Callosum Raises $100 Million in Seed Round

A seed round of $100 million is a major announcement, even in the capital-intensive AI market. The British AI infrastructure startup Callosum plans to use the funding to build a platform that connects different AI models, chips, and computing infrastructures. The round is led by Atomico, with participants including Plural, DCVC, and the UK Sovereign AI Fund.

$100 million already in the seed phase

Callosum is tackling one of the biggest bottlenecks of the current AI boom: available computing power—and, above all, its efficient use.

The company argues that the next stage of AI development will not be determined solely by increasingly powerful models or individual chips. What will be decisive, it says, is which combination of model and hardware can solve a specific task most efficiently in terms of cost, speed, and energy consumption.

Callosum is now receiving $100 million to build this infrastructure. The lead investor is Atomico. Plural, DCVC, and the UK Sovereign AI Fund, as well as other investors and business angels, are also significant participants in the round. According to Callosum, this is the UK Sovereign AI Fund’s first investment.

Not One Chip for Everything

Callosum refers to its core concept as “heterogeneous intelligence.” Instead of always running AI workloads on a specific hardware architecture, individual tasks are broken down and then assigned to the exact combination of AI model and chip that can handle them best. Callosum aims, in a sense, to create an overarching orchestration layer for an increasingly fragmented market.

This is because, alongside established GPU architectures, specialized AI chips are emerging worldwide that promise significant advantages in terms of speed, power consumption, or cost for specific workloads. For companies, however, this diversity also increases complexity.

Cerebras and Rebellions Become Partners

In parallel with the funding round, the company is announcing an initial series of compute partnerships. Cerebras Systems plays a central role in this effort. Together, the companies aim to enable particularly low-latency AI inference for complex multi-agent systems. South Korean AI chip developer Rebellions is also among the new partners. Other infrastructure companies and hardware manufacturers are also on board.

Callosum does not intend to replace a specific chip or become a hardware manufacturer itself. Rather, the platform is designed to ensure that different hardware components within an AI system can be deployed where their individual strengths lie.

AI workloads are broken down into individual tasks

Technologically, Callosum refers to this as “programmable heterogeneity.” AI workloads are broken down into individual subtasks and then assigned to different models and chips based on criteria such as cost, power consumption, and speed.

This could be particularly relevant for complex agent systems. Instead of having a single large model handle all tasks, different specialized models could work together on hardware best suited to each.

Callosum aims to make this so-called “tailored inference” accessible to businesses via an API infrastructure. According to the startup, the approach is already being used in fields such as cybersecurity and finance.

$100 Million Bet on the AI Infrastructure Market

The exceptionally large seed round demonstrates how significantly the competition surrounding the infrastructure behind generative AI has shifted. For a long time, the dominant question was who would gain access to sufficient GPUs. However, with an increasing number of specialized chips and models, a new problem is emerging: How can this diversity be combined as efficiently as possible?

At the same time, the company is betting that the future of AI infrastructure will not be shaped by a single dominant hardware platform and a universal model.

The participation of the UK Sovereign AI Fund also lends the funding round an industrial policy dimension. Callosum argues that a heterogeneous infrastructure could make countries less dependent on individual hardware suppliers. High-performance AI systems should be able to run on a variety of available chips, rather than being completely tied to a single manufacturer or architecture.