NVIDIA's $250M compute gift for UK AI agent safety

NVIDIA is donating $250 million in compute to UK researchers

On October 13, 2026, NVIDIA pledged five years of compute, worth $250 million in total, to UK research into the safety and security of AI agents. Compute here means the GPU processing power that AI models need to train and run. Researchers will tap it through the cloud rather than receive hardware or money.

It is compute access, not cash

The donation works out to $50 million of compute access each year. Five years of that adds up to the $250 million headline figure. The machines will sit in the UK, and researchers will reach them through cloud providers that partner with NVIDIA.

Most of it goes to the wider UK research community

The bulk of the compute is meant for the broad community of UK researchers working on AI safety and security. A smaller share goes to three groups:

• BOLD, led by the University of Oxford
• SOFAIR, led by UCL (University College London)
• Imperial College London

Huang said in the announcement that AI capability and safety have to move forward together. He framed the pledge as a way to let UK researchers build and test safeguards for AI systems that keep getting more powerful.

The research focuses on keeping AI agents safe

All of the donated compute is earmarked for work on AI agent safety and security. An AI agent is a system that takes a goal from a person, plans its own steps and carries them out. NVIDIA lists three priorities for the program:

• Secure agents
• Security for open models
• Defenses against threats seen in real deployments

What BOLD, SOFAIR and Imperial will work on

Each of the three groups comes at agent safety from a different angle. BOLD is building methods that let AI systems learn and adapt while working with people. SOFAIR studies new open source AI architectures and easier ways to build and deploy AI.

Imperial's share covers safety engineering and robotics. UK AI Minister Kanishka Narayan called the pledge a vote of confidence in British AI research and tied safe, secure AI to the UK's G20 presidency.

Sandboxes, benchmarks and NVIDIA's own safety tools

The compute will also cover the shared infrastructure that safety research depends on. NVIDIA names three uses:

• Secure sandboxes, isolated environments where agents can run without touching real systems
• Reproducible research
• Benchmarks that score how safe a model or agent is

Some of the work will run on NVIDIA OpenShell, the company's open runtime for agents, and on its Sentry reference design. NVIDIA says it wants the findings to feed into safeguards across the wider computing industry.

Huang also previewed $2 billion for Europe

Speaking in Paris on October 13, Huang said NVIDIA would announce $2 billion in new European deals within about a week. Bloomberg reported the remarks, and The Decoder carried them the next day.

The Decoder puts NVIDIA's European spending over the past year at about $8 billion. Tech Funding News reported that the company sizes its total European commitment at nearly $10 billion, without breaking it down into equity and infrastructure.

Four French startups get NVIDIA stakes

In Paris, Huang named four French startups that NVIDIA is backing. Tech Funding News reports that all four are minority stakes, and NVIDIA did not disclose any amounts:

• Doctolib, a healthcare booking platform
• Genesis AI, a robotics developer
• Pigment, which makes AI-driven business planning software
• Entalpic, which uses AI to discover new materials

Huang argued that open-weight models, whose trained parameters anyone can download and modify, give Europe a way to catch up. He also pushed back on calls to slow AI development.

Five UK startups are next

On October 13, NVIDIA also used its blog to announce planned investments in five UK startups. Here is where each is based and what it builds:

The blog gives no amounts for these stakes either. The $250 million compute donation is a separate research program, not part of the startup investments.

Key terms are still undecided

The details of both the UK donation and the European money are still unsettled. For the UK, NVIDIA has fixed the size and the rough delivery route, and little else.

Cloud providers and commercial terms are open

NVIDIA says three parts of the UK program have yet to be agreed. The release lists them as:

• Which cloud providers will host the compute
• The technical architecture
• The commercial terms

What is fixed is that the compute will run through UK-based NVIDIA Cloud Partners on infrastructure inside the UK. NVIDIA presents the pledge as a follow-on to its work with the University of Bristol on the Isambard-AI supercomputer.

The $2 billion has no named targets yet

As of October 16, 2026, NVIDIA had not said where the $2 billion for Europe will go. Huang is scheduled to give a keynote at GTC Berlin on October 21.