NVIDIA (NVDA) Weighs 8-Layer and 12-Layer HBM4 Mix for Rubin Ultra

NVIDIA Reconsiders Rubin Ultra Memory

NVIDIA is weighing a change to the memory design of Rubin Ultra, the next-generation AI accelerator the company has slated for 2027. A report in Korean industry media that circulated on Monday says NVIDIA is evaluating a shift from an all-12-layer HBM4/HBM4E configuration toward a mixed build that pairs 8-layer and 12-layer stacks. Samsung Electronics and SK hynix, the two suppliers named in the report, have set their mass production and shipment target for 8-Hi HBM4 at the second quarter of 2027.

The logic behind the re-evaluation is physical. HBM, the memory class used alongside AI GPUs, stacks DRAM dies vertically to raise bandwidth; 8-Hi and 12-Hi count the dies in each stack. A 12-layer stack carries more memory capacity and total throughput per package, but the taller build brings two costs: more heat and lower manufacturing yield. In AI server duty, a 12-layer stack that cannot shed heat trips protective downclocking, and a part that keeps throttling can deliver less sustained compute than an 8-layer stack running at full speed continuously. Yield is the second constraint: with demand for AI chips still climbing, a low-yield 12-layer process risks becoming the bottleneck that caps Rubin Ultra shipments.

On that reading, NVIDIA would keep both stack heights in qualification, a dual-track contingency that gives up some per-package memory spec in exchange for stable output across its server fleet and supply chain. NVIDIA has positioned Rubin Ultra as the highest-end accelerator of its Rubin generation, aimed at the largest AI server clusters; Vera Rubin, the generation's first platform, is already in the supply pipeline. A leaner per-GPU memory spec could also multiply demand downstream: cloud operators chasing the same total compute would buy more GPUs and build more racks, tightening advanced packaging capacity and lifting orders for liquid-cooling and power hardware.

The report moved the memory side of the trade first. COINOTAG data shows SK hynix up 2.98% over the last 24 hours to $978.99, a print that runs straight into the $977.09 resistance level, which COINOTAG's composite scoring rates 80 out of 100. The wider setup is cautious: the daily RSI reads 35.3 and the daily trend is sideways, so the gain is pushing into overhead supply rather than breaking out. A close above $977.09, or an official configuration decision from NVIDIA, would change that picture.

The HBM4 Record Before Q2 2027

The second quarter of 2027 is not where HBM4 begins. Samsung started HBM4 mass production and commercial shipments on February 12 of this year, and NVIDIA presented Samsung, SK hynix and Micron as the HBM4 suppliers for its Vera Rubin platform at the GTC Taipei keynote on June 1. What the new report changes is the shape of the Rubin Ultra order, not the existence of an HBM4 supply chain: the open questions are which stack height the accelerator takes and when 8-layer volume reaches NVIDIA in quantity. On that point the record is thin. NVIDIA has not officially announced Rubin Ultra's final memory configuration, and the Q2 2027 date is a reported production and shipment target, not a confirmed order book or shipment record. The report likewise leaves allocation volumes, pricing and the exact split between 8-layer and 12-layer parts undisclosed.

The competitive map behind the report is three-sided. SK hynix leads on thermal control and yield, and holds the largest HBM market share with the fastest production ramp; Samsung counters by bundling its foundry and advanced packaging into a single-supplier offering; Micron has built capability in both 8-layer and 12-layer HBM4 and is actively competing for allocations. Micron's dual capability means the mixed configuration, if confirmed, does not lock any of the three out of Rubin Ultra. The HBM4 generation also redraws one layer of the supply chain: the base die at the bottom of the stack moves to a logic process, a shift that elevates TSMC, and SK hynix has already said it will have TSMC fabricate that die. For the listed trade, COINOTAG data shows NVIDIA (NVDA) last at $212.31, just above its nearest support at $210.44, scored 100 out of 100.