NVIDIA GTC 2026 opens
NVIDIA’s GTC 2026 opened in San Jose today — Jensen Huang is set to unveil new AI inference chips and CPUs, including the Vera Rubin accelerator that targets CPU bottlenecks. The conference also confirmed HBM4 supply links with Samsung and SK Hynix, signaling tighter memory-compute co-design for next-gen AI hardware opens reported.
Vera’s CPU is an 88‑core Arm‑based “Olympus” design with 176 threads and FP8 support, which NVIDIA bills as the platform’s memory-and-coordination engine (nvidia.com). Each Rubin GPU is paired with sixth‑generation HBM4 stacks providing roughly 288 GB of on‑package HBM4 and about 22 TB/s of memory bandwidth per GPU in early public specs and previews (tomshardware.com). NVIDIA’s Rubin platform release states “extreme codesign” across Vera CPU, Rubin GPU, BlueField‑4 DPU, ConnectX‑9 SuperNIC and NVLink to deliver up to 10× lower inference token cost and 4× fewer GPUs for MoE training versus Blackwell, per the company press summary. (investor.nvidia.com) Multiple industry reports say NVIDIA designated Samsung Electronics and SK Hynix as exclusive HBM4 suppliers for Vera Rubin, with SK Hynix taking about 70% of the allocation and Samsung about 30%, while Micron is reportedly excluded from the initial Vera Rubin HBM4 roster (kedglobal.com). NVIDIA has begun shipping Vera Rubin samples to select customers and public coverage places full production and broader shipments in the second half of 2026 as partners complete qualification testing (tomshardware.com). Industry coverage also details supply‑chain moves: Samsung reportedly completed HBM4 verification runs at target rates and is moving to mass production while NVIDIA teams have visited packaging sites as SK Hynix ramps wafer allocation for the Vera program. (parameter.io)
