Nvidia tests lower-memory Rubin Ultra GPU designs amid HBM shortage

Nvidia ↗ 2.27% has pushed back against a wave of reports suggesting its next-generation Rubin Ultra GPU and Vera CPU platforms are being quietly scaled down, telling Tom's Hardware that "Our roadmap is intact." The denial comes as the chipmaker navigates an intensifying shortage of high-bandwidth memory that has forced it to evaluate alternative designs for its flagship 2027 AI accelerator.shanethegamer+1

Testing Lower-Memory Variants

According to The Information, Nvidia has been testing at least three versions of the Rubin Ultra GPU over recent weeks, some featuring less memory than the company originally announced at its GTC 2025 developer conference, where CEO Jensen Huang said each chip would include 1TB of HBM4E memory spread across 16 stacks. The company is considering lower-memory configurations in part because it may not be able to secure enough advanced memory chips to support the original design.tradingview+1

Separately, SemiAnalysis reported in late June that Nvidia cancelled the original four-compute-die design of Rubin Ultra in favor of a dual-die layout that is easier to manufacture, citing concerns about substrate warpage when connecting four large dies and sixteen memory stacks on a single package. The Kyber NVL144 rack meant to house 144 Rubin Ultra chips has also reportedly slipped more than 12 months to 2028.linkedin+2

Huang's Public Denial

Speaking at a developer event in Tokyo on July 15, Huang directly addressed the delay narrative. "The reports are not true. Vera Rubin is already in production. Giant amounts of production incoming," he told reporters, according to Bloomberg. However, as Tom's Hardware noted, Huang did not specifically address reports about Rubin Ultra or the Kyber NVL144 rack system — a distinction that has kept some analysts skeptical.finance.yahoo+3

Memory Costs Compound the Pressure

The HBM supply crunch extends beyond Rubin Ultra. TrendForce reported in late July that Nvidia is halving the SOCAMM memory modules on its Vera CPUs inside the Vera Rubin NVL72 rack, cutting from 192GB to 96GB per CPU to contain costs. Without the adjustment, memory could account for roughly 29% of the system's estimated $2.1 million bill of materials, well above Nvidia's preferred 20% threshold.trendforce+1

For AI companies, less memory per GPU could mean deploying more chips to run large models — a dynamic that may benefit memory suppliers like Micron Technology ↘ 0.44% even as it complicates the broader AI infrastructure buildout.tradingview