Rubin Ultra's Shorter HBM Stacks May Shift the Bottleneck

On September 13, 2026, SemiAnalysis reported that Nvidia had decided to use 8-high HBM stacks for Rubin Ultra, cutting memory per GPU from 288GB to 192GB. If HBM wafer supply is the binding production constraint, the same wafer allocation could support more HBM stacks and accelerator packages, shifting incremental demand toward packaging, inspection and metrology.[1]

HBM, or high-bandwidth memory, stacks multiple DRAM dies vertically and places those stacks beside a GPU in an advanced package. A 12-high or 8-high designation describes the number of memory dies in each stack. More layers increase capacity per stack but consume more DRAM dies cut from the same wafer.

The reported Rubin Ultra configuration keeps eight HBM sites but moves each site from 12 layers to eight. That reduces the DRAM die count per GPU from 96 to 64 and cuts capacity by one-third while leaving the number of stacks unchanged. SemiAnalysis also reported that the shorter stack does not reduce bandwidth. Nvidia would therefore be trading capacity for the ability to build more GPUs, rather than simply accepting slower memory.[1]

How the bottleneck could move from wafers to packaging

If HBM wafers constrain 2027 accelerator output, moving from 12-high to 8-high stacks could produce roughly 1.5 times as many stacks from the same number of DRAM dies. Each stack still needs a base die, stacking and testing, and it still enters a 2.5D package with the GPU. HBM bit output would not rise, but work charged per stack and per finished package could increase by about 50%.[1]

Another constraint could interrupt the chain. On September 10, TSMC CFO Wendell Huang said CoWoS capacity remained tight and that TSMC was outsourcing some advanced packaging to outsourced semiconductor assembly and test providers. If packaging, substrates, power or logic dies are the true bottleneck, freeing HBM supply may not produce more finished accelerators.[2]

Companies that could be affected

Amkor Technology (AMKR) provides semiconductor packaging and testing, and management has disclosed a multiyear advanced packaging and test partnership with Nvidia. Advanced Products represented 82.8% of fiscal 2025 revenue. If accelerator package output rises, its 2.5D and high-density fan-out lines could benefit through utilization, revenue and operating profit. The thesis fails if planned accelerator units do not increase.[3]

Camtek (CAMT) sells inspection and metrology equipment used in advanced packaging. The company said that, of more than $600 million in orders booked during the first half of 2026, more than 50% came from OSATs and more than 20% from HBM manufacturers. More stacks and packages could require more back-end inspection steps and equipment orders, but the scale depends on whether customers actually add capacity.[4]

Onto Innovation (ONTO) supplies equipment for 2.5D packaging and three-dimensional HBM metrology. The company reported $343 million of second-quarter 2026 revenue, with about half from Advanced Packaging and Specialty Devices. It also received a $200 million Dragonfly G5 order from one OSAT customer, mostly scheduled for 2027 delivery. A higher 2027 package plan could support orders and backlog conversion, but Onto is already in a rapid growth cycle, so its growth cannot all be attributed to this specification change.[5]

What would confirm or disprove the chain

The first test is confirmation from Nvidia or a memory supplier that Rubin Ultra uses eight 24GB, 8-high stacks. Camtek and Onto's October order, backlog and 2027 delivery updates can then show whether back-end demand is growing faster than HBM bit forecasts. Amkor's advanced-packaging utilization outlook provides another checkpoint. A final 288GB configuration, evidence that wafers are not the bottleneck, an unchanged GPU unit plan, or a binding CoWoS constraint would disprove the chain. Full transmission may take two to five quarters.

This analysis identifies a potentially overlooked transmission chain. It is not a stock recommendation.

Sources

[1] SemiAnalysis · September 13, 2026 · “Long Live The Short King: Why 4-Hi...” · https://newsletter.semianalysis.com/p/long-live-the-short-king-why-4-hi

[2] Sina Finance · September 10, 2026 · TSMC CFO comments on CoWoS capacity and packaging outsourcing · https://finance.sina.cn/7x24/2026-09-10/detail-inirixkn8570079.d.html

[3] Drillr · July 27, 2026 · Amkor Technology FY2026 Q2 earnings call summary

[4] Drillr · August 10, 2026 · Camtek FY2026 Q2 earnings call summary

[5] Drillr · August 6, 2026 · Onto Innovation FY2026 Q2 earnings call summary