$MSFT u/g at Stifel: "Azure... Azure business accelerated vs expectations in June as the business benefited from incremental efficiency derived capacity and what we believe was a step-up in OAI contribution stemming from acceleration in OAI's business over the course of the quarter."

Mizuho on $CRDO: "very positive $MU / CRDO": "Our checks indicate GB300 ramps could continue, significantly stronger than previously expected, into 2H26E and into 2027E - Very Positive MU-CRDO. With GB300 at ~$4M/rack and 145KW, tight memory and land/power/shell (LPS) we see a fat BW tail driving higher rack shipments into 2027E vs VR at ~$8M/rack and ~200KW/800V-400V dual-rail racks. We believe this is consistent with $NVDA recent comments "August Blackwell shipments up 27% m/m" and 2027E GPU shipments up 2x y/y.The upside for connector suppliers comes from increased number of racks/cables, offsetting the downside of lower ASP at 100G/lane vs 200G/lane. For NVDA, we est. ~10% of OctQ units (20% of revenue) on VR/1.6T."

Wells Fargo on $WULF: "839 MW Contracted Backlog Worth $1 5/Share." multiple catalysts worth watching for WULF; (a) initial +400 MW (or greater) Muskie lease; (b) IG-backstop and project financing secured for Hawesville (leased to Anthropic); (c) closing and regulatory approvals secured for Morgantown; and (d) NY state moratorium potentially lapsing by July 2027, clearing more NY expansion pathway. WULF seems well-equipped to meet, if not beat, its 250-500 MW annual delivery target in the years ahead.

Edgewater on $MU / $SKHY: "Anthropic expected to direct purchase ~20% of memory in CY27 & grow % in CY28, in DRAM roughly same size as NVDA." "Anthropic and OpenAI are expected to start buying direct in CY27 – Anthropic is claiming 40-50B 1Gb demand next year with ~20% bought direct. The balance would be from OEM, CPU/GPU, neocloud partners, but the volume and % direct should grow in CY28+."

JPM on Neoclouds: "Across the industry, longer-term bare metal pricing seems to be around $10-12/W...We believe achieving $15/W+ pricing on bare metal (or bare metal adjacent) deals involves having access to near-term capacity and high-end GPUs with shorter contract terms and low customer prepayments, if at all. While SpaceXAI was able to sign $30-50/W deals, we believe most high end deals for neocloud providers will be in the $15-20/W+ range given pricing comments made by CRWV, $NBIS, and $IREN. The higher end is achieved through shorter contract duration and smaller customer pre-payments. Additionally, industry commentary about the resiliency of GPUs once they are past burn-in testing and $CRWV recent comment about signing A100s point to GPU useful lifespans being longer than current estimates of ~5-6 years. With the primary operating costs for the neocloud business being power and GPU depreciation, an elongation in useful life represents a meaningful uplift in profitability."

Jefferies on $AAPL: "Our tracking of six key markets shows iPhone 18P/PM's delivery lead time since pre-order began on Sep 12 is significantly shorter YoY...We see this as a reliable indicator of weak sales. Last year, iPhone 17PM/17's lead time was much longer vs the 16PM/16, which proved to be a reliable indicator of strong sales."

Mizuho on opticals: We see CW laser capacity ramping sharply from ~83M units in 2025 to ~427M by 2029E (Ex-7), driven by SiPho pluggables at 1.6T/3.2T and the initial NPO/CPO transition into 2027-2029E as we see yields improving. We see $COHR profitability lagging as we believe it relies on external lasers for it's leading-edge products (also sourced from LITE) and its business model remains more transceiver-heavy which is lower-margin vs LITE, which we see as more laser-oriented. On margins, LITE continues to ramp GMs sharply from low-30% in early 2024 to the low-50% range in 2026-2028E, while COHR rises more gradually to only the low-40s%, resulting in a 10-11ppt gap. We believe ELS pluggable modules housing a laser source could start to ramp as well. ELS as the back plate pluggable component retains serviceability in case of a component failure, while enabling laser light to transmit the converted optical signal and advance more efficient optical architectures. $LITE has noted orders for ELS have begun, with we believe potentially 2x content opportunity compared to laser chips alone. Ex-China, we believe COHR is the largest transceiver assembler with ~17% market share, differentiated via deep InP vertical integration and high in-house component supply. We note some tailwinds with $NVDA, incl. $2B investment and multi-year CPO/CW supply agreement through YE-2030E. We estimate LITE's Cloudlight business (acquired Nov 2023) is ramping strongly toward ~$900M C26E revenues and ~$1.2B by C27E, but remains smaller at >6% market share. Similar to COHR, LITE is leveraging CW/EML in house manufacturing, but we believe it is entirely reliant on in-house laser chip supply, with a path to mid-30% GM. FN is an important assembler, serving NVDA's designed transceivers alongside merchant OEM contracts. $AAOI (NC) is another key domestic assembler, operating three Taiwan facilities (~795K sq ft) with a new Pearland, Texas facility (800G/1.6T) and an estimated SOP in early 2027E.

