Callosum raises $100M for AI workload optimization
The money pouring into AI is increasingly spent after the model is trained — on the steady, unglamorous work of running inference at scale. A new funding round today shows where a slice of that budget is heading.
Callosum Ltd. has raised $100 million to build software that makes AI applications cheaper and faster to run. The London-based startup's seed round was led by Atomico, with Plural, DCVC, and the UK Sovereign AI Fund all participating — a notable lineup for a company that only closed a $10.25 million raise back in February. Callosum sells a cloud service called Tailored Inference, and the pitch is deceptively simple: most AI tasks are really several smaller steps strung together, and not every step needs a frontier model.
The product breaks a task into standalone modules, then sends each one to the model best equipped to handle it. Routine work gets routed to lightweight, inexpensive algorithms; the genuinely hard steps go to the largest models. It goes one layer deeper, too — after deciding which model should run which part, it places each model on the chip that runs it most efficiently. According to Callosum, that approach completes some inference tasks 3.7 times faster than GPT-5.6 Luna while producing better output quality, and trims infrastructure cost along the way. Those are the company's own benchmarks, so they deserve a skeptical eye until independent tests land.
The round lands alongside a partnership with Cerebras Systems, announced today. Cerebras debuted its newest wafer-scale chip, the WSE-3 Turbo, earlier this week, and Callosum will weave that hardware into Tailored Inference — letting customers run huge models on purpose-built silicon rather than general-purpose GPUs.
Why it matters: inference, not training, is where the recurring cost of AI lives, and the race to orchestrate that compute more cleverly is quietly becoming its own industry. Callosum's bet is that "how intelligently compute is orchestrated, not simply how much compute is available," as CEO Danyal Akarca put it, will define the next phase of AI. That thesis is attracting real money, but the open question is whether per-task routing actually beats a well-tuned single model once you account for the overhead of splitting and coordinating work — and whether enterprises want to fragment their stacks to chase the savings.
What to watch: independent benchmarks on the 3.7x claim, and whether the Cerebras integration ships beyond a press partnership.
Would you route your AI workloads across a dozen models to cut the bill, or keep things simple on one? Tell us in the comments.
