🟢 [E3533] HBM wafer starts growing ~50% annually translates to 100-200% more process steps per wafer, each requiring subsystem content. Memory testing is becoming a critical bottleneck — NVIDIA has reportedly tightened standards requiring suppliers to fully test HBM packages before sending to foundry for advanced packaging integration.

🟢 [E3519] Citrini created SpaceX Supply Chain basket with Taiwan-focused names deriving revenue from SpaceX/Starlink spending. RF/assembly vertical includes Universal Microwave (3491 TWO) for mmWave components, MTI (2314 TT) for terminal assembly. PCB layer includes Compeq (2313 TT) for ground stations. Energy includes Taiwan Cement (1101 TT) Molicel batteries for mission hardware.

🟢 [E3518] NOR flash supply is tightening with prices increasing as TSMC and Samsung cut 8-inch capacity. TrendForce projects 8-inch output decline in 2026. Macronix (2337 TT) reduced NOR production to increase MLC NAND output, with reports of ~30% NOR price hike in Q1 2026. Samsung's MLC end-of-life announcement (March 2025, final shipments June 2026) drives global MLC NAND capacity down ~41.7% YoY in 2026.

🟢 [E3517] Taiwan legacy memory names Nanya Technology (2408 TT), Winbond (2344 TT), and ESMT (3006 TT) are rallying as DDR4 faces severe global shortages — at one point trading at a premium to DDR5. Memory buyers engaged in panic buying in Q1. DDR4 pricing could rise through Q1, Q2, and potentially Q3 with double-digit quarterly increases as supply tightness spills over to DDR3.

💬 [E3539] CrowdStrike CEO George Kurtz stated 'AI agents operate with superhuman speed and access, making every agent a privileged identity that must be protected.' CrowdStrike acquired SGNL for continuous access control and Pangea for AI security. Rubrik acquired Predibase and launched Agent Rewind to enable organizations to undo mistakes made by agentic AI.

🟢 [E3513] Citrini sold SaaS positions within a month of purchase after AI-fueled competitive threats materialized. They note the selloff in software related to agentic coding advances, citing Claude Opus 4.5 bringing 'agentic coding into median consciousness.' SaaS represents a 'knife we aren't keen on catching' — a nuanced approach is required for underwriting longer-term SaaS growth.

🔴 [E3514] Cybersecurity stocks sold off alongside SaaS names as 'baby/bathwater' trade despite fundamentally different exposure. Citrini argues selling cybersecurity names because of agentic coding is 'stupid' — if AI agents proliferate across enterprises performing autonomous tasks, accessing systems, making API calls, cybersecurity becomes MORE important. CRWD, NET, RBRK, PANW, ZS were sold as if they were the same trade as TEAM and CRM.

🟢 [E3515] NVIDIA has traded sideways for 6 months despite dominance, as investors chase second/third-order winners in memory, optics, and packaging. Citrini argues the market has gotten so focused on supply chain plays that NVDA is now underappreciated. TPU threats exist but the pie is still growing. Trump's decision to allow NVIDIA chips to be sold into mainland China should drive upside surprise.

🟢 [E3516] Citrini positions for NVIDIA options contracts targeting guidance upside surprise. They note that any upside surprise would more likely result in sustained move as underweight investors add exposure, rather than a single-day pop. The 'Billions' analogy suggests sometimes the most obvious opportunities (like Apple in 2015, which has nearly 10x'd) are overlooked because investors think they're too basic.

🟢 [E3526] Citrini staying short consumer electronics names as memory squeeze not fully priced in. Looking for opportunities to add memory testing and subsystem names, potentially taking profit on FormFactor (FORM). Not reducing Teradyne (TER) yet — expecting upside from robotics segment next quarter while memory test continues to perform.

🟢 [E3525] Citrini Dynamic AI basket is up ~16% YTD with +278.60% performance since inception. Best performers are advanced packaging (Unimicron, ASX, AMKR, KLIC), memory (SK Hynix), memory testing (TER, FORM), and semicap (AMAT, ACMC, KLAC, ONTO). Staying long ONTO, KLAC, AMAT, adding ASM NA and 6525 JP on ALD tailwinds. Adding GFS to UMC but keeping lower weight until concrete momentum.

🟢 [E3527] Citrini has been constructive on small caps and ex-US equities since Q3 2025 State of the Themes. IWM has outperformed SPY by ~10% (22% annualized) since turning constructive. They see deep value in SMID caps aligned with durable trends despite historical headfakes. Current allocation focuses on memory, advanced packaging, semicap, and memory testing exposures.

