NVIDIA AI Updates: September 25, 2026

1. Predicted Structures for 2,800 Viruses Landed in the AlphaFold Database

NVIDIA, Google DeepMind, and EMBL-EBI released predicted 3D protein structures covering more than 2,800 viruses, with roughly 30 percent of the protein interactions new to science. The structures came from AlphaFold2 running on NVIDIA’s BioNeMo Inference Runtime, predicting viral protein complexes on GPUs instead of through crystallography. NVIDIA also open sourced the BioNeMo Structure Prediction Pipeline so researchers can run the same process on their own targets, and the dataset is free through the AlphaFold Database with per-prediction confidence labels. Partners include CEPI, Seoul National University, Sungkyunkwan University, the Swiss Institute of Bioinformatics, and the University of Glasgow. Source

2. Fusing the MoE Kernels Doubled Biological Model Training Throughput

NVIDIA published a BioNeMo recipe for mixture-of-experts training that attacks three specific costs. Transformer Engine’s GroupedLinear submits all expert matmuls as one grouped GEMM rather than looping per expert, MXFP8 block-scaled 8-bit precision cuts memory against BF16 with hardware acceleration on Blackwell, and a Sequential API fuses GroupedLinear, ScaledSwiGLU, and routing-weight scaling into a single kernel so quantized intermediates never materialize. On eight B200 GPUs with expert parallelism and FSDP2, the recipe hit 9,050 tokens per second per GPU against 4,096 for the Hugging Face baseline, a 2.21x speedup. Source