STX

Post argues the AI compute stack is increasingly memory-bandwidth constrained (“memory wall”), with 2026 shifts (inference>training, agents, growing KV cache) driving demand for high-bandwidth DRAM (HBM) and for NVMe/QLC enterprise SSDs to offload idle KV cache. Supply response is slow (2–3 years to build fabs) and near-term HBM capacity is constrained until ~2027; HBM production also “cannibalizes” standard DRAM wafer capacity.

Post argues the AI infrastructure buildout has multiple “floors” of supply-chain constraints. Author claims memory was the key bottleneck in 2025 (more than GPUs/models), cites a large gain in a memory position (“SNDK”), and asserts photonics is the next emerging chokepoint. Actionable mainly as a thematic signal (memory scarcity / photonics constraint), with limited concrete tickers beyond NVDA and the mentioned memory stock symbol.

Anecdotal social post alleging an OpenAI Codex-related bug caused runaway log-file writes that killed a Samsung 9100 PRO 4TB PCIe 5.0 SSD in ~13 months; highlights potential software-induced write amplification risk and notes a sharp price increase for the same SSD model.

Bloomberg Asia Trade highlights: (1) SK Hynix preparing a very large (~$29B) US Nasdaq listing aimed at attracting US AI investors; (2) renewed Red Sea/Yemen shipping security risk after a reported attack; (3) OPEC+ agrees to another modest output quota increase; (4) Hon Hai (Foxconn) sales beat on continued AI demand; (5) Korea begins 24-hour trading for the won (market-structure/FX liquidity angle); (6) commentary that a “Burry is right about memory chipmakers” view may be in play (memory-cycl

Lecture snippet focuses on LLM inference mechanics—especially KV-cache growth during long-context + tool-call workflows—and the resulting systems bottlenecks. Key technical signal: inference scaling is increasingly constrained by memory capacity/bandwidth and storage hierarchy (GPU HBM → CPU DRAM → SSD), not just raw GPU FLOPs. Mentions industry “rumblings” (unverified) about OpenAI buying up SSD/DRAM, and references Nvidia plus emerging inference-focused chips (e.g., Groq, which is private).

Source claims a modest PC/laptop unit-growth outlook (+1% to +2% YoY in 1H26) driven by order pull-forward and a distributor-level inventory build ahead of 2H price hikes, followed by “large production cuts.” Net implication: near-term shipment/supportive revenue recognition risk (pull-in) but increased probability of a 2H26 digestion/correction that could pressure OEMs and the PC component supply chain.

Post claims China’s YMTC is ahead of the “big 3” in NAND (400-layer soon) while China’s CXMT lags the big 3 by 3–4 years in HBM. Implication: potential NAND oversupply/price pressure risk for incumbents, while HBM leadership remains with existing top memory players (likely SK hynix/Samsung/Micron).

HOYA describes its role as a key supplier of advanced glass/photomask products across (1) EUV lithography mask blanks/photomasks for semiconductor scaling, (2) photomasks for LCD/OLED flat panel displays, and (3) glass substrates for high-capacity HDDs used in cloud/data centers. Actionable implication: positioned as a picks-and-shovels beneficiary of (a) continued EUV adoption and leading-edge wafer-fab capex, and (b) nearline/cloud HDD capacity growth; more mixed linkage to consumer display cy