Cleveland on $AVGO: Demand still eclipses supply 6-12 months. Units capped by TSMC. ASP / package complexity on next-gen TPU, MTIA, Jalapeno, 1.6T is the upside they are not baking in from Samsung foundry.

UBS Hot Chips 2026- Model size + KV cache keep going up. Roadmaps are now capacity, bandwidth, and scale-up — not peak FLOPS. $GOOG TPU 8i (inference, 2-die) is a memory-bandwidth chip: 384MB on-die SRAM, 288GB HBM3E, 19.2Tb/s inter-chip. Built because reasoning / MoE / agentic decode is now more memory-bound than training. TPU 8t ( $AVGO, training): 9,600 chips/Superpod, 121 EF FP4, 2PB shared HBM. Different product. Training and inference silicon have diverged enough to justify two architectures. Same conference, same bottleneck everywhere else: Samsung LPDDR5X-PIM: MAC inside the DRAM banks. 614GB/s PIM BW (8x LPDDR5X), ~3x Llama-3.1-8B tokens. Cheap inference vs HBM. CXL pooling (Samsung demo): 3.35x LLM decode at 100K context. d-Matrix 3D-DRAM: META engineer on stage. 32GB/card, ~100TB/s, 1M-token context in a 72-card rack. LPX (Groq/$NVDA): no HBM at all. SRAM-only decode. LPX + Rubin ~3-5x on long-context agentic. $CBRS CS-6: 3D DRAM on wafer-scale SRAM. ~2028. Net: TPU 8i is Google saying inference is a memory-movement problem. PIM / CXL / 3D-DRAM / SRAM decode are everyone else saying the same thing. AVGO still has the training Superpod. The duration is the memory stack sitting under both.

SIG Networking Expert call notes: “1.6T” on the compute tray is usually two ConnectX-9 800G ports. VR200 doubles optical intensity to 2:1 — second leg of the cycle without needing CPO. Spectrum-X CPO is optional. Hyperscalers will take pluggable. Laser bookings look 2× demand. The real AEC/ACC driver is the custom adjacent-rack design (NVL72 + switch rack), not SuperPOD. Meta is already on AECs. Bidirectional NVLink-6 kills redrivers. Rubin Ultra staying 2-die = incremental, not a networking event. Google choosing 300G/lane PAM4 on TPU v9 is the quietest big signal of the year. UALink is getting squeezed. Ethernet scale-up is what OpenAI and AMD are actually building. ... etc. $NVDA $CRDO $SMTC $LITE $COHR

Mizuho on SemiAnalysis (rumor mill) channel checks: "Significant Feynman de-spec risk (potentially ~50%). 16/20Hi HBM + hybrid bonding still highly uncertain. GOOGL already scaled TPUv8 HBM3e to 8Hi." "Lower HBM content per GPU/ASIC = 20-30% more shipments possible. Aggregate DRAM bit demand for training/inference/KVCache actually rises. Better GPU/ASIC margins." "That extra silicon volume is pure upside for copper AEC, NPO, SiPho/InP. $CRDO and $LITE the clearest beneficiaries. CPO still 2-3 years out." "800V looks delayed — data centers pushing back hard on rapid power architecture change + staff retraining. Moving to OCP ORv3 Diablo 400V bipolar instead. More VRMs, smoother ramp." "2027 is still a monster year: xAI/ $SPCX 10GW + the $500B Neocloud financing facility = real tailwind. $GOOGL morphing into a major Neocloud supplier with accelerating TPU ramps. Mizuho modestly cuts PTs ( $MU $1,300 / $SNDK $1,875 / AMAT $590 / MKSI $330 / LRCX $365) on multiple compression from the de-spec uncertainty — but still sees memory tight and overall bit demand higher. Net: de-spec is real, but it unlocks more systems, more copper, more optics, and keeps the 2027 power + financing flywheel intact.

