Twelve months after Isambard-AI switched on, Prof. Simon McIntosh-Smith, Director of the Bristol Centre for Supercomputing (BriCS) sat down with Emma Hindley MBE, Deputy Director for Public Compute and the AI Research Resource (AIRR) at the Department for Science, Innovation and Technology (DSIT), for a special anniversary episode of BriCS's podcast, Node to Node, hosted by Dr Claire Thorne of Deep Science Ventures.
What followed was less a progress report than a genuinely candid conversation about what it took to build the system, what it's already being used for, and what a scientist's job starts to look like when an AI can run the analysis, write the code, and flag the mistakes in data before even being asked.
Listen and watch the video podcast, or read on.
A crazy idea
The build itself was a fast-moving, high-pressure affair. McIntosh-Smith described a “crazy idea” at the outset, a small but formidable team at BriCS backed by the innovative leadership at University of Bristol, and delivery partners HPE (who built the system) and NVIDIA (who supplied its GPUs) rising to a deliberately ambitious challenge, to build one of the world's most powerful AI supercomputers, in record time.
Sustainability was built in from day one and at launch, Isambard-AI has ranked the fourth most energy-efficient supercomputer ever built anywhere in the world. BriCS pays a premium to run it entirely on UK renewable energy, across solar, wind, hydro and battery, with nothing else in the mix.
Why does sovereignty matter?
For McIntosh-Smith, sovereign AI compute stopped being a nice-to-have the moment access to major commercial AI models started being switched off overseas with no warning. If a government or a business is relying on that access for something critical, that's a serious vulnerability, which is exactly why the UK needed the ability to build and run its own AI capability. Before Isambard-AI existed, entire categories of AI research simply weren't possible in the UK — training a large language model from scratch, for instance. A year in, that threshold has been crossed, and crossed at pace: the system has already welcomed close to 4,000 users, a ramp-up McIntosh-Smith admitted has outstripped even BriCS and DSIT’s own expectations. "It was all pent up," he said, "waiting for something like this to come along".
Stories from a sovereign AI supercomputer
When asked to pick out the projects that had stayed with them over the year, both guests reached across the country and across disciplines.
For McIntosh-Smith, one of the earliest and most personally memorable was UK-LLM, a UCL-led project by Prof. Pontus Stenetorp, training one of the UK's first homegrown large language models. Before Isambard-AI officially opened, the team were already queuing up to use it, building AI language support for Welsh, Scots Gaelic, Cornish and other British languages routinely overlooked by the large international models, alongside domain-specific work like English case law.
Closer to home, he pointed to a Bristol project monitoring the health of dairy cattle using cameras dotted around a farm. Cows instinctively hide signs of illness to avoid predators, making early diagnosis notoriously difficult, but AI trained to spot subtle behavioural changes can now alert a farmer before a problem becomes serious, benefiting animal welfare and farm economics alike. It's one of McIntosh-Smith's personal favourites, not least because livestock health is far from the first place people expect to find cutting-edge AI at work.
Hindley's standout was Prima Mente, which is developing models to help doctors understand how the brain changes during the earliest stages of Alzheimer's, in far greater detail than has previously been possible. Having spent three years of her own PhD in neuroscience advancing a similar question by increments, she was candid about the scale of the shift, saying ‘I spent three years advancing knowledge by a fraction of what this can do in an hour.’
The project's use of sensitive NHS blood test data was also, Hindley noted, a clear example of why a secure, trustworthy sovereign system matters, as patients simply wouldn't want that data handled anywhere less safe.
Both guests highlighted two projects from the UK AI Security Institute (AISI) that had piqued their interest. The first, run jointly with Anthropic and The Alan Turing Institute, tested how easily a language model's training data could be deliberately corrupted, known as a “data poisoning” attack. The result surprised everyone involved as it took only a few hundred documents to compromise a model, far fewer than expected, a finding McIntosh-Smith described as invaluable precisely because it shows the AI community what to guard against. The second project examined how depictions of AI fed into a model's training data shape its later behaviour, with Hindley drawing a striking parallel to her own work on inclusive education, that role models matter, and they matter earliest of all.
Elsewhere, the two discussed projects investigating heart disease at the cellular level in search of new drug targets, and public health research spanning respiratory illness, child mortality, and even air pollution. All of this work made possible, both agreed, by the sheer scale of NHS health data available in the UK, a resource McIntosh-Smith said UK researchers' international peers are frankly envious of. A large early project led by Imperial College London, Nightingale AI, combines this data at a scale rarely attempted before.
Inspirations
For all the scale of the science, some of the most memorable inspirations are small in size. BriCS's Lego version of Isambard Kingdom Brunel makes cameo appearances throughout the case study videos, there is a nod within BriCS's own name and the supercomputer itself is built in a genuinely modular way.
McIntosh-Smith brought a physical prop to the recording - an actual NVIDIA GH200 Grace Hopper Superchip, one of 5,448 installed inside Isambard-AI, worth in the region of £25,000 each. Despite housing that much computing power, the system's data hall occupies a room only around 12 metres square and surprises many of those who have a tour, a fraction of the football-pitch-sized data centres such power once required.
Of course, none of this happens without people. BriCS has grown from nothing to a team of around 35 people, in 2 years and McIntosh-Smith was quick to single out Dr Sadaf Alam, the centre's Chief Technology Officer, as one of the field's leading visionaries in supercomputer design, “the brains,” in his words, behind what BriCS has actually built. Both Simon and Sadaf were recently featured in the SCW75 list of 75 influential global figures in computer science.
What's next?
BriCS is preparing to launch an inference service, allowing the system to serve AI models directly to a wide range of users, an increasingly critical capability as demand for fast, affordable AI "tokens" grows. Fittingly, some of the underlying research making that possible has come from Doubleword, one of the companies backed by the UK's Sovereign AI Fund, working on how to generate those tokens faster and more cheaply on Isambard-AI itself.
The Bristol AI Data Facility, BRAID, is currently under construction. This is a large, high-speed data storage system being installed alongside the main supercomputer. And behind the scenes, BriCS is already planning what comes after Isambard-AI, which is after all, the fourth in a series of Isambard supercomputers built at University of Bristol over more than a decade, each one bigger, faster and more efficient than the last. Whether the next iteration becomes the world's most powerful machine, is now less a technical question than a question of appetite for investment, and power supply.
Every call for access to Isambard-AI has been oversubscribed so far, a sign of real demand rather than a reason to slow down. The plan is to keep expanding access, not narrow it, meeting the UK’s need for AI compute to drive economic growth and real societal benefit.
Apply for time | Press enquiries | LinkedIn | Instagram | YouTube | Bluesky