Petaflops

• Nvidia launched Blackwell Ultra B300 GPUs on April 25, 2026, at GTC US, delivering 20 petaflops FP4 performance for training models up to 10 trillion parameters. • New chips feature 4.5TB HBM4 memory and NVLink 6 interconnect, reducing training costs by 40% for hyperscalers like Microsoft Azure. • Announcement bolsters Nvidia's 92% AI accelerator market dominance amid US export controls on China sales.

• Nvidia unveiled Blackwell Ultra GPUs on April 9, 2026, at GTC US keynote, delivering 50 petaflops for AI training in a single rack. • Chips target US hyperscalers with 4x inference speed over prior gen, shipping Q4 2026 at $50,000 per unit. • CEO Jensen Huang called it 'the dawn of agentic AI,' boosting Nvidia's dominance in $200B US AI hardware market.