NVIDIA AI Updates: May 25, 2026

1. Huang heads to TSMC to lock in Vera Rubin capacity as CoWoS ramp dictates the second half of 2026

NVIDIA. Jensen Huang’s Taipei stop on the way to GTC Taipei is centered on a meeting with TSMC chair C.C. Wei to secure Vera Rubin production allocation, with NVIDIA already pre-committing more than 50 percent of TSMC’s CoWoS advanced-packaging capacity through 2027. TSMC is scaling CoWoS output from roughly 35,000 wafers per month to a 120,000 to 140,000 monthly range by the end of 2026, and capacity is described as sold out through 2025 and well into 2026, leaving little room for non-NVIDIA orders. Huang also addressed memory pricing pressure with SK Hynix, Micron, and Samsung as the six-chip Vera Rubin platform, which pairs the Vera CPU and Rubin GPU with NVLink 6, ConnectX-9 SuperNIC, BlueField-4 DPU, and Spectrum-X switching across roughly two million parts per system, enters volume production. Source

2. NVIDIA pitches Vera Rubin NVL72 as 3.5x training and 5x inference over Blackwell, with cost per inference cut to one-seventh

NVIDIA. The Vera Rubin NVL72 configuration ties together 36 Vera CPUs and 72 Rubin GPUs in a single rack and is being marketed at Computex with a 3.5x training and 5x inference improvement over the current Blackwell generation, alongside a roughly seven-fold reduction in cost per inference. That positioning underpins NVIDIA’s Q2 FY2027 guide of about $91 billion, which assumes zero China data center revenue after the company’s accelerator share there collapsed from around 95 percent to effectively zero and not a single licensed H200 has been delivered since the December 2025 framework. Huang’s public message in Taipei tied the ramp to about 150 Taiwanese supply-chain partners and warned of a very busy second half for the island, while reiterating that conceding the China market does not make strategic sense as Huawei’s Ascend 950PR targets 750,000 units and roughly $12 billion of 2026 revenue. Source