The platform is designed as a whole: CPU, accelerator, switches and network chips built together. In parallel, every hyperscaler is building its own.
About this eventSix new chips as one AI supercomputer: Vera CPU, Rubin GPU, NVLink 6, ConnectX-9, BlueField-4, Spectrum-6.
Rubin shows that the unit of progress is no longer one GPU but the whole rack: compute, memory, networking and control are co-designed. NVIDIA's claimed gains apply to the platform and selected workloads; actual cost and efficiency will become clear after cloud-partner deployment.
Source: NVIDIAAbout this eventMeta said its new accelerator is already in production and is the first of four in-house silicon generations being deployed over two years.
About this eventBy GTC 2026 the platform grew to seven chips in production; partner availability lands in H2.
About this eventGoogle, Amazon and Microsoft brought their own accelerators into cloud and production systems, widening the choice beyond NVIDIA GPUs.
About this eventThe bottleneck shifted from compute to data delivery: 2026 server demand is constrained by DRAM and NAND supply, while most HBM3E capacity was allocated in advance.
About this eventCopper can no longer carry the required connectivity — interconnects move to light.
About this eventSilicon designed for a single architecture rather than a class of tasks: multiples of efficiency at the price of total inflexibility.
Primary material behind this dossier: papers, lab publications and official reports.