〔Day Digest〕Samsung Hikes Foundry Up to 15% on AI Squeeze, TSMC CoWoS Spills to Intel — 30Y Yield at Post-2007 High and Thompson's 'Railway Bubble' Reprice the Capex Stack

0. Weekly Arc

Foundry and packaging are pricing for scarcity while the cost of capital is repricing the entire AI capex stack ^[1][2][3][4]. Samsung raised wafer prices on SF4/SF5/SF8 by 10-15% in July as AI demand outran capacity, and TSMC's CoWoS lines are now so full that back-end orders are spilling to Intel's Malaysia plant — a break with the prior ecosystem norm ^[1][3]. TSMC's >70% foundry share versus Samsung's 7% in Q1 2026 sets the asymmetry ^[1]. Yet Stratechery's Ben Thompson, arguing AI is the 19th-century railway bubble, landed as the 30-year US Treasury touched a post-2007 high and the Nasdaq fell 1.33% ^[5][4]. Net: a supplier-friendly tape colliding with a capital-cost ceiling.

1. Supply: Foundry, Packaging, Memory — and the China Stack

• **[ESCALATED] Samsung foundry pricing:** 4nm SF4 and 5nm SF5 wafer prices up 10-15% for new orders in July, with ~10% on 8nm, applied to Chinese mainland and US customers ^[1][2]. BNK Investment & Securities analyst Lee Min-hee says the move could return Samsung's foundry to profit "as early as next year," earlier than the market expected, as TSMC tightness pushes customers toward Samsung and Intel ^[1].
• **[NEW] Samsung 1d nm DRAM:** internal target is R&D complete by September 2026, mass-production transfer by December 2026, line build in H1 2027, and first mass production by end-2027 ^[6]. Source: TheBell via The STAR Market Daily — single-source.
• **[NEW] TSMC CoWoS overflow:** back-end orders being routed to Intel's Malaysia facility to serve a shared mega-customer; ASE Technology Holding, Unimicron, and Gudeng flagged as beneficiaries ^[3].
• **[NEW] JCET high-aspect-ratio TSV:** completed sample builds for 2.5D/3D advanced packaging, silicon-based interconnect, and heterogeneous integration, targeting high-compute/high-memory system-level integration ^[7].
• **[NEW] Helitai + State Grid Fujian Comprehensive Energy:** five-pillar strategic cooperation covering computing-power/electricity coordination, energy efficiency, market trading, and green-power certificates ^[8].
• **[NEW] Tianjin Data + Hygon:** strategic cooperation on domestic chip-and-database co-optimization, framed as a "domestic compute and data substrate" ^[9].
• **[NEW] Shenzhen microelectronics equipment LP:** Runke Investment Management (Shanghai), Guoxin Capital (under Guosen Securities), Guangzhou State-Owned Capital Industrial Fund, Fujian Chantuo Qihang, and Huafu Growth Investment (under Huafu Securities) formed Runke Microelectronics Equipment (Shenzhen) LP with about 410M RMB in capital ^[10].

2. Demand & Application Layer

• **[NEW] Kuaishou Q2:** revenue 35.5B RMB (+1.4% YoY); core commerce +7.4% YoY; adjusted net profit 3.9B RMB; DAU 412M and MAU 797M, both record highs ^[11]. Kling AI alone generated over 850M RMB in Q2, up over 200% YoY, with the 3.0 series adding native 4K video output and an MCP/CLI for agent batch creation ^[12].
• **[NEW] ECB President Christine Lagarde (Geneva, World Economic Forum HQ):** "Europe cannot miss the AI revolution"; survey data shows eurozone firms will invest roughly 9% of total investment in AI this year ^[13].
• **[NEW] Top 15 AI tools by site visits, June 2026 (AITools.xyz, web-only):** ChatGPT 5.3B visits (more than the other 14 combined), Google Gemini 1.1B, Anthropic Claude 968M ^[14].
• **[NEW] Optoelectronic chips:** CAE foreign academician Chang Ruihua (CUHK Shenzhen), a VCSEL specialist, answered "yes, certainly" when asked whether a "Nvidia of optical chips" will emerge, arguing AI compute competition is shifting from single-GPU performance to cluster interconnect ^[15].
• **[NEW] Zhipu founder Tang Jie:** argued trillion-parameter runs were an "industry detour"; GLM-5.3, sharing GLM-5.2's 753B total and 40B active parameters plus one month of long-horizon RL, scored 60 on the AA Evaluation Index vs GLM-5.2's 53, a 7-point gain ^[16].
• **[NEW] Ant Group** established Shanghai Ant Zhibao Technology, fully owned, with scope covering AI general-application systems and AI application software ^[17].
• **[NEW] Luming Robot** launched Lumos NexCore, integrating task definition, data ingestion, training, evaluation, deployment, and continual learning; compresses robot-skill development from "weeks to days" ^[18].
• **[NEW] ChatGPT Ads** expanding across Europe ^[19].