🟢 [E3552] ASMPT (522 HK) is key bonding-equipment supplier with deep SK Hynix relationship. SK Hynix placed order for 30+ TC bonders, targeting ~100 total for HBM4 with meaningful ASMPT portion. Samsung reportedly in discussions with ASMPT for TC bonders. ASMPT supplies both TC bonders and hybrid bonding tools to TSMC. CXMT could use ASMPT TC bonders for HBM production given Hong Kong/Singapore ties.

💬 [E3528] To catch the most asymmetric AI trades, investors needed to understand things beyond 'GPU-deep' and have strong views on supply chain dynamics 6-12 months ahead. The generalist understanding that sufficed previously will rapidly begin changing. Semiconductor complexity can easily leave investors behind, robbing them of understanding the full picture while ensuring proper positioning.

🟢 [E3505] HBM4E will introduce hybrid bonding, moving from thermal compression to direct copper-to-copper pad bonding. Samsung confirmed hybrid bonding adoption in HBM4E with yields reportedly highly mature. BESI is positioned as key beneficiary — potentially securing 50-60% of SK Hynix demand and 80%+ of Micron's hybrid bonding equipment needs.

🟢 [E3521] DDR4 is facing severe global shortages, trading at premium to DDR5. Samsung, SK Hynix, Micron, and CXMT are phasing out DDR4 as they pivot to DDR5/HBM. Memory buyers engaged in panic buying in Q1. Citrini expects DDR4 pricing to rise through Q2 and possibly Q3, with double-digit QoQ increases. DDR3 shortages are now emerging as spillover.

🔴 [E3553] Hanmi Semiconductor (042700 KS) currently holds ~90% share in HBM4 TCB equipment but faces significant risk from large-scale hybrid bonding adoption in HBM4E through HBM5. SK Hynix is jointly developing second-generation hybrid bonding equipment with Hanwha Semitek, threatening Hanmi's dominant position.

🟢 [E3520] UMC partnered with imec for silicon photonics process supporting 800G and 1.6T pluggable optics, with risk production planned for 2026-2027. The imec iSiPP300 process supports LiNbO₃ modulators enabling 400G+ per lane performance. UMC is trading as a commodity player but building specialty portfolio in power semiconductors, mixed-signal, and photonics.

🟡 [E3512] Memory bears have salient points — supply gluts are likely to happen at some point (author guesses 2027) as Samsung and SK Hynix historically lean in, increase capacity and inevitably glut the market. Human behavior is predictable and greed wins out even with cartel-like dynamics. However, the authors argue this misses how new capacity comes online — new capacity requires new capex spending, which is another player's revenue.

🟢 [E3559] Silicon photonics addresses AI cluster bandwidth constraints as east-west data center traffic explodes. Scaling copper beyond 224Gb/s is practically impossible — higher speeds shrink electrical reach and require more retimers adding power and cost. PICs enable 400+ Gbps with light pulses through silicon waveguides. Co-packed optics may eventually integrate into GPU/ASIC architecture but thermal management challenges remain.

🟢 [E3554] PDF Solutions (PDFS) provides manufacturing intelligence for fabs via Exensio platform — normalizing data across FDC, test, assembly, packaging into single semantic model. Has upside to Intel 18A node and advanced packaging (EMIB/Foveros). Uses 'forward deployed engineers' model making contracts stickier. Not generic enterprise software so relatively isolated from SaaS selloff.

🟢 [E3551] Powertech (6239 TT) has DRAM testing capacity nearly fully utilized with NAND also high. Guided 2026 capex over NT$40B. Q1 2026 expected much better than Q1 2025 despite normal seasonality weakness. PiFO (CoWoS-L alternative) and FOPLP give Powertech a path to incremental packaging orders, with revenue opportunity visible around 2027. Meta reportedly showed interest in PTI PiFO.

🟢 [E3560] Data centers transitioning to 800 high-voltage DC architecture, boosting demand for power semiconductors just as TSMC and Samsung cut lagging edge production. Global Foundries licensed GaN technology (650V and 80V) from TSMC specifically for data center power including 800V HVDC. This creates dual tailwind of photonics ramp and 800V DC power conversion wave.

🟢 [E3501] Memory, storage, semicap, and advanced packaging sectors have performed exceptionally well, reinforcing the supply/demand dynamics described in Citrini's 'Post-Traumatic Supply Disorder' thesis. HBM content per GPU is rising fast — NVIDIA's Blackwell Ultra GB300 contains $4k of HBM3E. The ability to scale MEMS manufacturing and solve hybrid bonding physics will ultimately determine the pace of AI advancement.

