MS: Copper Foil Growth Engines → HVLP4 copper foil supply shortage: Morgan Stanley forecasts HVLP4 demand rising 218% YoY in 2027 to 3,270 tonnes/month, exceeding capacity by 31%. NVIDIA Rubin, AWS Trainium3/4 and Google TPU v8 are driving adoption, while 18–24-month qualification cycles constrain supply. HVLP4 processing fees command >7× conventional foil. → Chinese suppliers challenging Japanese dominance: Mitsui Kinzoku is projected to retain 32.4% of global HVLP4 capacity in 2027, followed by China's Defu at 19.6%. Defu has achieved ~62% HVLP4 yields and shipped 190 tonnes in September 2026, with a 50ktpa PCB foil expansion supporting further market-share gains.

MS: HDDs vs Memory Cycle → “HDDs appear to be at the point in the upcycle where memory was a year ago.” → “Demand is accelerating, customers are broadening, pricing is entering a steeper part of the S-curve.” → While memory price growth is expected to decelerate and margins flatten, MS sees HDD pricing power strengthening, with more meaningful margin expansion and earnings upside still ahead

Nomura: DeepSeek Expert Call → DeepSeek V5: Next-generation flagship model reportedly targeting 1.5–2× the parameters of V4 Pro, with improvements in complex reasoning and long-horizon agent tasks. → Dual-model strategy: V4 Pro pushes frontier intelligence, while V4.1 Flash drives the majority of API revenue through lower-cost inference. → Pricing: DeepSeek offers a 50% discount during off-peak hours to improve compute utilisation, maintain competitive pricing and support margins. → NVIDIA dominance: DeepSeek primarily relies on NVIDIA GPUs for large-scale pre-training. Domestic accelerators are mainly used for adaptation testing and limited-scale training. → Chinese AI chips: Domestic accelerators face significant challenges in interconnect bandwidth, software compatibility and cluster stability. Substitution is expected to happen faster in inference than training. → Inference optimisation: DeepSeek's MLA architecture reduces KV-cache HBM requirements to approximately one-quarter, lowering inference costs. Infrastructure engineering and deployment optimisation are becoming as important as model architecture. → Competitive landscape: Expert favours DeepSeek (cost efficiency), Moonshot AI/Kimi (complex reasoning) and ByteDance (distribution and ecosystem). Future differentiation increasingly depends on inference economics and application integration.

Nomura: AI Lab Concerns vs. Memory Fundamentals → Remains bullish on memory, expecting supply constraints to persist through at least 2028. Views market concerns around CXMT's 4F² transition as excessive. → OpenAI + Anthropic combined revenue estimated to grow +507% YoY in 3Q26, despite recent sequential growth concerns. → Argues AI revenue deceleration is partly distorted by reporting differences, customer migration and seasonality, rather than deteriorating underlying demand. → 4Q26 revenue hurdle: Combined OpenAI/Anthropic revenue needs to reach $43.9bn (+61% QoQ) to meet market expectations, versus $29.7bn at unchanged September run rates. → AI profitability underestimated: AWS, Microsoft Intelligent Cloud and Google Cloud reported operating margins of ~39%, 41% and 36%, respectively. Nomura estimates ~60% EBITDA margins across these cloud businesses. → Anthropic reportedly reached adjusted operating profitability in 3Q26, following an adjusted operating profit in 2Q26. Nomura expects further operating leverage as revenue scales. → Cash flow ≠ profitability: Rapid infrastructure investment creates a timing mismatch between upfront compute spending and subsequent revenue generation, potentially exaggerating concerns over AI economics. → Memory demand extends beyond AI labs: Nomura estimates OpenAI and Anthropic account for only ~5GW of the 29GW global AI installed capacity expected by YE26.

Goldman Sachs: Japan Semiconductors / Electronic Components → GS describes the current cycle as one of the largest and longest in history, with 2027 expectations strengthening and visibility into 2028 improving. Yet sector valuations have compressed to ~17x FY27 and ~14x FY28 earnings as investors increasingly price in the risk of a post-2028 capex slowdown. → Disco (6146): Contract liabilities (primarily customer advances) surged 46% QoQ to ¥73bn, suggesting accelerating order momentum that isn't captured in reported bookings. GS expects advanced packaging, EMIB-T, hybrid bonding and CPO to drive further shipment growth. → Capacity rather than demand is increasingly the constraint. Tokyo Seimitsu (7729) has orders running at ~1.5x production capacity, with meaningful expansion constrained until 2028. GS argues that the key differentiator is increasingly suppliers' ability to convert strong orders into revenue and margin expansion, beyond demand exposure alone.

