NVIDIA compute to test AI agent safety: $250M for Britain over five years

How it was before. On Jun 23, 2026, the UK government gave BOLD and SOFAIR up to £60 million over six years, plus £2 million each to hire at least ten PhD students. The plan was doubled at the time: from one lab and £40 million to two labs and £60 million.

NVIDIA compute now comes on top: £189 million at the ECB reference rate of Oct 13, 2026, or about £38 million a year. It will be delivered through UK NVIDIA Cloud Partners hosted in the country, but there is no application process yet, and according to Jensen Huang, the provider, architecture and commercial terms are still being negotiated.

The stack is already open. On Sep 28, 2026, NVIDIA released the Open Agent Safety Platform. OpenShell, an open-source runtime, sets execution boundaries for an agent and is expanding to Arm and Intel platforms. Sentry runs as an out-of-band watchdog on the BlueField-4 DPU: it quarantines and stops an agent within milliseconds.

Both parts are available on NVIDIA's developer page and on GitHub. The platform lists more than 100 organizations, including Anthropic, Microsoft and Salesforce. The UK program builds on the same stack: the sandboxes and benchmarks are the same ones that are publicly available.

In the same move, the institute's access was confirmed: the AI Security Institute gets models from OpenAI, Anthropic and Google before release. AI minister Kanishka Narayan said so in London on Oct 12, 2026. The agreements have been voluntary since 2023, and the decision rests with the developer each time.

UK government scientists are now reviewing the latest generation of models, including Gemini 4 Argon. Earlier, the institute evaluated OpenAI's GPT-6 Astra.

The government's £60 million is scheduled through 2032, and NVIDIA's compute through 2031.