目录
TL;DR
I. Mainstream Rubin Ultra Memory Reduced to 192GB
II. Max-Q Becomes Mainstream, While Max-P Retains Higher-Power Headroom
III. From 1TB HBM4E to 192GB HBM4
IV. HBM Supply and Cost Pressures Drive Memory Downgrade
V. Power Shortages Limit the Practical Value of Higher-Power Versions
本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读
The mainstream Rubin Ultra version is shifting its upgrade focus from single-chip performance to NVL576 scale-up. Tight HBM supply, data-center power shortages, and cost pressures are jointly prompting NVIDIA to lower memory and power specifications.
TL;DR
According to supply-chain sources, the mainstream Rubin Ultra SKU being previewed to key customers will retain peak theoretical FLOPS and HBM4, but memory will be reduced to 8-Hi and 192GB, below Rubin’s 12-Hi and 288GB; memory bandwidth will increase only slightly.
The mainstream Rubin Ultra version will have chip-level power consumption of approximately 1800W, the same as the initial Rubin products and below the 2300W two-piece-lid version. Under the current roadmap, its only clearly defined upgrade over Rubin is expanding the scale-up domain to NVL576.
NVIDIA may still launch a 2600–2800W Max-P version and a 1200W version for lower-compute workloads such as token decoding, but the 1800W Max-Q will be the mainstream product, focused on optimizing FLOPS per watt.
Rubin Ultra has undergone successive specification reductions from the 4-die, 1TB HBM4E 16-Hi configuration initially previewed at GTC 2025: it was first reduced to HBM4E 12-Hi, then the 4-die MCM was canceled, and it ultimately retained Rubin’s 2-die package while further reducing mainstream memory capacity to 192GB.
The roadmap adjustment reflects the intersecting constraints of HBM supply, power availability, and data-center deployment conditions. Reducing memory capacity helps expand GPU shipments under limited HBM supply while balancing relatively less-constrained TSMC front-end capacity against tighter HBM supply.
The downgrade can also substantially reduce bill-of-materials costs, offsetting the HBM price increases expected next year; lower power consumption should facilitate customer deployment, reduce power-delivery-system costs, and improve wafer yield.
From an investment perspective, Rubin Ultra’s core value is no longer a major uplift in single-card specifications, but system-level gains enabled by NVL576 scale-up. Key areas to validate include the NVL576 interconnect architecture, NPO deployment, customer adoption across power SKUs, and changes in HBM supply and pricing.
I. Mainstream Rubin Ultra Memory Reduced to 192GB
According to supply-chain sources, Rubin “Ultra” is being previewed to key customers with specifications below prior expectations. The mainstream Rubin Ultra SKU will maintain the same peak theoretical FLOPS and continue using HBM4, but its memory specification has been reduced to 8-Hi and 192GB, below Rubin’s 12-Hi and 288GB.
Rubin Ultra’s memory bandwidth will increase slightly, while chip-level power consumption will be approximately 1800W, the same as the initial Rubin products but below the 2300W two-piece-lid version.
This means that, under the current roadmap, Rubin Ultra’s only upgrade over Rubin is expanding the scale-up domain to NVL576. One planned system architecture for implementing NVL576 is to interconnect 8 racks, each equipped with 72 GPUs, using NPO for switch-to-switch cross-rack connections.
As in the past, NVIDIA is pursuing multiple development paths in parallel, but this system architecture is reportedly considered the most promising candidate at present.
II. Max-Q Becomes Mainstream, While Max-P Retains Higher-Power Headroom
Rubin Ultra’s logic die is designed to support higher power consumption. NVIDIA may still offer a 2600–2800W version, but the 1800W version will become the mainstream product. In addition, it may launch a lower-power 1200W SKU for workloads with lower compute requirements, such as token decoding.
SKUs above 2600W will be launched as Max-P versions, while the 1800W SKU will be classified as Max-Q.
Rubin has also used a similar product configuration: Max-P can operate at a higher thermal design power and deliver higher absolute FLOPS, while Max-Q operates at a lower thermal design power but is optimized for FLOPS per watt.
Source: SemiAnalysis
III. From 1TB HBM4E to 192GB HBM4
At the packaging level, Rubin Ultra continues the trend of successive specification reductions since it was first previewed at GTC 2025.
The initial design used 4 dies and featured 1TB HBM4E 16-Hi. Subsequently, the HBM4E specification was first reduced to 12-Hi, after which the 4-die MCM design was also canceled, leaving Rubin Ultra with the same 2-die package as Rubin.
Some of the specification reductions stemmed from manufacturing challenges. However, the current Rubin Ultra specifications are even below Rubin’s, indicating that these adjustments also reflect deliberate product trade-offs by NVIDIA.
Source: SemiAnalysis
IV. HBM Supply and Cost Pressures Drive Memory Downgrade
NVIDIA’s roadmap adjustment may be a response to the intersecting challenges of memory supply and data-center infrastructure availability.
On the memory side, this adjustment follows the same logic as the earlier reduction of Rubin Ultra from 16-Hi to 12-Hi and the reduction in the number of SOCAMMs allocated to each Vera: limiting memory usage per accelerator enables higher aggregate output under constrained HBM supply while balancing relatively less-constrained TSMC front-end capacity against tighter HBM supply.
Reducing the memory specification can also materially lower bill-of-materials costs, thereby offsetting the HBM price increases expected next year.
Even so, NVIDIA has secured more favorable per-capacity HBM pricing from SK hynix than other customers. In exchange, NVIDIA will provide GPU allocations for SK Group’s AI data-center plans.
V. Power Shortages Limit the Practical Value of Higher-Power Versions
On the power side, AI data-center models indicate that power shortages will persist, prompting NVIDIA to lower product power specifications to facilitate customer deployment.
Rubin Ultra’s tape-out project is code-named GR150, and its design can accommodate significantly higher power consumption than GR100, which is used in the first-generation Rubin, but most customers may be unable to fully utilize the additional power budget.
In power-constrained environments, the 1800W Max-Q SKU is the better choice because it can deliver the best FLOPS per watt. In other words, even if NVIDIA positioned the higher-power 2600W Rubin Ultra as the mainstream SKU, most customers would not operate near its maximum power consumption in practice.
NVIDIA can therefore reduce power-delivery-system bill-of-materials costs by lowering the power consumption of the mainstream version, while improving overall wafer yield by using a higher proportion of functional wafers.