Nvidia rolls out Rubin GPU, Vera CPU and pursues US dark‑fiber network

Nvidia unveiled its next‑generation GPU architecture, dubbed Rubin, which will be fabricated on TSMC’s 3 nm N3P process. The design features 144 streaming multiprocessors (18,432 CUDA cores), HBM4 memory up to 288 GB and bandwidth exceeding 4 TB/s, with a planned market launch in the second half of 2026 and mass production in 2027. Rubin is part of a broader “Vera Rubin” platform that also includes Nvidia’s Vera CPU – an 88‑core processor aimed at AI agents – and BlueField‑4 DPUs, with the platform already in production ramp‑up across more than 350 factories in 30 countries and over 300 partner deployments.

Separately, research firms reported that Nvidia is acquiring extensive "dark fiber" assets across the United States. The planned network could involve up to 100 fiber pairs per route, offering an estimated capacity of about 7.6 Pb/s, and may require $5‑10 billion of investment over the next three years to build a proprietary high‑speed AI backbone.

Nvidia CEO Jensen Huang also emphasized Taiwan’s role as the global epicenter of electronic manufacturing and AI supply chains, citing the island’s advanced semiconductor, packaging and AI‑server capabilities as critical to the industry despite new factories abroad. Analysts note that surging AI compute demand is reshaping the optical‑module market and driving structural upgrades in high‑speed interconnects.