GG

[GG @Jiwai_]

Morgan Stanley: Humanoid Semis Demand

Strategic Role: Semiconductors are foundational to physical AI. As robots become more intelligent, semi content per robot increases.

Market Projections (OEM only):
2025: $2.2B
2030: $24B
2040: $570B
2050: $1.8T (~800x growth from current levels)

Key Components:
Compute & Perception: CPUs, GPUs, memory, connectivity for motion planning and real-time processing.
Specialized: Vision processors, sensor interfaces (LiDAR, cameras, tactile), motor control ICs, power/analog chips.

Trends: New robotics-specific silicon wave. NVIDIA’s “3 Computer” approach (datacenter training, simulation, on-robot inference). Humanoid semi content ~$1.5k–$3k per unit.

Explosive long-term opportunity for semiconductors in robotics/physical AI.
$MCHP $SONY $AMD $NVDA $QCOM $ON $AMBA $STM $ALGM $SNPS $CDNS $TXN $ARM

[GG @Jiwai_]

Goldman Sachs on High-End CCL:

High-end CCL has become one of the fastest-growing segments in the AI hardware stack, with demand significantly outpacing supply additions.

NVIDIA VR200: Expected to require 3.5x more CCL dollar content than GB300, driven by higher-grade M8+ materials and additional midplane boards.

Trend extends to AI ASICs and high-speed switches, steadily increasing CCL content per system.

Strong structural tailwinds for CCL in AI infrastructure.

[GG @Jiwai_]

高盛大幅上调中际旭创（Innolight，https://t.co/LROnSY6mEJ）目标价至2581元：

硅光子（SiPh）光模块出货增长（AI服务器机柜扩容驱动）。

市场扩张至 scale-up、scale-across 和 scale-out。
产品组合升级带来更好盈利能力。

盈利调整：2026-2028年盈利预估上调 65%/108%/119%。2026-2028营收/��利CAGR预计 64%/69%。

定位：全球最大光互联解决方案供应商，提供高速、低功耗、低延迟方案，服务云端与AI基础设施。

主要风险：800G+需求不及预期、新型器件量产慢、市占率正常化、地缘政治、供应链问题。

[GG @Jiwai_]

Morgan Stanley: Optical vs Copper in AI

Scale-Out: Small adoption (10–15%) in 2027; CPO low single-digit in 2028, scaling meaningfully in 2029 (Feynman era).

TAM Drivers: Capex 2x higher than expected last year + exploding cluster sizes (8 → 72 → hundreds → thousands) as multipliers.

Supplier Exposure:
Lumentum: Most positive (Google transceiver exposure, less CPO risk).
Coherent: Balances transceiver business with CPO opportunity.
Corning: Less direct impact from CPO; GlassBridge still early-stage.

Copper Position: Strong runway through 2026–2027 (retimers, better SerDes). Hybrid environment expected before full optics shift. Ethernet (Broadcom) and UALink (Marvell, Astera Labs) key debates.

Keysight: Agnostic winner — benefits from multiple architectures, higher speeds, and frequent changes requiring advanced testing.

Copper dominant short-term; optical scale-out ramps later. CPO more relevant in scale-up.

$ALAB $NVDA $AVGO $MRVL $CRDO $SMTC $APH $KEYS $COHR $LITE $GLW $GOOGL

[GG @Jiwai_]

BofA: Kimi K3 Release & AI Compute Race

Kimi K3: Moonshot’s new 2.8-trillion-parameter MoE model with 1M token context window. Positions China closer to U.S. frontier labs.

Pressure on U.S. Labs: Google Gemini 3.5 Pro reportedly delayed; OpenAI, Anthropic, Google need larger training runs, more RL, synthetic data, and faster cadences.

MoE Impact: Boosts demand for memory movement, routing, latency, and interconnects.

NVIDIA Advantage: GB300 NVL72 offers up to 25x performance-per-watt improvement on leading open MoE models.

Bullish for Semis: Open models drive infrastructure spend on GPUs, HBM, networking, and efficient inference. EDA (Cadence, Synopsys) remains critical despite open-source mentions.

Token Usage: Chinese labs leading weekly token volume; U.S. enterprise AI adoption at 55% (Anthropic 42.4%, OpenAI 39.5%).

