To summarize GFHK's note: Rubin's SOCAMM2 has been nerfed once again — down to 64GB. But is that bearish? If anything, it's the opposite. Working off a 192GB SOCAMM2 configuration, NVIDIA had asked the memory makers for a corresponding volume of SOCAMM2 — and the memory makers came back telling them they could only supply 60–70% of that volume. This isn't a demand problem, it's a supply problem. The spec was cut because they couldn't get the SOCAMM2 supply. What about HBM? Contrary to what a certain research house published, Rubin Ultra is not being considered with 8Hi only. An HBM4E 12Hi SKU is also under consideration. It is a step down from the original 16Hi — but everyone already knew that, right? NVIDIA needs 12Hi if only to compete with AMD. The 8Hi is there to accommodate SKU requests from a subset of customers.

Behind AI’s Rapid Growth Lies Memory. SK hynix showcased #HBM4# and #SOCAMM2# on @NVIDIA’s Vera Rubin at TSMC OIP 2026 and received TSMC’s Partner of the Year Award for the second consecutive year. 🤝 🔗 Read more: #SKhynix# #TSMCOIP#

Apple may have no choice but to buy older, bigger footprint LPDDR from China. Here's why. In 2027 AMD, Qualcomm & NVIDIA are now bidding for the same mobile memory that goes into your iPhone. $AMD EPYC Verano server CPUs adopt LPDDR5X on SOCAMM2 (2027). $QCOM AI200/AI250 carry 768GB of LPDDR per card (HBC) $NVDA Vera CPUs attach massive LPDDR5X via SOCAMM2 in Vera Rubin racks. "One of the most promising developments is Low Power Double Data Rate 5X (LPDDR5X) memory deployed using the recently JEDEC approved Small Outline Compression Attached Memory Module (SOCAMM2) form factor, which promises to combine the proven power efficiency of LPDDR5 mobile memory with the modularity and serviceability required in server environments." -An Excerpt from AMD Article

At #DTF2026# in Seoul, @SKhynix showcased its full-stack AI memory portfolio, from #HBM4# and #SOCAMM2# to server DRAM and SSD solutions, reaffirming its collaboration with Dell Technologies. 🔗 Read more below. #SKhynix# #AIMemory# #AIInfrastructure#

RAMBUS $RMBS Q3’26 EARNINGS HIGHLIGHTS 🔹 Revenue: $218.3M (Est. $213.2M) 🟢; +22% YoY 🔹 Adj. EPS: $0.84 (Est. $0.79) 🟢; vs $0.62 LY 🔹 Product Revenue: $114.3M; +22% YoY, +15% QoQ 🔹 Cash From Operations: $102.5M (Est. $96.7M) 🟢 Q4 Guide: 🔹 Revenue: $227M-$233M (Est. $227M) 🟡 🔹 Adj. EPS: $0.84-$0.92 (Est. $0.84) 🟡 🔹 Product Revenue: $127M-$133M 🔹 Royalties Revenue: $68M-$74M 🔹 Contract & Other: $26M-$32M 🔹 GAAP EPS: $0.68-$0.76 Revenue Mix (Q3): 🔹 Royalties: $77.0M 🔹 Contract & Other: $27.0M Announcements: 🔹 Expanded product and IP offerings for next-generation AI systems 🔹 Industry-leading PMIC for LPDDR6 SOCAMM2 Other Q3 Metrics: 🔹 Non-GAAP Oper Income: $102.2M (Est. $96.2M) 🟢 🔹 Non-GAAP Net Income: $91.7M (Est. $86.7M) 🟢 🔹 Non-GAAP Operating Margin: 47%; vs 44% LY 🔹 GAAP EPS: $0.69; vs $0.44 LY 🔹 Cash & Marketable Securities: $776.9M Key Updates: 🔹 Repurchases of common stock: $135.4M Comments: 🔸 “We continue to expand our product and IP offerings to address the growing performance requirements of next-generation AI and data center infrastructure.”

Korean media report that TLB, a Korean manufacturer specializing in memory module substrates, has begun developing SOCAMM3 substrates for NVIDIA’s Feynman platform. Industry sources estimate TLB’s SOCAMM substrate revenue at around KRW 70 billion this year, with expectations that it could double to roughly KRW 140 billion next year. SOCAMM2 substrates already take twice as long to manufacture and sell for twice the price of conventional memory module substrates. SOCAMM3 substrates are expected to command even higher prices than SOCAMM2.

