In July 2025, the UK's most powerful AI supercomputer switched on in Bristol.
Isambard-AI represented a £225 million bet that the UK needed its own sovereign compute power to compete on the global AI stage. This power was placed into the hands of UK researchers, SMEs and start-ups tackling the problems that matter most.
A year on, that bet is paying off. Powering around 1000 projects from over 4000 users, BriCS’ Director, Simon McIntosh-Smith, admitted usage has outstripped even BriCS's own expectations. "Demand was pent up," he said, "waiting for something like Isambard-AI to come along".
Here are ten of those user stories.
1. Seeing inside the cell to treat heart disease | University of Bristol | Dr John Lees & Dr Danielle Paul
Cardiomyopathy is the most common inherited heart disease, the leading cause of sudden death in adults under 35 and is often undetected until it's too late. Dr John Lees and Dr Danielle Paul, British Heart Foundation Research Fellow, are using Isambard-AI to understand how the proteins that build our cells can go wrong.
Danielle takes microscope images of real cells, while John uses Isambard-AI to computationally fill in everything the microscope can't see, reconstructing the parts of the cell invisible to direct observation. A simulation of this kind would take roughly 50 years on a standard research computer.
The pair and their teams are now looking to extend the work into peptide inhibitors, used to disrupt disease pathways in conditions from cancer to Alzheimer's, evidence that foundational protein research on one disease can ripple outward into many others.
2. Putting AI financial forecasting to the test | University of Manchester & UCL | Dr Ebal Rahimikia & Professor Hao Ni
Generalist "Time Series Foundation Models" promise to forecast anything, from heartbeats to house prices to financial markets, even without training on that specific domain. The FinText project team decided to test that claim.
Using time series data from 94 countries, the team built over 600 finance-specific models and ran them against real-world outcomes in a genuine zero-shot test, using 50,000 GPU hours on Isambard-AI to pre-train and fine-tune models with GPUs, then benchmarking the results on standard CPUs. Models trained specifically on financial data outperformed general-purpose forecasters, a finding with implications far beyond finance, into energy, healthcare and weather forecasting.
The project is fully open access, with models and methodology published for other researchers to build on.
Watch the case study | Read the case study
3. Teaching AI to understand human movement | University of Bristol | Professor Dima Damen
What if technology could prompt someone with dementia to complete a task at the exact moment they feel unsure? Professor Dima Damen's team is turning thousands of hours of first-person ‘point-of-view’ (POV) video into models that recognise actions like "cooking" but further understand the fine-grained detail of how an action happens for an individual - the specific hand grasp, the stage of the task, the level of skill etc.
One hour of video contains as much data as hundreds of thousands of pages of text. The team has already used 180,000 GPU hours on Isambard-AI, building a 3D digital twin of how people live their lives.
This could be life-changing for the elderly or those with dementia as they can be prompted to complete the next step if memory fails them. The team are also focusing on the future of robotics, with their in-depth models able to train humanoid robots in years to come.
As Dima puts it, ‘Isambard-AI is so central to our work that if it goes on holiday, my students can too.’
4. Building sovereign language models for British languages | UCL Centre for Artificial Intelligence | Professor Pontus Stenetorp
Most commercial LLMs are built around American English. UK-LLM is building large language models from scratch that speak the UK's own languages, including Welsh, Irish, Scottish Gaelic and Cornish, working with partners like the Welsh-language Unit at Bangor University.
Creating a multilingual LLM from the ground up is something only a handful of groups in the world can do.
Beyond language coverage, the team is probing a fundamental question - how much do LLMs generalise from their training data, versus simply memorise it? Answering that has already taken more than half a million GPU hours on Isambard-AI, laying groundwork that other non-English-speaking nations can use to build their own sovereign AI.
5. Mapping the air we breathe | University of Manchester | Professor David Topping
In partnership with NVIDIA, Professor David Topping's team took Earth-2, a generative AI system built for European weather and climate modelling, and retrained it to model air pollution across the UK, at unprecedented resolution.
The team compressed Earth-2's continental 50km-resolution view down to a 2–3km UK-specific model, turning a process that once took days into one that runs in seconds to minutes, using just two days of compute on a single 8-GPU node. It's a striking contrast to some of Isambard-AI's more compute-hungry projects, and fittingly light-touch for an environmental initiative.
The ambition is global: give any city, anywhere, the ability to generate its own street-level pollution model from local data and using just a small compute budget.
