--- type: "Topics" locale: "en" url: "https://longbridge.com/en/topics/43151501.md" description: "Rumors: Let the bullets fly for a while$Micron Tech(MU.US)$SK Hynix(SKHY.US)I. Core Facts that are 100% True ✅1. The 4-chiplet Rubin Ultra solution has indeed been completely canceled.SemiAnalysis obtained internal engineering documents from major clients, confirming that NVIDIA's initially planned top-tier configuration for GTC 2026—4 Die + 16-Hi HBM4E—has been permanently scrapped due to yield issues with TSMC's CoWoS-L advanced packaging warpage. The mass-production version will uniformly adopt a 2-Die dual-chiplet architecture, which is a consensus across the entire industry supply chain.2. Only CPU-side SOC AMM standard memory was cut; GPU HBM remains untouched.The biggest truth distinction on the internet: The single-rack CPU companion memory was slashed from 55TB to 28TB (capacity halved), but the total capacity of the 20.7TB HBM built into the 72 GPUs remains unchanged. Jensen Huang specifically debunked rumors in a Korean interview, stating that NVIDIA will not reduce HBM usage on the GPU side.3. The 1800W Max-Q mainstream version + 2600-2800W Max-P flagship tiering is the official established roadmap.Max-Q is the main production model for the vast majority of cloud providers, while the ultra-high-power Max-P version is exclusively supplied to top-tier customers like Microsoft and Google and is not publicly sold. This power tiering plan fully aligns with NVIDIA's public technical documentation.4. The core purpose of the cuts: To ship more GPUs using limited HBM capacity.Global HBM yield and capacity represent a hard ceiling. By compressing non-core CPU memory costs and abandoning extreme per-card parameters, NVIDIA aims to increase overall GPU shipment volumes. This supply chain strategy is an explicit adjustment direction within NVIDIA.II. Parts with Serious Misleading Information or Half-Truths ⚠️1. Incorrect Statement: “Rubin Ultra HBM cut from 288GB to 192GB”This interpretation only exists in retail investor rumors. Officially, the 12-Hi 288GB HBM version has not been canceled: the 288GB high-spec HBM is retained only for niche Max-P flagship SKUs, merely ceasing to be widely available across the market. The 192GB configuration is for the general circulation version and does not represent a performance downgrade.2. Rehashing Old News with Exaggerated BearishnessRumors about the cancellation of the 4-chiplet solution had already circulated in the supply chain as early as March 2026. SemiAnalysis's re-publication in July sensationalizing "chip shrinkage" constitutes repetitive 炒作 of old information, deliberately amplifying a pessimistic narrative.3. Performance Has Not Been HalvedThe dual-chiplet version achieves minimal performance gap compared to the original 4-chiplet solution through architectural optimization and NVL576 super-cluster interconnect design. The claim in the article that "performance is only 1/4" is an extremely one-sided statement.III. Completely False Self-Media Rumors ❌1. Rumor: Overall HBM demand has decreased.Although the stacking layers of HBM in single-card circulation versions have decreased, the significant surge in GPU shipment volumes ensures that the total bit demand for HBM across the industry continues to grow at over 50% annually. Long-term locked-price contracts for SK Hynix and Micron have seen no downward adjustments.2. Rumor: NVIDIA's technology has regressed.Specification adjustments are proactive system optimizations under the triple constraints of data center power, packaging yields, and material costs, not a lack of technical capability. The industry's competitive focus has shifted from single-card VRAM comparisons to multi-card cluster interconnect support.3. Rumor: The logic of the storage trio has broken down.Only the demand for low-end general-purpose DRAM (SOC AMM) has been compressed. The scarcity barrier and premium pricing structure of high-end HBM remain completely unchanged." datetime: "2026-08-03T09:27:43.000Z" locales: - [en](https://longbridge.com/en/topics/43151501.md) - [zh-CN](https://longbridge.com/zh-CN/topics/43151501.md) - [zh-HK](https://longbridge.com/zh-HK/topics/43151501.md) author: "[嘉禾暖光学习财商](https://longbridge.com/en/profiles/17154496.md)" --- # Rumors: Let the bullets fly for a while$Micron Tec… ### Related Stocks - [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md) - [SKHY.US](https://longbridge.com/en/quote/SKHY.US.md) - [SKHYV.US](https://longbridge.com/en/quote/SKHYV.US.md) - [MU.US](https://longbridge.com/en/quote/MU.US.md) - [TSM.US](https://longbridge.com/en/quote/TSM.US.md) - [MSFT.US](https://longbridge.com/en/quote/MSFT.US.md) - [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md) - [NVD.DE](https://longbridge.com/en/quote/NVD.DE.md) - [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md) ## Comments (5) - **深蹲达人 · 2026-08-03T11:59:43.000Z · 👍 1**: It was a cyclical stock anyway, SK Hynix has already halved, if I had seen through it earlier, it wouldn't have been this miserable 😂 - **嘉禾暖光学习财商** (2026-08-03T12:07:41.000Z): It's a cycle, but who knows when the cycle will end 😂😂 - **杰聪 · 2026-08-03T09:49:43.000Z**: Don't understand - **吃练睡 · 2026-08-03T09:40:40.000Z**: It feels like this is also bearish for the CPU. - **嘉禾暖光学习财商 · 2026-08-03T09:28:10.000Z · 👍 3**: The underlying industry facts in this research report are accurate, but the title and conclusions present a deliberately one-sided bearish narrative:No need to panic and liquidate positions in the HBM 'Big Three' (SK Hynix, Micron, Samsung);The only true negatives are low-end general-purpose DRAM an