Kimi K3 accelerates the compute race and remains structurally positive for AI semiconductors.

$CDNS $NVDA $GOOGL $SNPS

[GG @Jiwai_]

Citi: EML Output Trends

Industry Shift: Rapid transition from 100G EML to 200G EML and CW (Continuous Wave) by 2028 across listed companies.

AXT, Inc.: 200G EML output surges from 13.54M chips (2025) to 157.82M chips (2028).

Chinese Players: Strong expansion in CW output. Source Photonics (Dongshan Precision) from 14.66M (2025) to 161.77M (2028).

Sumitomo Electric: Maintains high production, especially 200G EML from 28.93M (2025) to 128.58M (2028).

Accelerated upgrade to higher-speed lasers driven by AI optical module demand.

[GG @Jiwai_]

Goldman Sachs: Innolight & Optical Modules

Speed Migration: 1.6T moving to volume production; 3.2T ramp in 2027. 1.6T expected to be mainstream by 2027 (45.72M units).

SiPh Adoption: Penetration 60% (800G), 80% (1.6T), 100% (3.2T) by 2026. SiPh revenue share from 28% (2025) to 62% (2028E).

Innolight: Leading 1.6T pluggable supplier, projected 50–60% (or up to 70%) market share in 2026. Capex +155% YoY in 2026 for capacity.

ASP Trends: 1.6T ASP from $840 (2025) to $435 (2028E); 3.2T starts at $1,023 (2027E) → $768 (2028E).

TAM: Global optical module market from $34.2B (2025) to peak $72.6B (2027E). High-speed (1.6T+3.2T) revenue from $3.9B (2025) to $44.9B (2028E).

Rapid shift to higher-speed SiPh-based modules driven by AI demand.

[GG @Jiwai_]

Morgan Stanley: Power Semiconductors

$/kW Trend: Power semi content per kilowatt rising sharply across generations.

Key Progression:
B200: $112/kW
GB200/GB300: $118/kW
Rubin: $128/kW
Rubin Ultra: $170/kW
Feynman: $191/kW (+$79/kW from Rubin)

Rubin Breakdown (Power Semi Allocation):
PSU: 30%
VRM (lateral) - 2nd stage: 28%
IBC - 1st stage: 22%
BBU/UPS: 10%
NICs: 5%
eFuses/etc.: 4%

Significant increase in power semiconductor content as AI architectures scale, driven by higher efficiency and power delivery needs. ($TXN $STM $ON $DIOD $ADI $NXPI)

[GG @Jiwai_]

Morgan Stanley: CXL Market Update

Raised long-term CXL market forecasts due to faster industry adoption and elevated memory shortages.

2030E Estimates:
CXL MXC chips: US$2.1B (up from previous $990M)
CXL switch chips: US$1.9B (up from previous $664M)

Significant upward revision reflecting stronger demand for CXL solutions to address memory bandwidth and capacity challenges in AI systems. ($PENG $MRVL $ALAB)

[GG @Jiwai_]

Morgan Stanley: ZAM Memory

Cloud Memory Spending: Projected to reach $418B by 2030 (8% CAGR from 2026). Share of total cloud capex rises to 40% in 2027 (from 12% in 2023 pre-AI).

Bandwidth Lag: Memory bandwidth improvements (+14% from DDR5-5600 to DDR5-6400) lag far behind token growth (320x on major clouds from Apr 2024–Jun 2026).

Spending Peak: Absolute cloud memory spend peaks in 2028E at ~$520B, then declines as % of total capex to 25% by 2030E (non-memory capex hits $1.28T).

DRAM S/D: Deep structural shortage persists through 2028E (deficit -10% to -16% in 2026E).

NAND S/D: Moves from surplus to tight (deficit -9% in 1Q26E, moderating to -3% to -4% later). Transitioning to enterprise storage/LLM KV caches.

AI driving sustained memory tightness and high cloud spend share through late 2020s. ($DRAM $MU $SNDK $WDC $EWY)

[GG @Jiwai_]

BofA: 3Q Memory Contract Price Check

Server DRAM: +20–30% QoQ (stronger than consensus ~20% or less), led by high-speed LPDDR5.

Spot Demand: Healthy, driven by commodity DDR5 and legacy DDR4; July prices rising MoM.

