NVIDIA AI Updates: October 20, 2026

1. Germany Signs a Physical AI Declaration With NVIDIA as Huang Names Four New German Startup Investments

NVIDIA. At the Invest in Germany Summit in Berlin, Chancellor Friedrich Merz, the Federal Ministry for Economic Affairs and Energy, and the Federal Ministry for Research, Technology and Space signed a Joint Declaration of Intent with NVIDIA to work on a joint initiative to advance AI models for physical AI. The declaration formalizes NVIDIA’s work with the government’s “Industry 4.0 Platform,” whose stated goal is to bring industrial AI into widespread use across German manufacturing. Jensen Huang also announced planned investments in Agile Robots (robot arms and the Agile ONE humanoid), Avelios Medical (an AI-native hospital information system), Mirelo (AI-generated sound effects, speech, and music for video), and Reflex Aerospace (satellites), part of a nearly $10 billion NVIDIA commitment across Europe. NVIDIA also cited a Siemens showcase at its Erlangen electronics factory, where a fully virtual production ramp-up cuts engineering effort by up to 85%. The announcements precede Huang’s GTC Berlin keynote on October 21. Source

2. NVIDIA Runs a 104-Qubit Fault-Tolerant Quantum Simulation on 8,192 GPUs at Jülich

NVIDIA. NVIDIA researchers simulated a 104-qubit fault-tolerant quantum processor on JUPITER, Europe’s first exascale supercomputer at the Jülich Supercomputing Centre, scaling to 8,192 GH200 GPUs. The run modeled magic state cultivation, a technique for producing the magic states consumed in fault-tolerant computation that NVIDIA says needs 90% fewer qubits than earlier methods, under realistic hardware noise (prior Jülich work simulated 50 qubits without noise). The circuit was expressed with the CUDA-Q fault-tolerant logical layer and run with cuQuantum, after scaling tests on a GB200 NVL72 at the NVIDIA Accelerated Quantum Research Center. NVIDIA calls it the largest simulation of its kind to date and says it lets researchers study noise effects on error-correction designs before hardware exists. Source

3. European Broadcasting Union Expands Its Sovereign AI Framework With NVIDIA, Microsoft, and Nebius

NVIDIA. The European Broadcasting Union (more than 115 member organizations in over 55 countries) is expanding the sovereign standards it introduced with NVIDIA last year to cover sovereign cloud providers, open model developers, and systems integrators. The initiative will define requirements and reference architectures for where broadcaster agent workloads run, how data is governed, which models are used, and how systems stay secure. Microsoft contributes Sovereign Cloud options including Azure Local and Foundry Local for offline processing, Nebius offers European infrastructure for training and serving models, NVIDIA Nemotron serves as the customizable open model base, and Qvest handles integration. At the same Berlin event, German producer UFA described using Adobe Frame.io’s GPU-accelerated semantic video search to find scenes across 15 years of footage. Source

4. Jane Street and NVIDIA Publish cuTile Mamba-2 Forward Kernels Up to 1.82x Faster Than Triton

NVIDIA. Jane Street engineers, working with NVIDIA, implemented the Mamba-2 forward pass in CUDA Tile (cuTile), NVIDIA’s Python DSL for tile-level GPU programming. The three kernels (chunk_cumsum, fwd_pass_state, and a fused compute_out) run 1.14x to 1.82x faster than the open-source Triton kernels in mamba_ssm 2.3.2.post1 on a single B300 GPU, with larger gains at longer sequences and bigger state sizes. Benchmarks used PyTorch 2.13.0, CUDA 13.3, and cuda-tile 1.5.0. These are kernel-level results only; a follow-up post will cover the backward pass. The code is in NVIDIA’s TileGym repository. Source