AI for Science: Up to 1,000,000 GPU Hours Now Open to UK Research Teams
The Department for Science, Innovation and Technology (DSIT) has opened a compute call offering 200,000 to 1,000,000 GPU hours on AI Research Resource systems. The focus: AI-led projects that push scientific discovery in engineering biology, frontier physics (including nuclear fusion), materials science, medical research, and quantum technology. Proposals that develop new AI models or virtual systems for future automated and autonomous scientific discovery are in scope. Deadline: 4pm, Sunday, December 21, 2025.
Who can apply
- UK-based applicants only
- Universities and research organizations eligible for UKRI funding
- Public sector research organizations
- Charities
- Registered businesses
What's on offer
- 200,000 to 1,000,000 GPU hours on AI Research Resource systems
- Support for high-ambition AI projects that enable automated and autonomous scientific discovery
- Priority areas: engineering biology, frontier physics (e.g., nuclear fusion), materials science, medical research, quantum technology
- Proposals in AI-driven research and scientific discovery are also eligible
Why this matters for your lab
This is serious compute for training frontier models, scaling simulations, and building virtual labs. Think foundation models for materials and molecules, inverse design for catalysts, high-fidelity surrogates for plasma control, multimodal models for medical imaging, or quantum error mitigation learned via RL. If your team has a dataset, a hypothesis, and a plan to scale, this call removes the usual compute ceiling.
Plan your GPU request with numbers, not optimism
- Map out model size, precision (FP16/8-bit), batch size, and expected training steps.
- Budget for pretraining, finetuning, hyperparameter sweeps, ablations, and evaluation.
- Example conversions:
- 1,000,000 GPU hours ≈ 1024 GPUs for ~41 days, or 512 GPUs for ~81 days, or 256 GPUs for ~163 days.
- 200,000 GPU hours ≈ 512 GPUs for ~16 days, or 256 GPUs for ~32 days, or 128 GPUs for ~65 days.
- Include contingency (10-20%) for retries, scaling issues, and longer-than-expected convergence.
Data, governance, and reproducibility
- Data rights: confirm licenses and consent; outline any sensitive data handling.
- Security: note export controls, bio-safety, and clinical governance where applicable.
- Reproducibility: specify seeds, exact commits, environment snapshots, and logging.
- Efficiency: show that you've considered sparsity, LoRA/QLoRA, mixed precision, and checkpointing to use compute responsibly.
- Outputs: plan for datasets, checkpoints, code, and benchmarks (open or restricted as appropriate).
What a strong proposal looks like
- Clear scientific question with concrete AI contribution (model, method, or virtual system).
- Specific compute plan with milestones and success metrics (loss, validation score, physical fidelity, clinical endpoints).
- Well-defined datasets, data quality checks, and augmentation/simulation strategy.
- Validation plan beyond accuracy: calibration, uncertainty, robustness, and domain shift tests.
- Risk register: safety, misuse, dual-use concerns, and mitigation steps.
- Team readiness: roles, prior results, infra experience, and maintenance plan for artifacts.
Example project directions (to spark scope, not limit it)
- Engineering biology: generative protein design with structure constraints; lab-in-the-loop active learning.
- Frontier physics: fusion plasma control surrogates; differentiable simulators for reactor design.
- Materials science: LLMs for crystal graphs; multi-objective inverse design with property predictors.
- Medical research: multimodal pretraining on imaging + EHR; foundation segmentation for rare pathologies.
- Quantum technology: error correction policy learning; noise-aware compilers via sequence models.
Key dates
- Submission deadline: 4pm (UK time), Sunday, December 21, 2025
Useful references
Quick prep checklist
- Eligibility confirmed and organization type matches the call.
- Scope fits priority areas or AI-driven scientific discovery.
- Compute plan with realistic scaling schedule and fallbacks.
- Data governance, ethics, and safety risks addressed.
- Reproducibility and artifact release plan written.
- Clear impact for UK science and your field.
If your team needs to level up fast on training at scale, curriculum design for AI upskilling, or toolchains, see curated options at Complete AI Training.
Your membership also unlocks:
