Jibang Consulting: DRAM supply continues to be tight in 2027 Nvidia (NVDA.US) evaluates the reduction of Rubin Ultra HBM configurations

The Zhitong Finance App learned that according to TrendForce Jibang Consulting's latest memory industry research, since the DRAM shortage pattern will continue until 2027, and there are variables in the verification progress of the original HBM4e, starting in the third quarter of 2026, NVIDIA (NVDA.US)'s HBM configuration for Rubin Ultra has changed from HBM4e 12hi to parallel evaluation of designs such as HBM4e 8hi, HBM4 12hi, and HBM4 8hi, etc., which have not yet been finalized. In addition to NVIDIA, some cloud service providers (CSPs) are also considering reducing the HBM capacity design of the next generation of self-developed ASICs.

TrendForce Jibang Consulting said that the recent shortage of memory supply has caused the DRAM specifications of AI chips to be lowered several times. Since the first half of 2026, CSP and server OEMs have successively lowered RDIMM capacity; recently, NVIDIA also considered that the LPDDR5X shortage will continue until 2027, and decided to halve the SoAMM capacity of next-generation Vera Rubin SuperChip modules.

For Rubin Ultra, Nvidia used HBM4e 12hi as the benchmark design from 2025 to the first half of 2026, and was open to evaluate other lower specifications until the beginning of the third quarter, and discussions are ongoing. This move mainly addresses two supply-side bottlenecks: First, overall DRAM supply will remain tight in 2027, which will limit the wafer production capacity that the original factory can allocate to HBM. Second, there is still uncertainty about the HBM4e 12hi verification and mass production yield improvement process.

TrendForce Jibang Consulting pointed out that NVIDIA's main goal in the Rubin Ultra generation is to increase I/O speed, while the secondary goal is to pursue growth in the scale of GPU shipments. If it is finally decided to downgrade the HBM specification, it is estimated that the trend will be to start by adjusting the number of DRAM stack layers (DRAM Stack Layers).

In summary, whether the HBM4e can complete verification and mass production as scheduled will determine whether the I/O speed of the Rubin Ultra can be upgraded from 8-11.7 Gbps of the previous generation Rubin to 14-16 Gbps, or whether it can only be improved to 11-12 Gbps through HBM4 design optimization. On the other hand, in a given generation, the number of stack layers will determine the balance between the HBM carrying capacity of a single GPU and the number of GPUs that can be shipped.

TrendForce Jibang Consulting believes that subsequent specification finalization will also depend on the original wafer supply allocation. As far as supply and demand are concerned, it is estimated that HBM's shipping level will increase by 50-60% per year in 2027, which is still insufficient to meet demand side growth. In the face of a pattern of short supply and demand, TrendForce Jibang Consulting estimates that in 2027, HBM's bargaining power will still be biased in favor of suppliers, and the sharp rise in HBM unit prices has become an industry consensus. AI chip manufacturers will face the double pressure of shrinking HBM supply and rising procurement costs, and the motivation to adopt lower capacity HBMs will increase accordingly.