Neural networks, a model for building computing that mimics what we understand about how the brain works, has long been one of the key concepts to underpin the development of artificial intelligence. Now, a new UK startup called Callosum has raised a seed round of $100 million to apply neural concepts to how computing infrastructure is built and operates.
The goal: to make AI, and other applications that require heavy compute workloads, more efficient — faster, cheaper to run, and less energy- and chip-intensive — than they are today by creating an integration platform that lets multiple vendors work together to power compute more efficiently and make lighter work of training and applying AI models, and eventually other kinds of computing tasks that also require a lot of processing power.
The industry, said Jascha Achterberg, the CTO, is at an inflection point right now. “We have a variety of vendors and a variety of abilities that are complementary to each other,” he said in an interview with Resilience Media. “We want a market environment where complementary vendors can exist. The goal is interoperability… and our early results show that it is possible.”
Callosum is announcing the funding with a long list of secured partners.
They include chip companies Cerebras, Rebellions and Axelera AI alongside bigger players like NVIDIA, AMD and Intel; the server company Supermicro; integration specialist Normal Computing; cloud giant AWS; Furiosa and many more. Some like Normal have yet to actually release products themselves; the idea is that by working earlier with them, the integrations will be part of the development, Achterberg, who co-founded the company with CEO Danyal Akarca, said in the interview.
Notably, the list does not include any of the AI model companies: these will come over time, he said, likely starting with open-source and open-weight companies rather than those that are proprietary.
Callosum’s aim may be to reduce costs and be more efficient, but that effort will clearly require a big investment of its own. A $100 million seed round is one of the largest ever for a startup in Europe and speaks to what it might need to raise in subsequent rounds.
If the funding amount and partners don’t already spell out some of the startup’s ambitions, its investor list is also very eye-catching and underscores how heated the market is right now, not just for AI but what comes next, since the AI industry feels like an untenable bubble with high operation costs that are being carried by investor money, not profit.
The seed round is led by Atomico, one of the biggest VCs in Europe; others in the round with “significant participation” include Plural, the London-based VC that has carved out a reputation for itself in defence, sovereign and resilience technologies; the US investor DCVC. The UK Sovereign AI Fund is also a major investor, and that is also a big deal. It’s the government organisation’s first named investment after announcing a £500 million fund to invest in more independent, sovereign UK AI computing operations earlier this year.
Akarca and Achterberg met when they were neuroscience researchers working on their PhDs at Cambridge University, a path that Achterberg said he originally thought would take him to academia, not startups, but AI has demonstrably cracked open the question of how compute can possibly made the next quantum leap in power, and neural models appeared to be an obvious way to conceive of how to him, not least because of their close link with how AI itself has been built. The startup has already hired a bunch of people from across the industry and academia for its initial push and it’s hiring more.
Here is a rundown on what Callosum is doing, and why it will have caught the eye of investors and others in the industry; and why it matters in the context of “resilience tech” and digital sovereignty:
Callosum describes its core concept as “heterogeneous”.
This is not only a reference to how brains work when they fire many circuits and neurons to learn, remember, and apply learnings in real time. It is also a reference to how Callosum believes computing infrastructure should be built. AI works best and most efficiently, they argue, when multiple operators, in the form of chips and data centres and AI and other models, work in concert rather than individually.
This, as it turns out, touches on some elusive goals in the world of AI in particular around diversification and efficiency.
There are a number of companies competing across chips, foundational models, data centres, and other aspects of the ecosystem, but in that mix, there are some outsized companies that control the market. Nvidia dominates the processing space; Anthropic and OpenAI are pulling in huge numbers of users and funding and data that all help them build even more powerful models, and so on.
You can see how this would appeal to different stakeholders.
Vendors don’t want Nvidia to be the only name synonymous with processing. Users want more choice and thus lower costs in the software and hardware they buy. And investors want to make sure that their AI investments also have a fighting chance of success longer-term, and they want to find more, not less, founders innovating and building more technology.
Callosum’s concept gives potential space for multiple companies across the ecosystem to play a part and have specialised uses and users. (It calls this “tailored inference”, in opposition to what it describes аs standardised solutions.)
And if it can pull off building a genuinely useful integration layer, Callosum has a potentially important role to play as a platform provider — a holy grail in tech for companies that hope to stick around for the long haul.
Callosum already has a product page up on its site where it is inviting interested parties — potential customers who want to take compute, and vendors who want to be a part of its platform — to get in touch, although it’s not clear if it’s yet commercially launched.
There is a big opportunity here to play into not just more widely how AI and other resource-intensive compute develops, but specifically how it develops outside of the countries that are the biggest in AI today.
One big question that has hovered over countries in Europe and other regions outside of the US and China is how their own nations can develop systems that are truly independent rather than interdependent, in order to improve their own sovereignty in critical moments.
The UK may be doubling down on working with US-based technology companies, but even in times of peace, the US has shown that it is capable of dictating how those companies operate; and that’s before considering how the UK interfaces with countries that are more clearly adversarial.
It’s no surprise that the UK’s Sovereign AI fund is interested in Callosum. Backing a UK startup that’s building a platform that might offer a route for building systems that can be more flexible, with components that could be swapped in and out more easily too, could be one way to offset the lack of home-grown technology within the country using it.
“When workloads can move fluidly across many models and many chips, no single vendor holds a chokepoint over anyone’s access to intelligence,” notes Atomico. “Resilient access to intelligence is fast becoming a matter of national sovereignty, which is why governments and enterprises can’t afford to wait.”
Updated with comments from the CTO and co-founder.
