NVIDIA AI Updates: June 9, 2026

1. NVIDIA Details UK Sovereign AI Buildout Spanning Compute Partners and Inception Startups

NVIDIA outlined progress on the UK’s sovereign AI infrastructure one year after the country committed to becoming an “AI maker, not an AI taker.” Partners are expanding compute capacity, including Nebius scaling to 65 megawatts across three sites alongside BT and Nscale collaborations on sovereign data centers. Four NVIDIA Inception startups, Cosine, Cursive, Doubleword, and Prima Mente, are using the Isambard-AI supercomputer, built on 5,400 NVIDIA GH200 Grace Hopper Superchips, to develop coding platforms, self-improving systems, optimized inference, and biological foundation models. Source

2. NVIDIA Adds NVFP4 Training Support to JAX and MaxText for Blackwell

NVIDIA introduced NVFP4 training support in JAX and MaxText for its Blackwell architecture, reporting up to 1.73x speedup over FP8 baselines while maintaining accuracy. The NVFP4 format applies 4-bit mixed-precision quantization to transformer MLP layers while keeping attention blocks at higher precision to avoid degradation from softmax noise amplification. Empirical validation on Llama 3 8B and Llama 3.1 405B showed negligible final loss differences against the throughput gains on GB200 and GB300 hardware. Source