Callosum is the Intelligent Systems company, and our vision is to build the most performant and efficient computing systems possible by reinventing how we build them from first principles. We have previously shown how this allows for large scale AI to be run in a way that pushes the frontier of performance, while being more efficient. The same paradigm shift will change how we run any computing workload at scale, expanding beyond AI to fields such as scientific computing and any other high performance computing workload.

To change how we build and run computing systems, Callosum works together with a large network of partner companies and organizations across the globe. Callosum acts as the global heterogeneous integrator: We work closely with each and every one of our partners to bring out the unique capabilities of each of their technologies. We do this by bringing them into the systems-level-architecture in which their technology is strongest and ready to commercially enter the market.

"By integrating Cerebras into Callosum's platform, we're making ultra-low-latency inference available exactly where it creates the greatest impact, enabling customers to build AI systems that simply weren't practical before," said Andrew Feldman, CEO of Cerebras.

What we build with our partner companies is larger than the sum of its parts: by building new computing systems, we can run workloads at scale that were not possible before, increasing demand for all computing technologies. In order to do this, we look at our relationships with our partner companies not only as a set of 1-to-1 partnerships. Instead, we make new connections between complementary partners that have not worked together actively so far, where Callosum's technology is the integrator that brings these together across the globe. For some chips and hardware, this might come in the form of the AI workflow specialisation that a novel chip architecture is particularly suited to, but this extends much further to identifying complementarities as a function of hosting requirements, cooling characteristics, and energetic requirements.

Rebellions is a strong example of this global integrator role in practice: building AI systems where the right chip handles the right part of the workload, and turning that into a global partnership that reaches beyond any single market. As Sunghyun Park, CEO of Rebellions, put it: "Rebellions built our architecture around inference efficiency. Working with Callosum puts that architecture into systems alongside hardware chosen for different parts of the workload, instead of asking one chip to do every job. That's the difference between a partnership and a deployment that only works in one environment or geography."

The European chip company Axelera AI brings a complementary piece to that same picture: chip-level performance that needs to be matched by careful co-design around the physical realities of deployment, from power draw to thermal management. As Louis Mather, Vice President of Strategy at Axelera AI, put it: "Axelera AI builds the world's most efficient AI inference chips that bring better performance per watt to datacenters globally. Partnering with Callosum connects our technology with their unique ability to co-design systems that are optimised for both performance and modern server hosting characteristics, meaning power, thermal, and operational requirements, giving customers an outstanding experience wherever they deploy."

The goal of building a new kind of computing system is to bring it to all computing workloads around the world. To do so, we make these systems available via API-hosted workloads as highlighted in today’s product release of Tailored Inference, working with our hyperscaler and cloud partners in the backend. However, the beauty of heterogeneous computing systems is that they are adaptable and can use the right technology to fit into any deployment environment. Hence we are tightly collaborating with leading OEM and systems partners like Supermicro, alongside chip companies with a long history of enterprise deployments like Intel to allow enterprises to host their own on-premise heterogeneous computing systems, while also allowing them to coordinate their workflows across their local and cloud resources in a hybrid-cloud setup, to balance out cost, performance, privacy, expandability, and controllability constraints. This results not only in controllable infrastructure but improved token economics and performance.

Supermicro shows what this flexibility looks like at the infrastructure layer, giving customers a consistent way to deploy on-premise, across multiple cloud providers, or in a hybrid mix of both, without being locked into any one path. As Vik Malyala, Chief Business Officer at Supermicro, put it: "Supermicro is excited to work with Callosum and their cloud-agnostic orchestration frameworks for CSPs and high-level deployments. Our collaboration with Callosum is designed to give customers greater flexibility and faster deployment across multi-cloud and heterogeneous environments, while preserving choice across providers." Intel has been a key partner on the inference side of that same equation, where the improved token economics come into play. As Nicolas Dube, Intel Corporate Vice President, Data Center Systems and Solutions, put it: "The market has shifted to deliver fit for purpose inference with the right model size and context for every prompt. With agentic AI driving an acceleration of token spend, added to the growing adoption of open models, deploying an orchestrator that can route to heterogeneous systems is becoming critical for every customer. Callosum are leading the path to enable this transition. Intel is delighted to be partnering with Callosum as they enter their next rapid stage of growth!"

Heterogeneous integration becomes especially important as compute infrastructure projects grow more complex and ambitious. Callosum doesn't just provide the software for these projects; through our expertise of benchmarking, integrating, and running on novel hardware, we also act as a design partner throughout the entire lifecycle of a new compute cluster build. That kind of long-term backing is exactly what the UK Government's Sovereign AI Fund had in mind when Callosum became its first ever investment. As Joséphine Kant, Investment Partner of the Sovereign AI (SovAI) Fund put it: "We are extremely excited that Callosum was the first company we backed through the SovAI Fund. They're a key part of how we're approaching the UK's next generation of frontier computing infrastructure, bringing both systems design expertise and the software layer needed to make heterogeneous computing work in practice.”

