# Callosum Banks $100M Seed to Route AI Workloads Around Nvidia Lock-In By Ray Termsheet · 2026-08-25 · Deals & Funding · https://datacommenter.com/callosum-banks-100m-seed-to-route-ai-workloads-around-nvidia-lock-in/ About the author: Follows the money in the data economy: venture rounds, M&A, and strategic stakes across data vendors and infrastructure. > London-based Callosum raised $100 million in seed funding led by Atomico, with Plural, DCVC and the UK Sovereign AI Fund joining, just six months after a $10.25 million pre-seed —… Original reporting: [Tech.eu](https://tech.eu/2026/08/20/callosum-raises-100m-seed-round/) _AI-assisted commentary, editorially reviewed. Quoted excerpts belong to the original outlet._ Callosum’s $100 million seed round, led by Atomico with Plural, DCVC and the UK Sovereign AI Fund joining, is one of the largest seed checks Europe has ever written for a company that only shipped its first product today. The London startup, founded by Cambridge neuroscientists Danyal Akarca and Jascha Achterberg, is betting that the real bottleneck in AI isn’t model quality but how inefficiently compute gets allocated once a model is actually asked to do something. ### Before and after Before Tailored Inference, an enterprise AI workflow typically ran end-to-end on one frontier model sitting on one vendor’s chips — brute-force and, per Electronics Weekly, expensive enough that AI-native companies routinely burn half or more of revenue on inference alone. After it, the same workload gets chopped into discrete “blocks,” with simple steps shunted to cheap models and hard steps kept on frontier ones, each then matched to whichever silicon — Cerebras, Rebellions, Axelera, d-Matrix, AMD, AWS — runs that specific piece most efficiently. The timing with Cerebras is no accident: the chipmaker unveiled its CS-4 rack system on August 18, 2026, claiming up to 30x GPU inference speeds, and rolled straight into a Callosum integration announced alongside the funding. Callosum claims its Cerebras-powered financial-services deployments ran four times faster, cut compute costs 70%, and lifted task-success rates 10% versus a single frontier model on conventional infrastructure, according to Electronics Weekly — figures that are entirely self-reported and worth treating with the usual skepticism until an independent benchmark shows up. > A $100 million seed round for a routing layer, not a model, is a wager that the next chokepoint in AI isn’t intelligence itself but who controls how it gets deployed. The bigger tension: Callosum’s whole pitch is chip-agnosticism, yet its loudest launch partner is a single chipmaker getting prime billing on demo day. Watch whether Supermicro, HPE and the smaller silicon partners (Axelera, d-Matrix, Lumai, Tendrils) get equal airtime once customers actually deploy, and whether hyperscalers with their own inference-optimization ambitions decide to build this layer themselves rather than license it. > Callosum is challenging the assumption that AI development and greater intelligence will come through scaling a single AI model on identical chips, a process which demands high energy and capital costs and concentrates power in the hands of the likes of Nvidia, OpenAI and Anthropic. > — [Tech.eu](https://tech.eu/2026/08/20/callosum-raises-100m-seed-round/) [Read the full story at Tech.eu →](https://tech.eu/2026/08/20/callosum-raises-100m-seed-round/) --- Cite this analysis: https://datacommenter.com/callosum-banks-100m-seed-to-route-ai-workloads-around-nvidia-lock-in/ Cite primary facts: https://tech.eu/2026/08/20/callosum-raises-100m-seed-round/ Need the underlying datasets (alt data, market data, AI training data)? Source licensed vendors via Brickroad: https://brickroad.network More machine-readable access: https://datacommenter.com/llms.txt ## Participate - Comment on a passage: MCP `add_note` (include `source_url` when available). - Suggest an editorially reviewed correction: MCP `suggest_edit`. - Open factual questions: none.