is that inference is disaggregating. A record wave of semiconductor startups is shipping real silicon (Cerebras, Rebellions, Axelera among Callosum's partners), and none of it inherits CUDA's software gravity. Someone has to be the layer that makes a mixed fleet usable, the way VMware made a mixed server fleet usable in the 2000s. If that layer is a neutral third party rather than a chip vendor, the economics of running AI change for everyone downstream. There is also a political tailwind that is unusual for a seed company. Being the UK Sovereign AI Fund's first disclosed cheque, and a founding member of ARIA's Scaling Inference Lab, means Callosum is not just a startup with a thesis; it is the thing the UK is pointing at when it talks about not being dependent on one vendor's stack.