Callosum announced a pre-seed round of $10.25m in February. In August it announced a seed round of $100m. The two events are six months and roughly one order of magnitude apart, and the second one is still called seed.
The round was led by Atomico, with significant participation from Plural, DCVC and the UK Sovereign AI Fund, plus other investors and angels. Reported in sterling it came to about GBP 75m; in euros, EUR 85.4m. Dealroom puts the company’s total raised at $110.25m and describes the round as one of Europe’s largest at seed stage. No valuation has been disclosed.
What Seed Now Means
Stage labels in venture capital describe the company’s maturity, not the size of the cheque, and the two used to move together closely enough that nobody had to think about it. A seed round meant a few million pounds, a product that mostly worked and a first handful of customers. That correspondence has been breaking for several years in AI infrastructure, and $100m at seed is what the break looks like at full extension.
The label is doing real work here rather than being sloppy. Calling it a seed round says the company has not yet built the thing at the scale it intends to build it. What has changed is that the amount of capital required to find out whether the idea works is now enormous, because the experiment involves buying time on very expensive silicon. Investors have adjusted the number and kept the name.
The practical consequence for anyone reading UK funding data is that stage is no longer a usable proxy for size. A £2m Series A and a $100m seed can appear in the same week, and a chart that groups by stage will tell you very little about either.
The State Wrote Its First Cheque
Callosum says it is the first investment made by the UK Sovereign AI Fund, the GBP 500m vehicle the government launched in April 2026 to back British AI companies in areas judged strategically important for growth and national security. The company has also been named in the government’s GBP 1.1bn AI hardware plan.
Kanishka Narayan, Minister for Artificial Intelligence, framed the rationale in terms of efficiency rather than capability. “AI is nothing without the chips that underpin it, and the eye-watering demand for them is only going to grow,” he said. “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.”
That is a coherent position for a country in Britain’s situation. The UK is not going to out-spend the United States or China on compute, and it does not manufacture the chips. Backing the layer that decides which chip runs what is a bet on being clever about a resource you cannot control, which is the kind of bet a mid-sized economy can actually win.
State money arriving inside a large private round is becoming a pattern rather than an exception. It was also the shape of the battery-materials deal in which Nexeon raised GBP 100m with half of it coming from the state. The difference is proportion. In Callosum’s case the public money sits alongside Atomico, DCVC and Plural rather than carrying the round, which is the healthier version of the arrangement: the state is following commercial conviction rather than substituting for it.
What the Money Is Actually Buying
The company describes what it does as “heterogeneous intelligence”, which needs translating. The argument is that new AI models and the specialised chips built to run them are arriving at the same time, in growing variety, and that most real workloads are not best served by one model on one type of processor.
So the software breaks a workload into smaller compute tasks and routes each one to whichever model and chip combination runs it most efficiently, with efficiency defined against constraints the customer sets: cost, energy or speed. Callosum calls this “programmable heterogeneity”, and says its production APIs, branded Tailored Inference, are already delivering performance and cost improvements for customers in cybersecurity and finance.
Three things landed with the funding announcement. A set of global compute partnerships, headlined by what the company calls a flagship partnership with Cerebras Systems, plus arrangements with the chipmaker Rebellions and other infrastructure providers and manufacturers. The routing advances themselves. And the production rollout of the APIs.
The partnerships are the part that explains the size of the round. A routing layer is only as good as the range of silicon it can route to, and access to that silicon is bought, not asked for. Andrew Feldman, chief executive of Cerebras Systems, said that integrating Cerebras into the platform makes “ultra-low-latency inference available exactly where it creates the greatest impact, enabling customers to build AI systems that simply weren’t practical before”.
The Founders Are Neuroscientists
Callosum was founded in 2025 by Jascha Achterberg and Danyal Akarca, and emerged from stealth in February 2026. The founding team’s backgrounds span neuroscience and hardware-software co-design, with staff drawn from Cambridge, Oxford, MIT, Imperial College London, Microsoft, ETH Zurich and Intel, working across the UK and the US.
The neuroscience lineage is not decoration. The company’s founding argument is that intelligence is not one uniform process running on one uniform substrate, and that the industry’s habit of putting bigger models on the same chips is running into the limits of that assumption. A brain does not solve every problem with the same tissue. The commercial claim follows directly: nor should a data centre.
Whether that translates into a durable business is a separate question from whether it is a good idea, and the honest answer is that nobody knows yet. Orchestration layers sit between customers and hardware vendors, and both of those parties have reasons to build the layer themselves. The value has to be large enough, and the chip landscape fragmented enough, that neither side finds it worth replicating.
Why Investors Are Paying Now
Niklas Zennström, chief executive of Atomico, put the timing at the centre of the case. “As AI moves from training to inference, the question is no longer just who builds the best models,” he said. “It’s how we deliver intelligence efficiently, across an increasingly diverse landscape of models and chips.”
That distinction between training and inference is the whole thesis. Training is a capital project: you buy an enormous amount of compute, run it hard for a period and produce a model. Inference is an operating cost that never stops, incurred every time anybody uses the thing. Costs that recur forever attract optimisation in a way that one-off costs do not, and a few percentage points off a permanent bill compounds into a large number.
It also sits alongside a structural feature of the British economy that shows up elsewhere in the data. Britain’s exposure to AI is unusually high because of what it sells, being weighted towards services rather than manufactured goods. A country whose output is disproportionately professional and financial services is a country whose costs are disproportionately inference costs.
What Would Make This Look Cheap
The bear case is straightforward. Chip diversity could narrow rather than widen, leaving less to route between. The hyperscalers could fold the same capability into their own platforms and give it away. Customers could find that switching between chips is harder in practice than in a benchmark.
The bull case is that inference spending grows for years, stays fragmented across vendors who have every incentive to remain incompatible, and that a neutral layer becomes infrastructure the way content delivery networks did. In that world $100m at seed is not expensive, it is early.
What is measurable now is narrow: two founders, one year, $110.25m committed, a first state investment, and a set of named chip partners. The rest is a position on how the next few years of computing are organised. Atomico and the British government have both taken the same side of it, which is worth noting precisely because they are answering different questions.
About The Author
More Stories
LearnUpon Secured $50m Without Selling a Single Share
La Cordée Is Closing With $13 Million Owed and 210 Jobs Lost
Pensioners Would Carry Half the Private Health Insurance Rebate Cut