Six months after leaving stealth, London infrastructure startup Callosum has closed a $100m (£73.5m) seed round to settle a question the AI boom has made expensive: which model, running on which chip, should handle any given job. The round, announced on 20 August, was led by the venture firm Atomico, with the early-stage backer Plural, the US fund DCVC and the UK government’s Sovereign AI Fund all taking part. It ranks among the largest seed rounds a British company has raised.

The size of the cheque is the headline; the thesis behind it is the story. Callosum is betting that the next phase of AI will be decided less by larger models on more expensive hardware and more by software that works out how little compute a task actually needs. That is a move from capability to cost, and it reaches marketing teams the moment a supplier calls its product “AI-powered” without saying what that costs to run.

What does Callosum actually do?

Callosum sells software that sits between AI applications and the chips that run them, breaking each workload into its component tasks and sending every one to the model and processor best matched to it on cost, speed and energy use. The company calls the approach “heterogeneous intelligence”: rather than run everything on one model and one class of chip, it treats general intelligence as a spread of capabilities and matches each task to the cheapest option that can still do the job.

The problem it targets is real and growing. Most enterprise AI runs on a narrow base of hardware — Nvidia’s CUDA software controls an estimated 85% of the GPU market — which leaves buyers exposed to one supplier’s pricing and availability. Callosum’s case is that a simple request can run on a cheaper model and processor while a hard reasoning problem is routed to something more capable, and that the savings compound as usage grows.

Founded by the Cambridge researchers Danyal Akarca and Jascha Achterberg, the company emerged from stealth in February 2026 with a $10.25m round led by Plural, taking its total raised to roughly $110m. Alongside the seed round it named hardware partnerships with the US chipmaker Cerebras, the Korean firm Rebellions and the chip designer Axelera AI, widening the mix of silicon its software can route across. The company frames the mission grandly — its blog talks of redefining how humanity computes — but the commercial proposition is narrower: lower cost and better performance per task, without the customer managing the machinery underneath.

Why is the government its first equity backer?

Callosum is the first equity investment made by the UK’s Sovereign AI Fund, a £500m vehicle the government launched in April 2026 to take direct stakes in British AI companies at what it calls venture-capital speed. The fund named Callosum as its debut equity bet at that launch, alongside supercomputing access for six other startups, and the seed round now puts a figure on the commitment. Callosum has also been named in the government’s £1.1bn plan to support homegrown chip companies.

The logic is sovereignty as much as return. In the UK, “sovereign AI” has come to mean reducing dependence on US-owned compute and model layers, not simply funding national champions — and an orchestration layer that lets workloads move across many chips is exactly the capability the government would rather not see bought by a US hyperscaler before it matures.

Kanishka Narayan, the minister for artificial intelligence, framed the deal as a bet on efficiency, arguing in a statement that access to chips will matter less than using them well. The framing suits a fund under pressure to prove it can move at the industry’s pace: France has committed €109bn (about £94bn) to AI and the US has deployed tens of billions through its CHIPS Act, sums that make a £500m fund look modest.

What does the raise change for the AI tools marketers buy?

Callosum’s own numbers, published with its investors, show why the cost argument is landing. In work with Cerebras on agentic tasks in financial services, the company says its approach ran workloads four times faster, cut compute costs by 70% and lifted task-success rates by 10% against a single frontier model on conventional GPU infrastructure. Those are the company’s and its backers’ figures, drawn from one partnership in one sector and not independently verified — but the direction is the point.

For marketing teams, the move from “best model” to “cost per task” changes how an “AI-powered” claim should be read. The tools marketers buy — content generation, lead scoring, campaign analysis — increasingly compete on the terms Callosum is selling: latency, cost per action and which models sit underneath. A supplier that cannot say which models it runs, on what infrastructure, and at what operational advantage is describing a feature, not a capability. Marketing teams are already under scrutiny over what their AI spend returns, and buyers who can ask the infrastructure question are better placed to judge the answer.

There is a procurement thread too. UK marketers pitching to regulated buyers — in finance, health or the public sector — increasingly field questions about data residency and which AI systems handle customer information. A domestic, government-backed push towards heterogeneous, sovereign compute gives those conversations firmer ground, and gives vendors a sharper line than “we use AI”.

Callosum has not disclosed a valuation, and its headline performance figures rest on a single partnership in one industry. Whether the same gains hold across the everyday AI workloads inside marketing software remains untested.