HBM4 Transition: Shifting orders toward higher-priced HBM4 from HBM3e.

LTAs: Represent <50% of DRAM sales (non-LTA 60–70%), allowing solid QoQ increases (5–10%), often settled even under existing contracts.

Rush Orders: OEMs rushing to buy ahead of peak 4Q season, confirming >20% QoQ increase for commodity DRAM and NAND.

ASP Outlook: BofA forecasts +21% QoQ for 3Q DRAM ASP (more optimistic than TrendForce).

Spot Market: DRAM spot prices up for 8 consecutive weeks; NAND rebounding (1Tb wafers +4% WoW).

Strong pricing momentum continuing into 3Q, defying earlier skepticism. ($DRAM $MU $SNDK $WDC $EWY)

[GG @Jiwai_]

UBS: Materials Sector (H2 2026 Price Hike Visibility)

Ranking (Highest to Lowest):
MLCC Materials > MLCCs > WF6 > Semiconductor Wet Chemicals > Helium

MLCCs:
High-capacitance MLCC supply remains tight due to AI server demand crowding out capacity.
Yageo raised prices from July 1 across automotive, industrial, AI server, and consumer segments.
Upstream materials (e.g., release films) at high utilization → further upside.

Electronic Specialty Gases:
WF6 (5N-grade): Prices doubled YTD to RMB1,800/kg due to tungsten powder export decline (shortage for Japan/Korea).
Helium: Peaked in April, corrected to RMB250/m³; expected to stabilize at RMB200–300/m³ in H2.

Electronic Wet Chemicals & Targets:
EG-HF: Rising raw material costs (sulfur, anhydrous HF) + strong exports → continued price hikes.
Sulfuric Acid, H2O2: Upward trend since March.
Targets (Molybdenum, ITO): Indium (+70% YTD), Molybdenum (+30% YTD) raw material costs support price increases.

Focus shifting to materials with strong pricing power post-correction, led by MLCC-related segment.

[GG @Jiwai_]

Moonshot AI is about to release Kimi K3, a massive large language model with 2–3 trillion parameters — China’s largest AI model to date.

Expected to surpass Anthropic’s Opus 4.8 on mainstream benchmarks and will be released as an open-weight model available for free download.

The launch could challenge the industry consensus that Chinese AI models lag U.S. frontier labs by 8–12 months.

Moonshot is raising a new funding round at a valuation of ~$31.5 billion.

K3 is scheduled for release in the coming days.

[GG @Jiwai_]

SEMI: Global Semiconductor Equipment Sales Forecast
2026: Record $165.9B (+23.2% YoY).
Long-Term: Momentum through 2028 to $229.5B.
AI infrastructure demand — leading-edge logic, HBM, advanced packaging, and complex architectures.

Segment Breakdown:
Wafer Fab Equipment (WFE): +23.1% to $143.9B in 2026; +21.8% (2027), +14.1% (2028) → surpasses $200B.
Test Equipment: +31.0% to $15.3B in 2026 (after +55.3% in 2025) → $20.8B by 2028.
Assembly & Packaging: +9.6% to $6.7B in 2026 → $8.6B by 2028 (driven by heterogeneous packaging).

Geographic Trends:
China, Taiwan, and Korea remain top 3 through 2028.
China #1 but growth moderating after heavy investment.
Taiwan & Korea focused on leading-edge foundry and HBM.
Other regions accelerating in 2027–2028 due to regionalization and incentives.

Historic AI-driven equipment boom continuing through 2028.

[GG @Jiwai_]

Morgan Stanley: HBM & NAND Supply/Demand

HBM:
Market: From $3B (2023) to $94B (2027E), 128% CAGR.

NVIDIA GPGPU share: 90%+ through 2027E.

HBM intensity: GPGPU from 80GB (2023) to 317GB (2027E); ASIC from 40GB to 238GB.

2027E Production: SK Hynix 22,226M Gb, Samsung 21,239M Gb, Micron 10,372M Gb.

DRAM: Severe squeeze — sufficiency ratio -17%
(2026E) / -15% (2027E).

AI NAND:
Demand: From 205 EB (2025) to 609 EB (2027E).

AI share of total NAND: 18% (2025) → 41% (2027E).