Japan's bubble in the 80s and subsequent balance sheet deterioration meant it was unable to sustain capex at a time when Korea was aggressively expanding, leading to the brutal loss of its memory industry during the Lost Decades. This loss, although painful perhaps has left it far better positioned/diversified than Korea for the eventual bear market that follows AI? Korea now finds itself in a tricky situation (note: Korea doesn't have the same credit expansion so spillover would probably not be 3 Lost Decades grim). EM's have a tendency to be commodity dependent (DRAM/NAND), have concentrated industries (~50% of KOSPI is memory names) and volatile capital flows (refer to post below). The risk is that an eventual memory bust spills beyond equities into export income, employment and domestic credit, particularly given Korea's high household debt, while capital outflows put further pressure on the won. However, hopefully a scenario more akin to Taiwan emerges, where even a >50% drawdown in equities can be endured without undermining the industry's ability to sustain capex and maintain its technological dominance, avoiding the fate Japan suffered.

You can see that foreign money loves increased political certainty when looking at Japan (Feb Takaichi super majority) and Brazil as of late. Meanwhile Korea is the rollercoaster of doom.

UBS | Memory: Is the Wall of Worry Justified? → UBS sees the 19% memory sell-off as an opportunity, not a fundamental deterioration. Agentic AI is driving stronger-than-expected server DDR5 orders into 2027, while Nvidia's SOCAMM demand has reportedly risen above 50bn Gb, exceeding combined PC and GDDR DRAM bit consumption. HBM de-specing is also shifting demand toward conventional DRAM and NAND for inference/KV cache. → Supply fears look premature despite Samsung and CXMT expansion. UBS expects Samsung to add 120k wafers/month of DRAM capacity by 4Q27, potentially accelerating beyond 150k in 2028, but sees limited supply-demand impact until late 2028. Yap: Although memory fundamentals remain strong, it goes to show how significant a directional shift in news flow can be for price action when there are no more marginal buyers. Remember, tops form over time and should not be characterized by a single day, while fundamentals will often be a lagging indicator. Rather, the initial shift in directional headwinds can seem relatively insignificant at the time (particularly when a trade is extremely crowded). We saw this with the initial June sell-off on AVGO earnings and AI financing concerns, which have since been exacerbated as of late.

Probably the most bearish sentiment I've seen with the SPY 2.5% off highs. All types of investors from the tech bros to swing and position traders. It has been a narrative driven market as always occurs in markets experiencing exponential growth so assumptions can be short-lived when you have to stare poor price action in the face. VIX finally got a spike higher after being subdued for a few months and although we are far from wash out fears one must ask themselves with a rough macro backdrop (oil crept up to $100 again) where can money sustainably rotate?

Genuine question...is there any precedents where stocks that surge hugely in magnitude in a compressed period of time (i.e. SNDK +1000% in 180 days) don't endure pain for the years that follow the peak. I.e. The returns since the cycle low actually match QCOM in dot-com which took 20+ years to reach highs again, similar to Cisco. The saying goes that the stock market takes the escalator up and the elevator down. Well, what if you take the elevator up as well? For now it's play on but pls manage risk.

