Nvidia (NVDA) is weighing a radical step to deal with a shortage of advanced high-bandwidth memory chips, The Information reports; using less of it than planned in its next-generation GPU, the Rubin Ultra

• Recently, Nvidia has been testing at least three versions of the Rubin Ultra GPU, some of which have less memory than what the company initially announced, according to sources.
• Nvidia is considering the lower-memory versions partly because it may not be able to secure enough advanced memory to supply its original design, sources said.

Memory supply has been the binding constraint in the AI accelerator build-out for some time, and reports of a lead designer re-specifying a product around component availability fit a familiar pattern in past shortage cycles: vendors downgrade configurations, tier their SKUs, or reallocate scarce parts to flagship parts rather than halt production. The precedent in prior HBM tightness has been that allocation flows to the largest buyers first, so a signal that even the dominant customer is short of advanced memory points to capacity, not demand, as the marginal driver, with read-through to the concentrated memory supplier set and to second-tier accelerator makers further down the allocation queue. A reduced-memory configuration is a spec compromise with direct performance implications for memory-bound workloads, so the distinction worth drawing is between a tiered product strategy and a forced downgrade, since the former is margin-neutral segmentation and the latter concedes supply failure. Reports of this kind, sourced to a single outlet ahead of any confirmation, have historically been refined or partially walked back, and the follow-ons are any company or supplier comment, qualification timelines from the memory makers, and whether the original spec survives at launch. As unconfirmed reporting it sits in the category where initial price action has often outrun subsequent fact.