Callosum raises $100M seed round to challenge Nvidia's AI dominance

London-based Callosum has raised $100 million to route AI workloads across competing chip architectures, directly challenging Nvidia's dominance of enterprise inference.

Callosum, a London-based AI infrastructure startup founded by Cambridge neuroscientists, has raised $100 million in seed funding to route AI workloads across competing chip architectures, directly challenging Nvidia's dominance of enterprise inference.

"AI is nothing without the chips that underpin it — and the eye-watering demand for them is only going to grow," Kanishka Narayan, UK minister for artificial intelligence, 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."

The round was led by Atomico with participation from Plural, DCVC and the UK Sovereign AI Fund — the fund's first disclosed equity investment since its £500 million launch in April. Callosum has now raised approximately $110 million total, following a $10.25 million pre-seed in February. The company's platform routes each subtask of a complex AI workflow to the model and chip combination best suited for it, spanning Nvidia GPUs, AMD processors, Amazon Trainium and Inferentia silicon, and specialist chips from Cerebras, Rebellions, Axelera, d-Matrix and Lumai.

The bet is that inference — which Deloitte estimates will account for roughly two-thirds of all AI compute in 2026, up from a third in 2023 — is where the next battleground lies. Callosum's published benchmarks on financial-services agentic workloads show 4x faster execution, 70 percent lower compute costs and 10 percent better task success rates versus running the same workloads on a single frontier model on conventional infrastructure.

The Brain as Blueprint

Callosum was founded in 2025 by Danyal Akarca and Jascha Achterberg, computational neuroscience PhDs who met at Cambridge around 2019. Their thesis: the human brain achieves intelligence not by replicating one neuron type billions of times, but by combining many specialized cell types that work in concert. The company name references the corpus callosum, the neural bundle connecting the brain's two hemispheres — the integration layer between specialized systems.

"Big labs are currently betting that one model will rule them all. We think that's wrong, and our work proves this," Akarca said at the company's February launch. "Nature shows that real intelligence emerges from many systems working together."

The platform operates through a two-path architecture. A planning layer decomposes each incoming workload into subtasks and profiles each one's compute requirements — latency sensitivity, accuracy threshold, budget ceiling. A fast-path routing layer then dispatches each subtask in real time to whichever model and chip combination best fits those constraints. The system runs across AWS, Google Cloud and Microsoft Azure without requiring customers to restructure their cloud setups.

Why Sovereign AI Backing Matters

The UK Sovereign AI Fund's decision to take its first equity stake in Callosum reflects a structural problem: countries building AI infrastructure to reduce dependence on US hyperscalers can procure alternative chips — Korea's Rebellions ATOM NPU, the UK's Graphcore and Olix — but those chips are commercially useless if the enterprise software stack assumes Nvidia CUDA compatibility. Callosum's orchestration layer is the software that makes non-Nvidia silicon accessible without rewriting applications.

The fund, chaired by James Wise of Balderton Capital with Suzanne Ashman as managing partner, offers portfolio companies up to one million GPU hours on the UK's AI Research Resource supercomputer and equity investments typically ranging from £1 million to £10 million.

The competitive field is heating up. Cerebras on Aug. 19 doubled its CS-4 inference speed to 4,400 tokens per second without new silicon. Etched raised $700 million at a $21 billion valuation for its transformer inference clusters. UK chip startup Olix raised $312 million at a $3.3 billion valuation for photonic inference chips in August, while Fractile is reportedly in talks to raise approximately $600 million.

"By integrating Cerebras into Callosum's platform, we're making ultra-low-latency inference available exactly where it creates the greatest impact," Andrew Feldman, CEO of Cerebras, said.

Callosum plans to use the funding to scale its engineering team in London, accelerate US expansion and build out complementary hardware infrastructure. The company has also described a longer-term ambition to work with photonics-based interconnect companies that transmit data between chips using light rather than electrical pulses, addressing the communication bottleneck that grows more severe as heterogeneous chips need to talk across a rack.

The routing approach carries risks. Routing quality depends on the accuracy of the router's predictions about each subtask's requirements — academic research on model-to-model routing has documented a consistent tradeoff where higher accuracy tends to mean routing more queries to larger, more expensive hardware. Callosum's published benchmarks are the company's own figures; no independent third-party verification has been published.

For investors, the question is whether the orchestration layer becomes the standard interface for heterogeneous AI compute — and whether Nvidia's CUDA lock-in can be broken by software that makes alternatives viable. As inference costs become the dominant expense for AI-native companies, the ability to dispatch workloads to cheaper, specialized silicon could reshape procurement decisions across the industry. Ian Hogarth, Plural co-founder and former chair of the UK AI Safety Institute, described Callosum's "vision for a multi-model, multi-chip future" as potentially changing the competitive dynamics with the world's largest chip and model makers.

This article is for informational purposes only and does not constitute investment advice.