Callosum raises $10.25M to challenge single-vendor dominance in AI compute

While the AI world keeps revolving around a handful of giant platforms, and the wait for NVIDIA GPUs stretches to months, London startup Callosum has quietly closed a $10.25M funding round, prompting the industry to ask whether it is time to change the architecture of AI compute itself.

The round was led by Plural, a European early-stage venture fund that focuses on infrastructure technology. Among the angel investors is Charlie Songhurst, a well-known serial investor and former Microsoft strategist. By the standards of the AI industry, where rounds often run into the hundreds of millions, the sum looks modest. But the money backs an idea that could reshape the market for AI compute infrastructure.

Callosum develops software for multi-chip AI infrastructure. Put simply, the company builds an abstraction layer that distributes compute tasks across processors from different manufacturers: NVIDIA GPUs, AMD accelerators, specialized chips from Intel, Google or any other vendor. The startup's name is no accident: the corpus callosum is the brain structure that connects the two hemispheres and coordinates them. The metaphor fits: Callosum aims to be the link between disparate compute resources.

To see why this matters, look at the current state of the market. Today, AI compute is practically synonymous with NVIDIA. Jensen Huang's company controls roughly 80-90% of the market for GPUs used to train neural networks. Its CUDA ecosystem has become the de facto standard: developers write code for CUDA, cloud providers buy NVIDIA cards, and alternative chipmakers struggle to break through this vertical integration. The result is a shortage of compute capacity, dependence on a single supplier and rapidly rising prices. For companies training large language models, that means infrastructure budgets in the billions and strategic vulnerability.

This is where Callosum sees its niche. If a software layer can make efficient use of chips from different manufacturers, companies gain freedom of choice. They can mix hardware, optimize costs and reduce dependence on a single vendor. For the market, this could mean stronger competition among chipmakers: AMD, Intel and dozens of startups building specialized AI accelerators would get a real chance to compete not only on hardware specifications but also on availability for end users.

The European context of the deal matters too. Venture capital in AI infrastructure has traditionally been concentrated in Silicon Valley. By backing Callosum, Plural is effectively saying that Europe can produce not only consumers of AI technology but also builders of the infrastructure layer. This fits a broader trend: the European Union is actively promoting digital sovereignty, and companies that reduce reliance on American and Asian chipmakers fit that strategy neatly.

Callosum's road will, of course, not be easy. The CUDA ecosystem has been built over more than fifteen years, and its network effect is enormous. Attempts to create alternative frameworks, from OpenCL to AMD's ROCm, have so far failed to seriously shake NVIDIA's position. But the situation is changing. Growing demand for AI compute makes single-vendor dependence increasingly risky, and major cloud providers, from Google with its TPUs to Amazon with its Trainium chips, are already investing in their own hardware. The market needs software glue to hold this fragmenting ecosystem together.

Ten million dollars is a starting position, not a finish line. But the bet on multi-chip architecture reflects a deeper shift in the industry's thinking. The era in which one chipmaker could dictate terms to the entire AI market may be approaching a turning point. And if Callosum can prove that heterogeneous computing performs no worse than monolithic setups, everyone will feel the consequences, from cloud giants to startups training their first model.