🟡 [E3523] Memory bears argue 'this time isn't different' — Samsung and SK Hynix historically lean in, increase capacity, and glut the market. Citrini acknowledges they 'probably aren't wrong' and expects oversupply 'sometime in 2027.' However, new capacity requires capex spending which creates fab equipment revenue opportunities before oversupply materializes.

🟢 [E3522] NOR flash supply is tightening with prices rising. TSMC and Samsung cutting 8-inch capacity contributes to projected 2026 200mm output decline. Macronix (2337 TT) reduced NOR production to increase MLC NAND output, with ~30% NOR price hike reported in Q1 2026. Samsung announced MLC end-of-life in March 2025 with final shipments June 2026, causing global MLC NAND capacity to drop ~41.7% YoY in 2026.

🟢 [E3524] Cohu (COHU) saw ~75% utilization in Q3 with accelerating AI data center demand. Management raised 2025 Neon HBM inspection tool revenue to $10-11M with first HBM4-configured shipment (likely to Micron). Eclipse handler selected for next-gen AI processor testing at leading US manufacturer (likely Intel) handling ~3,000W power dissipation.

🟢 [E3502] TSMC expects to spend $190-200B in capex over the next 3 years. Intel could ramp to 100k wafers per month of 18A-P + 14A capacity by 2029-30. Intel Foundry investments are front-loaded in 1H with greater equipment purchases, suggesting Intel is becoming more confident in securing foundry customers despite market disappointment with Q4 earnings.

🟢 [E3503] AI workloads require memory across the full stack. NVIDIA GPUs keep 'hottest' data in HBM while pushing 'colder' Key-Value cache to DRAM and SSDs. This structurally lifts demand for HBM, DRAM, and SSDs as video and 'world model' workloads grow. AI labs pushing toward 100M+ token contexts require far more working memory during inference.

🟢 [E3504] At 3nm and below, semiconductor wafers require 1,500+ individual process steps, each representing yield loss risk and subsystem supplier content opportunity. Each new process node increases fab cost by ~30%. Advanced packaging equipment growth extends beyond test/assembly to include gas, vacuum, and thermal systems enabling capacity expansion.

🟢 [E3509] Silicon photonics is emerging as the only viable path forward for AI data center bandwidth. Copper cabling hits a physics wall at 224Gb/s per lane. Silicon photonics enables 400+ Gbps transmission. The market for photonic components is expected to inflect within two years, with risk production planned for 2026-2027.

🟢 [E3506] Memory testing requires MEMS-based probe technology to contact thousands of microbumps (40µm pitch) without damage. FormFactor (FORM) and Micronics Japan (6871 JP) are near-exclusive suppliers of this critical probe card technology. The market is rewarding companies that validate chips and memory at extreme speeds and densities.

🟢 [E3507] Citrini created a 16-name Memory Testing basket covering capital equipment (automated test equipment, wafer probes, thermal controls) and consumables (probe cards, test slots). Key names include Advantest (6857 JP), Teradyne (TER), FormFactor (FORM), Cohu (COHU), Aehr (AEHR), and socket suppliers WinWay (6515 TT), LEENO (058470 KS), ISC (095340 KQ), Enplas (6961 JP).

🟢 [E3508] Subsystems are 'picks and shovels for the picks and shovels.' AMAT, KLAC, TEL, LRCX tools are assembled from hundreds of subcomponents — vacuums, fluid delivery, wafer handling, gas boxes. Citrini called out ICHR and UCTT as subsystem beneficiaries back in July 2025. Subsystem vendors are positioned to outperform the WFE market they serve.

🟢 [E3510] ASM International (ASM NA) has highest exposure to Atomic Layer Deposition (ALD), which becomes exponentially more important as the industry transitions from FinFET to GAAFET. More 3D surfaces and tighter tolerances make film uniformity critical. AI chips as most compute-dense and power-hungry will be first/largest volume using GAAFET processes.

🟢 [E3511] Global Foundries (GFS) is assembling a vertically integrated silicon photonics platform through its Singapore Advanced Micro Foundry acquisition, Cisco partnership, and R&D hub. GFS licensed GaN technology from TSMC for 800V HVDC data center power with Navitas and ON Semi as customers expected to begin manufacturing in 2026. This dual tailwind of photonics and power creates re-rating potential.