Every AI agent needs memory to remember what it's doing, what it's learned and what comes next.​ Scale that to thousands of agents running simultaneously, and memory capacity can become a key factor in AI inference performance, especially for long-context workloads. That's why we're advancing solutions like our SOCAMM2 modules to create the memory foundation for the next generation of AI. Read more →

DIGITIMES: Memory boom stays strong into 2027 DIGITIMES argues the memory upcycle is accelerating as AI infrastructure drives demand for DRAM, NAND and HBM, with supply remaining tight until meaningful new capacity starts arriving in 2027–2028 The biggest structural change is that memory demand is increasingly tied to AI accelerator architectures rather than PCs and smartphones. HBM4 adoption accelerates through 2026–27, followed by HBM4E, while Nvidia’s Vera platform moves toward removable SOCAMM2 LPDDR5X modules Enterprise storage is also benefiting. PCIe Gen6 enterprise SSDs are entering production, with Samsung and Micron already ramping products and Phison and Silicon Motion expected to launch Gen6 controllers in 2027 Capacity additions are substantial but mostly back-end loaded: · Samsung: Pyeongtaek P5 production in 2028, Yongin DRAM in 2029, new Onyang HBM packaging in 2029. · $SKHY: M15X already ramping in 2026, Yongin DRAM from Q1 2027, P&T7 HBM packaging by end-2027, Indiana HBM packaging in H2 2028. · $MU: Tungluo expansion, Hiroshima DRAM, Singapore HBM packaging from 2027, new NAND capacity in H2 2028, Idaho expansion in 2027–28 and New York DRAM from 2030.

Micron is better positioned than both Samsung and SK Hynix in building out the memory-compute platform. - SK Hynix has the best current HBM execution and the tightest Nvidia relationship but it is geographically concentrated, not full-stack, and more exposed to any AI capex slowdown. - Samsung has the most raw resources and broadest technology portfolio but has lost HBM share in the past, has inconsistent execution, and its conglomerate structure creates strategic diffusion. - Micron has the most durable structural positioning for a 3-5 year platform transition: the only full-stack U.S. memory supplier, with SOCAMM2 competition rights, hyperscaler co-design relationships, broadest IP portfolio and the geopolitical moat that neither Korean player can replicate. $MU $DRAM

Semiconductor Supply Chain News Roundup, September 14 1. Samsung Electro-Mechanics submitted TGV module samples for its glass substrate to TSMC for reliability verification two to three months ago, but the evaluation schedule has reportedly been delayed by more than a month. As a result, the ordering schedule for some TGV related equipment that Samsung Electro-Mechanics had been preparing with its suppliers is understood to have been pushed back from the original plan. Samsung Electro-Mechanics had originally targeted glass substrate mass production around 2027, building a pilot line and running customer evaluations. Earlier this year it revised the target for full operation to 2028, but with reliability verification and mass production process stabilization taking longer than expected, equipment investment is also not proceeding as originally planned. 2. A senior executive at Samsung Electro-Mechanics recently conveyed to employees a policy of moderating the pace of investment in the glass substrate business for the time being. Rather than concentrating large scale investment on glass substrates, the direction is to prioritize investment in existing package substrate businesses such as FC-BGA and in high value added businesses with confirmed demand, such as MLCCs for AI and automotive applications. Looking at the current state of Samsung Electro-Mechanics' glass substrate work, yields in the initial process of forming holes in the glass through laser processing and etching have risen to 90% or above, but in the subsequent processes of metallizing and filling the holes with copper, yields reportedly drop to around 30%. After that, additional steps are still required, including chemical mechanical polishing (CMP), which flattens the substrate surface using chemical reactions and mechanical force, as well as cutting, and by that point yields become far too low. That said, some argue that this adjustment itself should not be read as a delay in the glass substrate business as such. The market has not yet opened in earnest, and the actual customers, global semiconductor companies, are still carrying out their own preparation and validation for glass substrate adoption. 3. SK hynix is moving to lock in its equipment supply chain. The approach is to share medium to long term demand forecasts earlier than before so that suppliers can secure in advance the equipment and key components needed for future investment. Previously, SK hynix shared investment plans with equipment makers covering about six months, or at most about a year. Going forward, it plans to present demand not only for next year but also for the year after, allowing suppliers to plan their production capacity and component procurement ahead of time. 4. SEMES, Samsung Electronics' semiconductor equipment subsidiary, is firming up the component supply chain for the hybrid bonder to be applied to next generation high bandwidth memory (HBM). Specifically, SEMES has signed a supply agreement with the Korean subsidiary of HIWIN, a Taiwanese motion control company, for the Single Axis Stage, a precision drive component to be used in the hybrid bonder, and is currently conducting tests. HIWIN's product is being validated against SEMES' hybrid bonder specifications, and testing is understood to be in its final stages. The single axis stage is a drive component that supports precise position control in the hybrid bonder. Because a hybrid bonder must align and bond chips at extremely fine pitches, it requires a high level of positional accuracy. SEMES is said to be conducting equipment testing after signing a supply contract with a customer for a D2W (Die to Wafer) hybrid bonder for HBM. The industry views that customer as its parent company, Samsung Electronics. 5. PCB orders for SOCAMM2 destined for major memory makers including Samsung Electronics and SK hynix are growing beyond the levels planned at the start of the year. The increase in orders for DDR5 RDIMM PCBs is especially pronounced. Compared with the order plans projected at the start of the year, the upward revision for RDIMM is understood to be even larger than that for SOCAMM2. An industry source cited by the press said, "For SOCAMM2, order volumes are growing beyond what we expected at the start of the year, and we expect them to increase further into the second half," adding, "RDIMM orders are rising even more steeply relative to the beginning of year plan than SOCAMM2." The source also noted, "Some customers have already begun discussing advance orders for next year's volumes."