Nvidia trims high-bandwidth memory allocation across product stack amid soaring HBM costs and persistent supply constraints.

NVIDIA, the world's largest artificial intelligence semiconductor company, is considering significantly reducing the capacity of high-bandwidth memory (HBM) used in its next-generation AI chips—a move that could slash memory capacity by up to 81% compared to initially announced specifications, or 33% below the previous generation. Even NVIDIA has struggled to manage HBM supply shortages and soaring prices, driving what amounts to a strategic "memory diet."

NVIDIA previously announced that its Rubin Ultra AI chip, scheduled for 2027 release, would feature 16 units of 12-layer HBM4E (seventh generation), delivering 1TB of total memory. However, since Q3 this year, the company has been testing products with significantly reduced specifications—either 8-layer HBM4E or the older-generation HBM4 (sixth generation). According to U.S. IT media The Information, some prototypes have total memory capacity of just 192GB, representing an 81% reduction from the original plan. Compared to the current mass-produced Rubin's 288GB capacity, this constitutes a 33% cut. NVIDIA has not yet officially finalized Rubin Ultra specifications.

The motivation is clear: memory costs have skyrocketed. In NVIDIA's Grace Blackwell (GB300) server product, memory accounted for 9.4% of total cost—$373,939 out of $3.9946 million. For the next-generation Vera Rubin (VR200), total cost increased 95% to $7.8031 million, while memory costs surged 435% to $2.0016 million, rising to 25.7% of total cost. TrendForce, a market research firm, noted: "AI chip suppliers face a dual challenge of limited HBM supply and high costs, creating strong incentives to adopt low-capacity HBM configurations."

This "memory diet" extends beyond servers. Microsoft, Dell, and Acer have recently reversed the trend of higher base memory by relaunching 8GB laptop models, finding it difficult to pass rising component costs to consumers. Google is reportedly reducing its upcoming Pixel 11 smartphone's base memory from 12GB to 8GB. Q2 shipments reflected this pressure: smartphone shipments fell 11% year-over-year; PC shipments declined 4%.

The immediate industry impact may be limited. HBM is essential for maximizing AI accelerator performance—reducing it excessively would require connecting more chips, potentially raising overall system costs. TrendForce forecasts 50–60% HBM shipment growth in 2027 but expects supply to struggle keeping pace with demand. General-purpose DRAM and NAND flash in PCs and smartphones face similar constraints, as the rise of "on-device AI"—running inference directly on consumer hardware—demands more memory.

Pricing concerns persist, however. SK Group Chairman Chey Tae-won recently warned: "If prices keep rising without building new factories, the market will shrink, and new players will inevitably enter. We need to increase supply to lower prices."