The United Kingdom aims to be a “AI maker, not an AI taker”. This is the rallying cry of the British AI Action Opportunities Plan and the message NVIDIA has reinforced during the visit of its founder and CEO Jensen Huang to the country: an ecosystem investing in sovereign computational infrastructure, applying physical and agentic AI in industry and robotics, and advancing
Here’s an overview of the most significant initiatives NVIDIA and its partners are driving across Britain.
1) Foundations: Isambard-AI and the computational backbone for sovereign AI
Funded by UK Research and Innovation and based on Grace Hopper (GH200), Isambard-AI — housed at the University of Bristol — has become the backbone for numerous national projects:
- UK-LLM: a collaborative initiative of UCL, the University of Bangor, and NVIDIA that trains reasoning models with Nemotron to support Welsh and English. The goal is to enhance public service delivery (healthcare, education, legal resources) in the languages of approximately 850,000 speakers.
- Nightingale AI: a multimodal foundational health model (Imperial College London) trained on clinical data from the UK and US, aimed at early diagnosis and personalized care.
- PolluGen: a high-resolution pollution dispersion model (University of Manchester) using NVIDIA CorrDiff and Earth-2 Studio to inform citizens and regulators about air quality.
- Ultrasound Foundation Model: (Queen Mary University of London) focused on ultrasound imaging and rheumatoid arthritis, with the ambition of creating a publicly reproducible model.
- Gen Model in Ego-Sensed World: (University of Bristol) analyzes visual data from >900 participants to understand everyday tasks and predict real-world interactions; it could support memory and independent living for patients with dementia.
- Electrostatic Foundation Models: (University of Cambridge and NVIDIA) the first models to “understand” electrostatics in chemistry at the atomic level over >200 million structures (OMOL/OMAT) with cuEquivariance, opening doors to materials and molecular simulations previously unfeasible.
Beyond hardware, the talent gap is addressed through SCAN and the NVIDIA Deep Learning Institute, as well as SCAN Springboard U.K., training developers and professionals for specialized roles.
2) Physical AI and robotics: from XR teleoperation to modular humanoids
The UK roadmap emphasizes physical AI and safe automation:
- Extend Robotics: safe and scalable deployment in automotive sectors combining XR teleoperation and training systems, using Jetson AGX Orin, Isaac Lab, and Isaac GR00T (robot skills learning).
- Humanoid (HMND 01): a modular humanoid robot designed for warehouses and retail, meant to integrate naturally into human environments.
- Materials Innovation Factory (University of Liverpool): trains models to predict material properties, employing Jetson Orin Nano for “robot scientists” within a fully automated laboratory.
- National Robotarium: a national hub combining NVIDIA frameworks to push applied research and accelerate spin-outs.
- Opteran: algorithms inspired by comparative neuroscience (insects/animals) to give robots robustness and natural efficiency.
- Oxa: full-stack autonomy for industrial and commercial fleets utilizing NVIDIA DRIVE, massive generation of synthetic data for training and validation in scenarios with unreliable GPS.
- Wayve: AV2.0, an end-to-end deep learning system capable of generalization to unseen environments without expensive sensors or HD maps.
Common pattern: deploying AI at the edge with platforms like Jetson, coupled with simulation, synthetic data, and digital twins to speed up real-world validation.
3) Life sciences: AI-first approaches to drug design and regulatory twins
A notable cluster of British biotechs and labs adopts an AI-first approach:
- Basecamp Research (BaseData): an evolving dataset 10× larger than public sources, powering foundational models in biology for programmable medicine.
- CEiRSI (University of Manchester): complex digital twins (on NVIDIA) for testing treatments across diverse populations.
- Isomorphic Labs: a drug design engine built on multi-therapeutic foundational models.
- Peptone: physical AI to exploit the entire proteome, focusing on intrinsically disordered proteins (historically “inedulcorables”).
- Latent Labs (Latent-X): in silico generative AI to create and test therapeutic molecules.
- Relation Therapeutics: a lab-in-the-loop platform for disease target discovery and fast-tracking drugs.
- Hologen AI: precise modeling of human biology and medical interventions to reduce times and costs.
- Oxford Nanopore: fast, rich, and affordable DNA/RNA sequencing for research and clinical use.
The message: biomedical data sovereignty, scaled on DGX and Grace Hopper, with microservices via NIM powering optimized models at lower cost and latency.
4) Agentic & generative AI: productivity, voice, and conversational agents
From financial LLMs to voice agents:
- Aveni: a financial LLM built on NeMo with agentic agents interacting with live systems, serving clients, and advising on risk with compliance and control.