Some datapoints from a memory supply-chain expert call a few weeks ago: → Conventional DRAM may now be competing with HBM for wafer allocation on economics. The call puts HBM LTA gross margins at ~58% and argues that tight conventional DRAM supply can make marginal capacity more attractive outside HBM. → The contracting model is getting much harder. Older LTAs were generally 2–3yr price/volume agreements with limited upfront commitment, while newer structures increasingly involve 5yr+ take-or-pay terms, deposits and price floors/ceilings. The bigger shift appears to be guaranteed volume and downside protection for suppliers rather than locking today's elevated prices for five years. → The call identifies Google, Meta, Microsoft and Amazon as Micron’s four largest contracted customers, with Apple, Alibaba and Tencent in the next tier. Automotive customers cited include GM, Ford and Tesla-related suppliers. → Suppliers are reportedly exercising much tighter control over the distribution channel. Distributors may be required to provide end-customer sales records, while manufacturers can conduct warehouse checks and potentially revoke allocation or authorisation if inventory is being hoarded rather than sold through. → SanDisk is said to be requiring NCNR terms alongside ~30% upfront deposits from some distributors, effectively pushing parts of the channel toward much harder long-term supply commitments. → Customer importance can also differ materially by supplier. The call gives an example of the same distributor being Tier 3 at Micron but Tier 1 at SanDisk, reflecting SanDisk’s greater reliance on large retail/channel partners and resulting differences in allocation and commercial terms. → HBM4/HBM4E is also being used as a pricing reset. The call cites HBM3/3E contracts around ~$530–550 per device versus ~$800–1,400 for HBM4/4E. The exact magnitude is difficult to verify given pricing varies materially by stack, density, customer and contract period. → Interestingly, the expert was less convinced that tight memory supply necessarily proves equally tight underlying AI demand. Data-center capacity can be built ahead of available power and utilization, making it difficult to distinguish actual consumption from capacity being secured ahead of future demand.

HSBC: Japan Repatriation + Carry-Trade Risk → Japan is becoming a potential source of global liquidity tightening. With 10Y JGB yields around 3%, the highest since 1996, Japanese investors finally have a more credible domestic alternative to overseas bonds. → Raises the risk of capital being repatriated back into Japan. HSBC flags faster BoJ tightening, renewed FX intervention or a GPIF shift toward JGBs as possible catalysts. → The Treasury market is especially exposed. Japan remains the largest foreign holder of Treasuries, so even a marginal allocation shift could pressure global yields. → The bigger risk is the yen carry trade. Higher Japanese rates + a stronger yen raise funding costs and reduce the FX profitability of leveraged positions elsewhere. → In that scenario, Japan tightening becomes a global story: Treasury selling pushes yields higher while carry unwinds force deleveraging across risk assets. I'd say more credible tail risk is if rising rates cause balance-sheet stress at regional banks + insurers, leading to forced foreign asset sales for liquidity. No reason to stress rn, look for a decoupling of banks from the long end during a period where the velocity of the move surges as a warning sign, as we witnessed in Davos earlier this year. Correlation breakdowns scare investors a lot, especially if they're macro or in quant.

Trend beginning to emerge now of Japanese car companies getting into humanoids. Longer-dated, yes but as discussed, these names are beaten up/cheap and have been affected by agentic age cost-push inflation. Could the next physical AI phase be their time to shine? Toyota has big factory automation plans from 2028, and it should come more into focus whether their manufacturing scale, factories and mechatronics capability allow them to become a meaningful robot producer and platform (beyond just a robot user).

JPM: ASICs, Networking & WFE Take the Baton → AI infra remains supply-constrained, not demand-constrained. Cloud capex estimates keep moving higher, now at roughly $953B / $1.41T / $1.54T for 2026–28. → ASICs are taking a much bigger share of AI compute. JPM sees custom AI ASICs at $60–70B in 2026, growing 40–50%+ CAGR, with ASIC/XPU units overtaking GPUs by 2027. → Broadcom is central to this shift. Exposure spans Google, Meta, OpenAI, SoftBank/Arm and potentially Anthropic, with JPM modelling AVGO AI revenue at $58B → $135B → $245B from FY26–28. → OpenAI ASICs are already moving into reality. First silicon is out of fab and functioning, Gen 1 is on track for YE26/CY27, Gen 2 is taping out and Gen 3 design has started. → Networking is becoming another major bottleneck. Broadcom’s Tomahawk 6 is already sold out for CY27, with demand rising across switching, optics and SerDes. → The WFE cycle still looks very strong. JPM models +31% growth in 2026 and +38% in 2027, driven by TSMC N2/N3, DRAM capacity and broader AI fab buildouts. → The bottleneck is moving into the physical stack. JPM explicitly highlights cleanroom constraints as the primary WFE bottleneck, alongside packaging and tool availability. → Next phase is increasingly about ASICs, HBM, networking, packaging and manufacturing capacity.