Callosum raises $100m seed led by Atomico to split AI jobs across many kinds of chips

Callosum, a London company building software that sends each part of an AI workload to the chip and model best suited to it, has raised a $100m seed round led by Atomico. UK reports put the figure at about £73.5m. That makes it one of the largest seed rounds raised by a European startup, and it comes only six months after the company came out of stealth.

Plural, DCVC and the UK Sovereign AI Fund also took sizeable stakes, alongside other investors and angels. Callosum said the deal is the first investment the government-backed fund has ever made. The company did not disclose a valuation.

Ten times the pre-seed, in six months

The round dwarfs what Callosum raised in February. Back then it announced a $10.25m pre-seed led by Plural, with 22 other backers and support from the UK's Advanced Research + Invention Agency (ARIA). Plural has now followed on, so total funding comes to more than $110m.

The company was founded by Danyal Akarca, its chief executive, and Jascha Achterberg, its chief technology officer. The two met while doing PhDs at Cambridge in neuroscience and AI,, and Akarca has described the company's thinking as inspired by the brain. Callosum says its founding team has worked across Cambridge, Oxford, MIT, Imperial College London, ETH Zurich, Microsoft and Intel, on everything from chip design and kernels to clusters and new model architectures.

What the software does

Most AI systems today run one large model on racks of identical processors. Callosum is betting against that setup. Its argument is that useful AI work increasingly involves several models with different strengths, and that a growing range of chips, from in-memory and analogue designs to photonic and neuromorphic processors, each handle some tasks far better than others.

Its software takes a workload apart into its component tasks and assigns each one to the model and hardware that will run it best, within limits the customer sets on cost, energy use and speed. The company calls this programmable heterogeneity and brands itself as an intelligent systems company. The commercial product, Tailored Inference, is sold through a family of APIs. Callosum said it is already in production with customers in fields such as cybersecurity and finance, though it did not name any of them or publish figures on performance.

Cerebras and Rebellions on board

Alongside the money, Callosum named its first set of hardware partners. The flagship deal is with chipmaker Cerebras, whose systems Callosum will use to offer very low-latency inference, including for setups where several AI agents work together.

"By integrating Cerebras into Callosum's platform, we're making ultra-low-latency inference available exactly where it creates the greatest impact, enabling customers to build AI systems that simply weren't practical before," said Andy Hock, chief strategy officer at Cerebras.

Chip designer Rebellions has also signed up, along with unnamed infrastructure providers and hardware makers. For newer chip companies, the appeal is a route to customers without having to replace the incumbent hardware outright. Rebellions chief executive Sunghyun Park said the partnership puts its chips into systems next to other hardware picked for different parts of a job, "instead of asking one chip to do every job".

That pitch matters in a market where AI compute is costly, much of it tied to Nvidia hardware, and where inference bills are a growing line in the budgets of companies building on large models. Software that can move pieces of a workload onto cheaper or faster silicon gives buyers more say over price and supply, and gives alternative chipmakers a way in.

A sovereignty play for Britain

The UK government is treating Callosum as part of its industrial strategy. Besides the Sovereign AI Fund cheque, the company says it is named in the government's £1.1bn plan for AI hardware. Its case to policymakers is that if frontier-level AI can run on whatever mix of chips a country can obtain, no single supplier controls access.

Kanishka Narayan, the UK minister for AI, said demand for chips will keep growing and that winning in AI will depend not only on getting hold of hardware but "on using them as efficiently as possible".

For European investors, the round is a sign that large early cheques are flowing into the infrastructure layer as well as into model builders. Atomico leading a nine-figure seed for a company still early in commercial deployment is a bet that the money spent on AI compute will shift towards whoever makes the hardware go further.

Callosum has not said how it will split the new money. The obvious test is commercial: naming paying customers and showing, with numbers, that routing work across mixed hardware beats running it on a single fleet of GPUs. The company is also recruiting, and calls the Cerebras and Rebellions deals its first wave of compute partnerships, which suggests more chipmakers will follow.