- ElevenLabs: hyperrealistic AI voices in over 70 languages, supported by DGX B200, for real-time agents, accessibility, and localization.
- PolyAI: conversational agents at scale using Riva ASR and NIM, capable of authentication, orders, billing, and phone reservations.
- Recraft: image generation/editing with TensorRT, professional creative workflows (marketing, mockups, graphics).
- Speechmatics: multilingual ASR on Dynamo-Triton and cuDNN.
- Synthesia: enterprise video platform (training, sales, support) with avatars and voice-over in over 140 languages, optimized with Dynamo-Triton.
The focus is on deploying agents and models into production with low latency, low cost (via quantization, TensorRT), and interoperability through NIM microservices.
5) Celtic languages and public services: a Welsh LLM with national ambitions
The UK-LLM project (initially known as BritLLM in 2023 and led by UCL) has released two previous models for UK languages and now presents a Welsh reasoning model trained with NVIDIA Nemotron (49B Llama Nemotron Super and 9B Nemotron Nano). The team crafted a new Welsh dataset by translating over 30 million entries from open corpora using NIM for gpt-oss-120b and DeepSeek-R1, training on DGX Cloud Lepton and hundreds of GH200s at Isambard-AI.
- Language validation: Bangor University (Gwynedd, the county with the highest % of speakers) provides cultural expertise and verifies translations and nuances (e.g., initial consonant mutations).
- Deployment: provider Nscale will expose the model via API for businesses and public sector.
- Public agenda: Prime Minister Keir Starmer emphasizes that reasoning in Welsh will enable bilingual public services and preserve heritage (the Cymraeg 2050 plan aims for 1 million speakers).
Methodologically, the Nemotron + NIM framework is reusable for other minority languages (Cornish, Irish, Scots, Scottish Gaelic) and even for African and Southeast Asian languages.
Why this matters: becoming an AI stack maker, not just a taker
The British plan is supported by three pillars:
- Investing in fundamentals (computing, data, talent, regulation),
- Driving cross-sector adoption (pilots and rapid scaling in public and private sectors),
- Being an “AI maker” with national champions in key layers of the AI stack (models, tools, hardware, services) to capture economic benefits, influence values and safety, and strengthen sovereignty.
NVIDIA plays the role of technology partner, orchestrating hardware (GH200, DGX, Jetson, DRIVE), software (NeMo, Riva, TensorRT, Dynamo-Triton, cuDNN, cuEquivariance), and services (NIM, DGX Cloud), while the UK’s academic-business ecosystem supplies data, use cases, and regulatory validation.
Challenges: energy, costs, talent, and reliability
- Energy and sustainability: GH200 and Isambard-AI clusters require stable power and efficient cooling. The footprint must be mitigated through renewables, heat reuse, load optimization (e.g., quantization), and data center efficiency.
- Total cost of ownership: balancing CapEx/Opex for supercomputing with public value (health, education, climate), prioritizing use cases with high social return.
- Talent: expanding training in systems, data, security, and MLOps (DLI, universities, bootcamps) to meet demand.
- Reliability and safety: strengthening assessment, alignment, data protection, and compliance (NHS, MHRA, ICO), especially for clinical multimodal models and autonomous agents.
What to watch in the next 12-24 months
- Public sector deployments: pilots of UK-LLM Welsh in healthcare and education; adoption by municipalities and legal services.
- Robots in production: XR teleoperation with skill learning (Isaac GR00T) in automotive and retail; expanded testing of modular humanoids.
- Health and pharmaceuticals: first regulatory twins and scaled in silico trials; integration of Oxford Nanopore with discovery pipelines.
- SME ecosystem: proliferation of conversational agents (Riva/NIM), voice (ElevenLabs), ASR (Speechmatics), and AI video (Synthesia) in banking, insurance, retail, and citizen support.
Conclusion
The UK is building a comprehensive sovereign AI strategy that combines cutting-edge computing, domestic data and languages, applied robotics, and AI-first biomedicine, with NVIDIA as a strategic partner. The outcome is not just a portfolio of projects; it’s a national architecture to produce — not just consume — the next generation of AI technology.
If the plan succeeds, we will see bilingual public services with AI, collaborative robots learning quickly in factories, better-informed diagnostics in hospitals, and SMEs capable of deploying agents and generative workflows within weeks, not years. The goal is clear: create local value with AI that reasons at home, speaks national languages, and respects regulatory frameworks.
via: blogs.nvidia and gov.uk