NVIDIA Cuts HBM Capacity for Rubin Ultra Amid Memory Shortage! How Chipmakers Are Coping

AI chip giant NVIDIA (NVDA-US) is quietly adjusting the design blueprint for its next-generation flagship GPU. The company has recently tested multiple "downsized" configurations for the new Rubin Ultra GPU, with some configurations having less HBM memory capacity than originally planned. This highlights how the AI boom is creating memory shortages that are now affecting chipmakers themselves.

Insiders reveal that NVIDIA has tested at least three versions of Rubin Ultra with modified HBM specifications in recent weeks. At last year's developer conference, CEO Jensen Huang announced that each Rubin Ultra chip would feature up to 1TB of HBM4e memory through a 16-stack architecture for unprecedented computing performance.

However, market research firm TrendForce points out that NVIDIA has deviated from its original plan of a single 12-layer HBM4e specification, now evaluating three alternative options: 8-layer HBM4e, 12-layer HBM4, and 8-layer HBM4. The final specification has not yet been decided.

TrendForce analyzes that two main factors are driving this specification reassessment. First, the industry widely expects the overall DRAM industry to face supply tightness in 2027, which will inevitably compress wafer capacity available for HBM production. Second, the 12-layer HBM4e is still in the testing and verification stage, and its yield rate after mass production remains uncertain.

Beyond HBM, memory shortages are also affecting other product lines. TrendForce reveals that due to the expected LPDDR5X chip shortage extending to 2027, NVIDIA has decided to halve the SOCAMM memory module capacity in its Vera Rubin superchip module.

Industry insiders point out the irony of this situation: it's precisely NVIDIA's own AI chips selling so well that have pushed the entire industry's memory demand to new heights, now causing supply shortages that are backfiring on NVIDIA itself, forcing the company to modify memory configurations to adapt.

The effects of memory shortages aren't limited to NVIDIA. TrendForce notes that major cloud service providers are also evaluating downward adjustments to the HBM capacity of their self-developed AI chips.

The firm also estimates that while global HBM shipments in 2027 are expected to grow by 50-60% year-over-year, this still won't be enough to fill the actual market demand gap.

In this supply-demand imbalance, TrendForce assesses that HBM suppliers will maintain pricing power throughout 2027, while downstream AI chipmakers will have to endure the dual pressures of "not being able to get inventory" and "rising costs."

The reduced memory capacity will have some impact on end users. Analysts believe that the reduced memory capacity could affect the chip's actual computing performance, and if companies use these downsized versions to train or run large AI models, they may need to purchase more GPUs than originally planned to make up for the performance gap.

Semiconductor research firm SemiAnalysis points out that the cost savings from HBM memory may be redirected to switches and optical interconnect components. By combining tiered storage architecture with optical interconnect technology, they aim to mitigate the impact of high HBM prices and tight supply.

This strategy aligns with Jensen Huang's previous public statement that when supply chains get stuck, while it gives upstream suppliers pricing advantages, it also forces companies to come up with innovative technical solutions to break through.

While NVIDIA may be able to partially compensate for the performance loss from reduced memory capacity by increasing GPU computing power and optimizing interconnect bandwidth, industry observers generally believe that overall system construction costs and cluster setup complexity will only increase.

For cloud giants like Microsoft, Meta, Amazon, and Google that are continuing to invest heavily in AI infrastructure, this also means that the construction costs of future data centers may need to be further revised upwards.