Hardware and Chips

Heterogeneous compute is a key ingredient to how we envision future computing systems ought to be built differently. Hence we have built what is one of the largest networks of chip vendors around the world. With the established ones, we collaborate on how to integrate them into heterogeneous workloads of the future. With new and upcoming ones, we partner throughout the entire development cycle of their hardware, being both a helpful development partner while prototypes are in production, and being ready to deploy the chips in our own infrastructure as soon as they are ready, with our software ecosystem able to handle them from day one of their go-to-market motion.

One partner we work tightly with is Normal Computing, whose thermodynamic computing architecture is exactly the kind of new silicon this partnership model is designed to support. Normal Computing and Callosum work together in a holistic engineering partnership, covering hardware deployment, workload optimisation, and software integration. Faris Sbahi, CEO of Normal Computing, has described what that partnership means in practice:"The data center is going to look completely different in two to three years. We're going to have hundreds, if not thousands, of different kinds of chips tailored to specific workloads. New silicon needs to be integrated into the systems that AI workloads run on, and that's exactly the problem Callosum is solving. They've been a design partner as we deploy our thermodynamic computing architecture.” That same partnership model applies just as directly to Tendrils, a new CPU company we work with from the earliest stages of hardware development. As Nils Cremer, Co-Founder of Tendrils Compute, put it: “Callosum provides us with the opportunity to iterate on and deploy our prototypes during the earliest stages of chip development”. From thermodynamic architectures to next-generation CPUs, these partnerships show how Callosum meets new silicon wherever it is in its development journey.

Beyond these highlighted partnerships, we work with many chip companies around the world. To start with our local ecosystem, the UK's hardware sector is growing rapidly, and we are excited to be working closely with emerging companies including Lumai, Tendrils, Oriole, Quantum Dice, and Optalysys. The US naturally has a major ecosystem of chips and hardware, and we are excited to work with many partners across substrates and specialisations, spanning both compute and networking, including Cerebras, d-Matrix, Intel, SambaNova, AMD, Normal Computing, Tenstorrent, GreatSky, and Mixx. Our partnerships extend well beyond the UK and US too, spanning Europe, Asia, and Australia. Cortical Labs in Australia and Rebellions in Korea were among our earliest partners as we expanded into this wider ecosystem. Since then, we've continued to grow across these regions, including partnerships with Axelera and FinalSpark in the EU and Switzerland, and Furiosa in Korea.

Original Equipment Manufacturers

Alongside our chip partnerships, we work closely with OEMs to build the physical infrastructure to turn heterogeneous silicon into deployed systems. Nation states and enterprises are increasingly seeking greater resilience from and more control over their AI inference infrastructure. Heterogeneous systems address these challenges by design. As the global heterogeneous integrator, Callosum is involved from the earliest stages as a design partner and stays involved during construction to later provide the cluster software and tailored workflows and inference services. Together with our partner Supermicro we are in a position to shape infrastructure projects from conception to deployment, for both enterprise and sovereign deployments.

Clouds and Infrastructure Partners

Like everyone else, we consume significant compute to host our advanced systems, and we’ve built deep relationships with the largest hyperscalers, using their resources for hosting while also working with them on deeper integration of our heterogeneous inference strategies into workloads running on their infrastructure. AWS was the first major cloud provider Callosum built on. Their infrastructure gives us access to an extremely diverse mix of compute, including their own silicon in the form of Trainium. "We're proud to support Callosum as they scale their heterogeneous compute platform on AWS. Their approach to orchestrating workloads across diverse hardware, including AWS Trainium, represents an exciting new direction for how infrastructure can be built and optimized. Congratulations to the team on this milestone." Nicolas Tarducci, Head of Solution Architecture for Startups EMEA, AWS

In order to provide the best heterogeneous integration across resources we sometimes need to deploy just the right node in just the right location with the right networking connections to other partners. In order to support such specialised setups, we have built relationships with infrastructure partners, specifically partnering with Era4, Deliverance AI, and Computacenter to support such builds across the energy, cluster, and rack infrastructure.

Public Entities

A new way of computing is not worth much if it isn't available to the students, researchers, and engineers who will build on it. To get our technology into the hands of the next generation of builders, we work closely with universities, public research organisations, and national infrastructure projects. We start this closer exposure to our internal technology with select partners in the UK, including the UK Government through the UK SovAI Fund, the public research initiative CommonAI through the Scaling Inference Lab, and leading university and research computing centres such as the University of Cambridge's Open Zettascale Lab, UCL's Advanced Research Computing Centre, the University of Bristol's Center for Supercomputing, and the University of Edinburgh's Parallel Computing Centre.

We could not build our technology without the many partnerships that we have built over the last couple of months, and we are grateful for the time and trust each of our partners have put in us. The partnerships above are what's ready to be shared today, with many additional ones to follow in the coming weeks and months.