Nvidia is reportedly considering a significant reduction in the amount of high-bandwidth memory (HBM) packed into its next generation of AI chips, as the cost of that memory climbs.
According to a report from finance.biggo.com, circulated via Google News, Nvidia is weighing cuts to next-gen AI chip HBM capacity "by up to 81%" as memory costs soar. A separate report from South Korean outlet 조선일보 (Chosun Ilbo) similarly states that Nvidia is trimming HBM capacity amid soaring costs.
Both items point to the same underlying pressure: HBM, the specialized stacked memory that sits alongside the processor in modern AI accelerators, has become an expensive bottleneck. Neither source item provides detail on which specific chip generation is affected, when a decision would be finalized, or which memory suppliers would be involved.
It's worth being clear about the limits of what's known here. These are press reports, not an Nvidia announcement, and the headline figure — up to 81% — is a reported upper bound on what's being considered, not a confirmed specification. Nvidia has not publicly confirmed the plans described.
Why it matters in plain terms: HBM is what lets an AI chip feed data to its processing cores fast enough to keep them busy, and how much of it a chip carries shapes how large a model that chip can hold and how quickly it can run. A meaningful cut would be a visible sign that memory prices are now steering the design of the industry's most important hardware — with knock-on effects for the memory makers who supply Nvidia, the cloud companies that buy its chips, and ultimately the cost of running AI services.