3. OpenAI and the Model-Layer Worry

• **[NEW] OpenAI's model development is slowing** on cybersecurity concerns, per Euronews ^[20].
• **[NEW] "OpenAI trails Anthropic as losses deepen,"** with Sam Altman pausing frontier training per CoinDesk — single-source, thin ^[21].
• **[NEW] Ed Zitron (24/7 Wall St.):** "What Happens If OpenAI Dies?" — argues OpenAI is running out of time ^[22]. All three OpenAI items overlap; only the cybersecurity-driven slowdown names an outlet, the rest are headline relays — flag as thin.

4. Structural Bear Case: Memory, Nvidia, and the "Railway Bubble"

• **[ESCALATED] Cathie Wood (ARK Invest):** does not own memory stocks; calls memory the most commoditized, most cyclical part of the semiconductor chain, and says today's high prices are a negative rather than positive signal ^[23].
• **[NEW] Ben Thompson (Stratechery) on Nvidia:** Nvidia's "hidden door" is a circular-financing arrangement in which it funds customer purchases and absorbs bad-debt risk — a de facto price discount invisible in the headline margin; if compute demand rolls over, the off-balance-sheet exposure returns to Nvidia's books ^[24].
• **[NEW] Ben Thompson on the AI cycle:** the AI infrastructure build is the 19th-century railway bubble — the technology survives, but capital timing is misaligned, with cash running from free cash flow to debt to equity (Google is now issuing equity, Nvidia is assembling a $500B project to tap pension and insurance float) ^[5]. Same analyst, same outlet, two pieces; treat as a single voice.
• **[NEW] Policy nudge:** Xinhua-affiliated coverage cites experts urging broader "patient capital" channels for compute-asset development ^[25].

5. Overnight Tape and What Would Falsify the Squeeze

• **[ESCALATED] US AI rout (overnight):** Nasdaq -1.33%, Philadelphia Semiconductor Index nearly -5%; memory, optical-communication, and AI-compute names broadly hit ^[4].
• **[ESCALATED] Asia follow-through:** Samsung Electronics and SK Hynix down over 6%; 2x-long Samsung and 2x-long SK Hynix leveraged ETFs both fell over 14%; A-share AI compute, optical modules, and memory blocks fell over 5%, with several leaders off more than 8% ^[4].
• **[ESCALATED] Rates and oil:** 30-year US Treasury yield touched the highest since 2007; 10-year also moved notably higher; oil hit a three-week high after the US paused Iran talks and the UAE suspended all trade and financial ties with Iran ^[4].
• Falsification test for the foundry-tightening thesis: if TSMC Q4 utilization slips or CoWoS order intake decelerates, Samsung's 10-15% hike stops looking like scarcity and starts looking like market-share defense ^[1][3]. Falsification test for the bubble thesis: any quarter in which hyperscaler free cash flow turns positive without further equity or project-finance issuance collapses Thompson's timing argument ^[5]. Source-quality flag: the "OpenAI trails Anthropic" framing ^[21] and the 1d nm DRAM timeline ^[6] are single-source and unverified.