Thread · 3 stories · Jan 5 – Jun 22 · concluded
Nvidia's Blackwell successor: launched in January on a 10x inference-cost claim, expanded at GTC into seven chips, formally shipped in June.
Jump to timeline ↓Vera Rubin arrived in January claiming 10x lower inference costs, 4x fewer GPUs for MoE training and 5x better energy efficiency than Blackwell. At GTC 2026 it grew into a seven-chip platform for robotics, inference and agentic workloads, alongside a projection of $1 trillion in cumulative orders through 2027.
In June Nvidia formally launched the platform in its CPU-plus-GPU form, pitched at converging AI with high-performance computing — the hardware that would later train GPT-6 Astra.
