SK Hynix printed 72% operating margins and management called it structural. Lam raised WFE to $140B with upside bias. TI confirmed the analog cycle is broad across all sectors and regions. Every layer of the chain is sending the same signal: AI infrastructure buildout is accelerating, not plateauing.
Management dismissed spot weakness as distribution channel flush, not demand deterioration. Customers signing multi-year LTAs prioritizing volume over price. Q2 guided DRAM ASP +20% QoQ, NAND ASP +30% QoQ. HBM4E sampling 2H26, mass production 2027.
↳ 72% margin unsustainable as Yongin cluster and M15X capex ramps. Samsung NAND yield parity still unconfirmed. If memory becomes a geopolitical bottleneck, supply chain pressure accelerates.
Demand broadening from HBM into NAND conversion ($40B pre-2028), DRAM 1C migration, and advanced packaging +50% YoY. Constraint is cleanroom capacity, not demand. Lam SAM expanding toward high-30s% of WFE.
↳ Lam gaining share in deposition and etch at key customers — relative negative read for AMAT and TEL, not just a rising tide story.
Pricing flat versus normal seasonal decline, potential price increases flagged for 2H26. TI entering Stage 1 and Stage 2 VRM sockets in data center — direct competitive pressure on MPWR and Vicor longer term.
↳ TI supply flexibility is a structural relative negative for specialty power peers. Validates ADI, MCHP, INFN upside on the industrial recovery.
TPU 8t (Zebrafish, training) built by MediaTek — not Broadcom, contrary to consensus. TPU 8i (inference) stays with Broadcom. TPU 8t carries 216GB HBM capacity. CPU-to-TPU ratio shifting to 1:1 from 2:1, implying higher Axion 2 content per cluster. Broadcom CoWoS-L capacity ~40K this year, 3-4x growth in 2027.
↳ Intel x86 partnership announced days before Axion emphasis — defensive deal for legacy Search/YouTube/Ads workloads, not a growth catalyst. X86 TAM trajectory is still down over time.
Claude Code at $10.95M ARR at SemiAnalysis before Opus 4.7 improved token efficiency. Goldman meetings with software companies say SaaS apocalypse is premature. CRM's McDermott cut guidance and quoted Buffett — bad signal. SAP raised AI revenues 50%.
↳ Medallia bankruptcy: Thoma Bravo's $6.4B 2021 buyout wiped out as debt converts to equity. AI disruption is real but this is also a peak-rate leverage story.
TI data center +90% YoY. SK Hynix LTAs extending 3 years. Lam WFE raised with upside bias. Microsoft securing 2GW+ via Nscale and Crusoe. The constraint across all layers is supply chain execution, not demand.
↳ Memory hierarchy inflection: NAND is no longer just storage. Enterprise SSDs for KV cache, QLC for AI data, SOCAMM for inference architectures — Lam called it explicitly.
China reopening trade conspicuously absent despite recent policy signals. NVDA vs custom ASIC debate under-discussed given TPU, Trainium, and inferencing chip proliferation. Enterprise software beat rates silent during earnings season.
Consensus: AI infrastructure capex continues, memory upcycle structural, soft landing intact. Crowded: long memory (SK Hynix +88% YTD), long semicap, long AI infra.
↳ Contrarian: credit sources arguing corporate debt maturities 2025-2026 force a reckoning. SK Hynix 72% margins unsustainable as CapEx accelerates. X86 secular decline underappreciated given Intel defensive deal.
Near-term: INTC foundry earnings (14A execution test), Samsung NAND yield recovery read, MU earnings memory cycle durability check.
Medium-term: SK Hynix CapEx acceleration read-through to ASML EUV orders. DeepSeek $44B fundraise close sets China AI benchmark. MCU price hike cycle — TI up to 85% on some analog lines.
↳ Tail risks: memory overbuild if HX/MU/Samsung all accelerate simultaneously (2027-2028 oversupply). Strait of Hormuz. AI ROI compression if hyperscaler margins face memory cost inflation without pricing power.
Full analysis, implications, relative trade read-throughs, positioning reads and catalysts — delivered by email at 07:00 HKT daily.