Key drivers: Google TPU, AWS Trainium, NVIDIA/AMD platforms.
NAND Overall: Surplus +2% (2025) → deficit -15% (2026E) / -9% (2027E). Non-AI demand flat.

AI driving extreme tightness in both HBM and NAND, with massive growth ahead. ($MU $DRAM $EWY $SNDK)

[GG @Jiwai_]

BofA: Optical FAUs (CPO Market Outlook)

Nvidia CPO Switch Shipments:
2026E: 14k units
2027E: 52k units
2028E: 190k units (meaningful volume takeoff)

TAM: FAUs in Nvidia CPO switches projected to reach US$8B by 2030E (based on 1.4M CPO switch units). FAU contributes 5–10% of CPO switch BoM, ASP $80–$200.

Technical Barriers:
Micron-level alignment (< ±0.5μm for NPO, < ±0.3μm for CPO) critical to avoid signal loss.
Automation essential as channel counts rise to 60–100.

GlassBridge (Corning): Early stage, unlikely for current/next-gen CPO due to edge coupling limitations, alignment issues, and wafer-scale testing challenges.

CPO/FAU ramp accelerating with Nvidia, but tight technical requirements create high barriers. ($GLW $MRVL $AVGO $HIMX $LITE $COHR $FN $NVDA $TSMC)

[GG @Jiwai_]

Morgan Stanley: AI Infrastructure Meets Humanoids

Humanoids Semiconductor Opportunity: Could reach $305B in semi sales by 2045 (~24% of BoM from semis, up from 5% today). Potential 1B units/year by 2045. Humanoids could represent 15% of semi spend by 2040.

Top European Picks:
Infineon & STMicroelectronics: Supply microcontrollers.
STMicro: AI vision, Time-of-Flight sensors, Mobileye experience.
Infineon: Inverters, onboard chargers, battery management. BoM per robot ~$500–$700+.

AI Infrastructure (Power):
Shift to 800VDC, GaN/SiC ~35% of server rack BoM.
Infineon highlighted for power efficiency, roadmap, and scale (300mm thin wafer). Strong in 400V/480V (Oberon for NVIDIA), TPUs, Trainium, ASICs. Power semi sales expected to grow from ~$1.5B to $3–4B by FY28.

Long-term bullish on semiconductors in humanoids and power delivery for AI infrastructure. ($STM $IFNNY)

[GG @Jiwai_]

长鑫科技CXMT：

发行价：8.66元/股，募资约 579亿（绿鞋全额 ~666亿），发行后总股本约 668亿股。

预计市值：发行价对应 ~5792亿；盘前 6U 计算对应 ~2.88万亿（约5倍发行价）。

申购日期：2026年7月16日；预计上市 7月27日。

业务：国内最大DRAM厂商（IDM模式），全球市占率 ~7.67%（全球第4）。主要产品：标准DRAM、移动DRAM + HBM 等下一代技术。

流通盘：上市初期实际流通股仅约 2.2%（低流通meme属性强）。

业绩：2025年首次盈利（营收618亿，净利19亿）。2026上半年预计营收1100-1200亿，净利500-570亿（DRAM涨价+产能释放驱动）。

估值：发行价对应2025年PE ~309倍；若2026上半年利润年化��极低（5-6倍）。

国产DRAM龙头 + AI存储逻辑 + 超低流通盘，短期情绪驱动强，中长期看DRAM周期和技术差距。外资定价已较高，A股开盘需关注情绪。

[GG @Jiwai_]

Samsung Electronics is considering outsourcing the back-end design of Google’s 2nm TPU I/O Die due to surging advanced process orders straining internal resources.

Google’s 10th-generation TPU (codename Icefish) is co-designed with MediaTek, with mass production planned for 2028. The compute processor will be made by TSMC on 1.4nm, while Samsung handles the 2nm I/O Die (responsible for HBM data transfer).

Potential partners: ADTechnology, Gaonchips (less enthusiastic, prefer full ASIC projects), and Alphachips (more eager).

Context: Samsung recently secured 2nm customers including Anthropic and DeepX, in addition to Google and Tesla. Back-end design contracts are lower value (tens of billions KRW) compared to full ASIC projects (hundreds of billions to trillions KRW).