London Startup Quietly Raises $10.25 Million to Challenge Nvidia's AI Dominance
In a move that could shake up the AI hardware landscape, a London-based venture has secured $10.25 million in a seed round. The company, Callosum, says its software can choreograph AI tasks across a mixed fleet of chips, from Nvidia GPUs to AMD processors and custom silicon from cloud providers. The round was led by Plural, a European early-stage fund backed by notable tech investors, with angel participation from industry figures and academics. The funding arrives as UK policy makers press to build a sovereign cloud that reduces dependence on US technology in AI infrastructure.
The founders, two Cambridge-trained neuroscientists, say their approach mirrors how the human brain integrates diverse cellular components to achieve intelligence. They argue that AI workloads do not need to run on identical hardware to reach peak performance; instead, orchestrating tasks across heterogeneous hardware can unlock efficiency gains and cost savings. This thesis underpins Callosum’s product roadmap as the AI market increasingly diversifies its hardware choices.
In parallel with private funding, the UK government’s Advanced Research and Invention Agency (ARIA) is providing grant support to accelerate R&D on integrating cutting-edge chip designs into Callosum’s platform. ARIA’s support comes as Britain positions itself as a leader in sovereign cloud projects designed to be independent of, or less reliant on, US suppliers. A spokesperson for ARIA confirmed the agency is backing the R&D effort, though it is not an investor in the current round.
The round’s lead investor, Plural, is co-founded by Wise co-founder Taavet Hinrikus and Ian Hogarth, who also chaired the UK AI Safety Institute. Other participants include Charlie Songhurst and Stan Boland, among a slate of angel investors who have backed startups at the intersection of AI, hardware, and cloud infrastructure. The funding signals growing interest in startups that can challenge the prevailing model of AI data centers built around large, uniform GPU clusters.
For Callosum, the core proposition is software that distributes AI tasks across multiple chip technologies in real time. That means a machine learning job could leverage Nvidia accelerators for certain layers, while other portions of the same workload run on AMD CPUs or specialized AI accelerators from newer players in the market. The idea is to extract the best performance and energy efficiency from each chip type, reducing the need to endlessly scale up a single technology to keep pace with model size and complexity.
The founders—Danyal Akarca and Jascha Achterberg—met during PhD studies at Cambridge around 2019. They say their research into the brain’s use of diverse neuron types and circuits informs Callosum’s software design. In their view, intelligence is a product of diverse, complementary components working together, not a single type of computation replicated billions of times. Callosum’s platform seeks to translate that insight into practical gains for AI workloads in the data center.
Industry observers say the timing is right for a hardware-agnostic orchestration approach. As models grow larger and training costs escalate, data center operators are increasingly exploring heterogeneous hardware strategies to optimize throughput and energy use. This dynamic has been amplified by the emergence of specialized chips from cloud providers and startups that aim to capture niches Nvidia has not fully dominated.
While the market has seen several startups targeting AI compilers and schedulers, Callosum’s emphasis on cross-chip orchestration sets it apart. If successful, its platform could appeal to cloud service providers, research labs, and enterprise customers seeking to hedge against supplier risk while squeezing more performance out of existing hardware investments.
The company’s leadership stresses that the money will go toward product development, customer trials, and expanding the technical team to support a broader range of hardware targets. They also caution that the field remains competitive and that broad adoption will depend on measurable improvements in efficiency, reliability, and ease of integration with existing AI pipelines.
The broader policy backdrop adds another layer to Callosum’s story. The UK’s sovereign cloud program is aimed at bolstering national resilience in AI infrastructure, reducing exposure to foreign supply chains, and enabling greater control over data governance and security. ARIA’s support signals government interest in scalable, hardware-diverse AI platforms that can operate across borders and public and private sectors alike.
Asked about the strategic implications, Akarca framed the funding as both a commercial milestone and a signal that a diversified hardware approach is becoming a practical alternative for AI workloads. “Our software is designed to treat each chip type as a resource with unique strengths,” he said. “The future of AI compute is not a single hammer; it’s a toolkit.” Achterberg added that the team is focused on proving real-world performance gains in collaboration with early customers and partners.
The funding round adds to a growing soundtrack of AI infrastructure bets across Europe and the UK. Investors see a potential path to reduced dependence on any single vendor while enabling faster experimentation with new chip designs, a critical factor as chip shortages and supply chain volatility continue to shape the AI landscape.
As the market weighs new entrants and established platforms, one question remains: can a software layer that deftly coordinates multiple chip types deliver consistent, scalable benefits at scale? If Callosum can demonstrate clear improvements in throughput-per-dollar and time-to-innovation for AI workloads, the exclusive: startup aiming break Nvidia’s hold on AI data centers could usher in a new era of hardware-aware AI computing.
In the coming months, Callosum plans to broaden pilots with large cloud users and research labs, while refining integration with a growing range of accelerator architectures. The outcome will hinge on real-world metrics—throughput, energy efficiency, and ease of deployment across hybrid data centers—and on whether the sovereign-cloud push gains momentum among enterprise buyers seeking to diversify risk and strengthen data sovereignty.
What This Means for AI Compute Buying Decisions
For CIOs and procurement teams, the Callosum approach could tilt budget decisions toward multi-hardware strategies rather than a one-vendor path. If the platform delivers measurable savings, buyers may reframe how they plan capacity for training and inference. The key will be interoperability, reliable scheduling, and predictable performance across different chip ecosystems.
Industry insiders are watching closely how regulatory and policy developments in the UK and Europe could affect vendor ecosystems and data governance. The sovereign cloud concept is still evolving, and how it intersects with global cloud markets will influence whether hardware-agnostic software gains traction beyond niche pilots.
Key Data Points
- Funding: $10.25 million seed round for Callosum
- Lead investor: Plural (European early-stage fund)
- Other participants: angel investors including Charlie Songhurst and Stan Boland
- Government support: ARIA grant funding to accelerate R&D on integrating novel chip technologies
- Focus: software that orchestrates AI workloads across Nvidia GPUs, AMD processors, AWS Trainium/Infernence, Cerebras, SambaNova, and future chips
- Strategic aim: reduce reliance on a single vendor and support a sovereign cloud strategy
Final Take
Callosum’s $10.25 million raise places it squarely at the intersection of AI hardware diversification and national strategy on cloud infrastructure. The company’s success will hinge on delivering tangible performance uplifts across heterogeneous chip stacks and proving that a software-first approach can unlock a new level of flexibility for AI workloads. With ARIA backing and a notable investor roster, exclusive: startup aiming break Nvidia’s grip on AI data centers now faces a pivotal test: translate theoretical gains into real-world wins across multiple chip ecosystems.
Discussion