6. Spotting disease in dairy herds | University of Bristol Veterinary School | Professor Andrew Dowsey
Cattle are prey animals which means they instinctively hide signs of illness. That makes early detection of conditions like mastitis and lameness nearly impossible for even the most experienced farmer watching 200 animals at once.
Professor Andrew Dowsey's team is using 60 unobtrusive video cameras and Isambard-AI to watch a 200-strong herd around the clock.
The AI has learned the subtle behavioural tells that give away an ailing cow, such as withdrawing to the herd's edges, grooming less, receiving more head swipes from neighbours. The AI can even identify individual all-black cows with no coat markings by body shape alone. The goal is a commercial system built on just a handful of simple cameras that any farmer could use, supporting agricultural efforts from the farm up, across the UK.
7. AI security - Can science fiction make AI misaligned with human interest? | Geodesic Research with the UK AI Security Institute (AISI) | Puria Radmard
If an LLM's pre-training data is full of stories where AI is depicted as deceptive or dangerous, does that shape how the model behaves once it "learns" it is an AI? Geodesic Research set out to find out, building open-source language models from scratch to control every part of the pre-training process.
The team manufactured fake news, research articles and fiction depicting AI both positively and negatively, then fed them into pre-training data. Bad role models did indeed produce misaligned behaviours like deception and power-seeking; good role models produced the opposite, and it took just 0.1% synthetic "positive AI" data to see a strong effect.
Since the project completed, major AI labs have begun publishing on pre-training safety, with OpenAI using Geodesic's own synthetic datasets in their experiments.
8. AI security - How easy is it to poison an AI model? | UK AI Security Institute, with Anthropic and The Alan Turing Institute | Alex Souly
Modern AI models learn from trillions of words scraped from the open web. In principle, anyone who can publish text online could try to plant a hidden vulnerability. This joint study investigated exactly how much malicious data it takes to install a "backdoor" in a language model during pre-training.
The prevailing assumption was that bigger models would need proportionally more poisoned data to compromise. The study using Isambard-AI found the opposite - model size barely mattered. As few as roughly 250 malicious documents were enough to install a working backdoor across a wide range of model sizes, a genuinely surprising result that reframes how the AI safety community thinks about training-data risk.
9. Diagnosing dementia from a blood test | Prima Mente, with Imperial College London | Hannah Madan & Dr Ivan Koychev
Dementia is the UK's leading cause of death, and progress against it has been held back by a basic problem - brain tissue has traditionally only been able to be analysed after death. Hannah Madan and Dr Ivan Koychev are working with a network of 20 NHS trusts, to change this, cross-referencing fragments of brain cells that leak into the bloodstream with post-mortem tissue samples to build foundation models that can read what's happening in a living brain.
The company's first model, Pleiades 1, can already diagnose Alzheimer's from a simple blood test. Its successor, Pleiades 2, is a 100-billion-parameter model trained on Isambard-AI using 80 trillion tokens across five data types. It aims to identify the roughly 25 distinct subgroups the team believes exist within Alzheimer's, each potentially needing its own targeted treatment. Access to Isambard-AI's one million GPU hours came through the UK's Sovereign AI Fund. Using UK sovereign compute was a critical factor for the project, given the sensitivity of NHS patient data involved.
10. Cancer Pathologists assisted by AI clinician | King's College London | Dr Chris Banerji & Dr Greg Verghese
UK hospital trusts process around 100,000 pathology cases a year, generating roughly two million enormous digital images that need rapid scanning for signs of cancer. Speed is one challenge, but trust is another, as most cancer-detection AI today is a "black box," offering a probability with no explanation or rationale.
Banerji and Verghese's system is built to reason the way a human pathologist does. First, identifying cancer type and stage, then weighing that against a patient's individual characteristics, rather than pattern-matching on pixels alone.
Crucially, the AI has to present its diagnosis and reasoning to a real clinician, who can correct it if necessary, feeding that correction back to refine future recommendations. As a sovereign, NHS-secure resource, Isambard-AI made it possible to train on sensitive patient data without it ever leaving UK shores.
Ten projects, one AI supercomputer, and an extraordinary range of problems, across heart disease, dementia, cancer, air pollution, animal welfare, financial forecasting, British languages, and the frontier questions of AI safety itself. What connects them is sovereign UK compute, allowing each team to do work that they couldn't do anywhere else, faster, at a scale previously impossible and without sensitive UK data ever leaving the country.
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.