How we built the UK's sovereign AI supercomputing stack

Isambard-AI has been developed using open-source software, open standards and community-driven technologies. This approach provides the flexibility to adopt future innovations without rebuilding the entire platform.

When the Bristol Centre for Supercomputing (BriCS) began work on Isambard-AI in late 2023, there was no data centre, no operational team, and no established national AI infrastructure. Less than two years later, at launch, Isambard-AI became one of the world's most powerful AI supercomputers, ranked amongst private and public compute alike.

The achievement extends far beyond hardware performance. Alongside delivering one of Europe's leading AI systems, BriCS has built and deployed an open, sovereign software stack designed to give the UK long-term control over its AI infrastructure while remaining flexible enough to evolve with a rapidly changing technology landscape.

Beyond a supercomputer

The global AI ecosystem is changing at extraordinary speed. New models, frameworks and specialised accelerators emerge almost weekly, while traditional infrastructure procurement cycles typically operate over years.

Rather than building around proprietary, vendor-specific technologies wherever possible, Isambard-AI has been developed using open-source software, open standards and community-driven technologies. This approach provides the flexibility to adopt future innovations without rebuilding the entire platform.

Researchers can access resources through familiar tools such as Jupyter notebooks, PyTorch, TensorFlow, Hugging Face and vLLM, while the underlying platform combines technologies including Kubernetes, Slurm, OpenCHAMI and open federated identity systems.

This layered architecture ensures that the UK's AI infrastructure remains adaptable, portable and resilient as AI technologies continue to evolve.

Delivering AI infrastructure as a service

A key innovation has been the development of AIRRPortal.

AIRRPortal is the single front door to national AI resources. Once a project is approved, accounts, storage and permissions are provisioned automatically, in seconds. Funded through the AI Research Resource, tiered access calls run from first-time Gateway grants of up to 10,000 GPU hours and SME Rapid Access of up to 20,000 to national AI for Science missions of up to a million GPU hours, part of a £1 billion plan to grow AIRR capacity at least twentyfold by 2030. 

Historically, gaining access to high-performance computing facilities could involve complex application processes and lengthy provisioning times. AIRRPortal is designed to remove these barriers.

Researchers, academics, start-ups, SMEs and public sector organisations can apply for resources through a unified system that connects multiple national facilities. Once approved, accounts, storage allocations and access permissions can be provisioned automatically in seconds.

The goal is simple: reduce the time between a research idea and access to compute from months to days.

As the AIRR ecosystem grows, AIRRPortal is expected to become a marketplace for a wider range of AI services, including inference platforms, trusted research environments, federated learning services and national AI data infrastructure.

Serving a growing national community

The impact of the platform is already being felt across the UK research and innovation landscape.

Thousands of users have accessed resources through the AIRR programme, supporting hundreds of projects spanning healthcare, drug discovery, climate science, advanced materials, robotics and AI development.

Importantly, access has been designed to extend beyond traditional academic research communities. The UK's national AI infrastructure now supports a broad spectrum of users, from government and universities through to start-ups and small businesses.

This reflects a central principle behind the programme: world-class AI infrastructure should be a national asset that supports scientific discovery, economic growth and innovation across society.

A decade of co-design

The success of Isambard-AI builds on more than a decade of innovation at Bristol.

The original Isambard system was the first production supercomputers built around Arm processors, helping demonstrate that alternative architectures could compete at scale in high-performance computing.

Bristol's influence also extends into networking technologies. The Slingshot interconnect used in modern supercomputing systems traces part of its heritage to Gnodal, the Bristol networking company acquired by Cray in 2013.

This history of co-designing technologies, rather than simply consuming them, continues to shape BriCS' approach today.

The objective is to deploy powerful hardware and help shape the technologies, software and services that underpin future computing infrastructure.

Open source as national infrastructure

A defining feature of the Isambard-AI programme is its commitment to contributing back to the wider community.

Several key components developed by BriCS as part of the sovereign stack have been released as open source software, including:

  • Clifton, a secure SSH connection management system.
  • Conch, a federated SSH certificate authority.
  • OpenPortal, a standardised interface connecting user portals and research infrastructure.
  • Contributions to OpenCHAMI, the Linux Foundation's open infrastructure management initiative.

These projects are already being reused by organisations beyond Bristol and are helping establish common approaches for next-generation research infrastructure internationally.

This model reflects a broader philosophy: national capability grows strongest when it is shared.

Supporting the UK's Sovereign AI ambitions

The launch of the UK's Sovereign AI Unit and wider government investment in national AI capability highlight the growing importance of secure, nationally controlled digital infrastructure.

Isambard-AI demonstrates how sovereign capability can be built around openness rather than isolation.

By combining open technologies, community collaboration and operational excellence, the platform offers an alternative to both proprietary vendor lock-in and dependence on overseas cloud infrastructure.

The result is a flexible foundation capable of supporting future generations of AI hardware, software and services while remaining aligned with national priorities.

A foundation for the future

Isambard-AI is the beginning of a broader ecosystem: a sovereign, open and evolving AI infrastructure designed to support research, innovation and economic growth across the UK, all delivered at pace.

As the AIRR expands, new services are integrated, and emerging AI technologies mature, the principles established through Isambard-AI of openness, portability, collaboration and public value will continue to guide the UK's approach to national AI infrastructure.