- Having raised $10.25 million in a pre-seed round in February, Callosum has now obtained $100 million in seed funding.
- Atomico led the seed round, with Plural, DCVC, and the UK’s Sovereign AI Fund also joining.
- According to the fund’s website, the investment in Callosum was its first, made in April.
Callosum, a London-based startup, has raised a $100 million seed round, one of the largest ever in Europe, led by Atomico. This follows its pre-seed round at a $10.25 million valuation led by Plural in February.
UK’s £500 million Sovereign AI Fund, along with some other unnamed investors and angels, took part. Since its foundation in 2025, Callosum has now raised approximately $110 million in funding. The company has not shared its valuation.
From neuroscience PhDs to a nine-figure seed
Callosum was founded by Danyal Akarca and Jascha Achterberg, who first began working together while pursuing their PhDs at Cambridge in neuroscience, computing, and AI. Their research has appeared in Nature journals, and both have held positions at Intel and Google DeepMind.
Callosum’s method is based on the founders’ academic backgrounds. They hold the view that, just as intelligence in nature arises from a variety of neurons, AI should not rely on identical chips.
Instead, their software breaks down AI tasks into separate steps and sends each one to the most suitable model and chip rather than using the same hardware for all purposes. As a growing share of industry spending shifts from training to inference, AI companies frequently spend half or more of their revenue on inference.
For instance, one step in an autonomous computing task may require fast pattern matching, while another may call for deeper reasoning. The company’s system therefore allocates each step to different hardware, thus preventing competition for resources. It claims that its solution is twice as accurate, seven times faster, and four times cheaper for complex tasks compared with using uniform hardware, even though it has not made the technical benchmark figures available.
Nvidia’s CUDA ecosystem, with its ~85% share of the GPU market, remains Callosum’s principal competitor. Cerebras, the wafer-scale chipmaker that went public in May 2026 at a valuation of nearly $56 billion, has now entered into a partnership with Callosum. Although Etched and SambaNova are also developing their own custom chips to improve inference efficiency, their methods are less extensive than Callosum’s chip-agnostic approach.
A 10x increase in funding since February
The company’s new funding will enable it to become a “global heterogeneous integrator” and establish new computing partnerships. Its key partnership with Cerebras focuses on achieving low-latency inference at scale, and it is also entering into a collaboration with the Korean chipmaker Rebellions.
“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 Andrew Feldman, Cerebras’ chief executive.
Sunghyun Park, chief executive of Rebellions, framed the deal as a structural stand against single-vendor lock-in: “Working with Callosum puts our architecture into systems alongside hardware chosen for different parts of the workload, instead of asking one chip to do every job. That’s the difference between a partnership and a deployment that only works in one environment or geography.”
Kanishka Narayan, the UK’s minister for artificial intelligence, tied the investment to national chip strategy: “AI is nothing without the chips that underpin it, and the eye-watering demand for them is only going to grow. In the race to develop and use AI, success will depend not just on having access to those chips, but on using them as efficiently as possible.”
Nvidia believes the total AI infrastructure market could reach at least $1 trillion by 2027, driven primarily by inference workloads. Callosum maintains that no single chip can meet this demand. The UK government has again supported this view since April, even if it has not stated so outright.
