Nvidia reportedly cancels quad-die Rubin Ultra GPU in favor of dual-GPU design, report claims — complex design purportedly scrapped over 'manufacturing execution concerns'

According to SemiAnalysis, Nvidia had originally planned to launch its Rubin Ultra AI accelerator in 2027 with a quad-chiplet design aimed at doubling performance. However, due to manufacturing execut... Read original →

Nvidia’s pivot from a quad-chiplet Rubin Ultra to a dual-GPU design isn’t merely a yield concession—it signals that 3nm chiplet integration has hit hard limits in both physics and supply chain execution. This delays HBM4E’s volume adoption, forcing memory suppliers like SK Hynix to recalibrate capacity plans. Liquid-cooled Kyber racks now serve as the performance backstop, marking a strategic shift from per-GPU FLOPS to system-level thermal orchestration in AI data centers. AMD stands to gain: its MI500 series could capture premium training workloads, especially as U.S. and EU policies favor locally resilient AI infrastructure with flexible chiplet ecosystems. Crucially, this reversal underscores that advanced packaging can’t indefinitely compensate for scaling bottlenecks. Over the next 18 months, the industry will pivot from ‘more chiplets’ to ‘smarter architectures,’ redirecting capex from wafer fabs toward cooling and power delivery systems.