UBS on $AVGO: "we are trimming ~$5B from our F2027 AI revenue number (now ~$130B) as our checks indicate we have been too aggressive on AVGO's TPU unit shipments and we now see 6.4MM shipments in CY27"

Rosenblatt lasers/opticals takeaways: "Demand for CPO/NPO has increased, and deployment timelines have not been pushed out." "Spot prices for 100G/200G EML and CW lasers have recently surged by HSD to LDD percentages on a sequential quarterly basis." "While Chinese-made lasers for 1.6T silicon photonics transceivers may become more competitive in the next 18 months, highpower and ULH power lasers for CPO, where the U.S. manufacturer's moat is deepest, are where the puck is headed" $AAOI: "On LTAs, AAOI is selectively holding off on signing, as customers want volume flexibility and price locks while AAOI wants pricing upside as shortages deepen, with leverage expected to shift further toward AAOI." "800G targeted up ~5x Q/Q in Q3; capacity sold out through at least H2 2027: Capacity is fully sold out through at least H2 2027, with new Hyper Scale inquiries being turned away." "1.6T ramping on the $200M order. Roughly $70- 80M of the announced $200M 1.6T order is expected in Q4, with the remainder into H1 2027, additional 1.6T customer qualifications underway, and future orders expected as a continuous stream, not all of which will be press-released" "Management pointed out that industry wafer capacity is nowhere near sufficient and investments are accelerating across all players;" "Management expects CPO to come before NPO/XPO integration, with Nvidia targeting CPO by the end of this year, and CW lasers are the primary focus as EML demand declines in relative terms" $LITE: "Chinese competition discounted for 2+ years. Substrate supply is secured via a 7-year LTA with a Japanese vendor and an agreement with AXT, with geographic fab diversity across the US, Japan, and Europe reducing China export control risk." "OCS revenue is expected to double Q/Q for the next several quarters, with past stumbles attributed to manufacturing and engineering ramp issues rather than design flaws. 1.6T transceivers are shipping to two major Hyper Scalers, with the first-to-market advantage driving better pricing and share, and management targets keeping transceivers at ~25% of the revenue mix while scaling higher-margin components faster." $COHR: "InP capacity doubled over the last 12 months and is expected to more than double again over the next 12 months, with 80% more lasers produced last quarter Y/Y" "More than 50% of shipped transceivers came off six-inch lines last quarter, with the mix expected to reach 75% six-inch by the end of next calendar year." "Coherent is shipping into every major Hyper Scaler OCS design currently in deployment, the liquid crystal component production bottleneck is resolved, and applications are expanding beyond scale out into scale up, including GPU sparing, dynamic compute allocation, and workload distribution across clusters" "Gross margin has expanded in 8 of the last 9 quarters, up more than 660 bps cumulatively, with Data Center transceivers (the fastest-growing segment but lowest margin) the primary headwind."

Wells Fargo on $AVGO: Estimate the following as it relates to $GOOGL: • Wells' estimates reflect an expectation that Google's TPU end-of-period (EoP) GW capacity deploy would ramp from an estimated ~2.7GWs exiting 2025 to 4.8GWs, 6.3GWs, and 8.2GWs exiting 2026, 2027, and 2028, respectively. • Google's external TPU sales commenced in 2Q26—estimated at $1.15 billion; growing to $21.9 billion in 2026 and $74.8 billion in 2027. Wells' estimates reflect Google TPU purchases from Broadcom growing from $19.0 billion in 2026 to $65 billion in 2027."

UBS on $MU: We do believe MU will ultimately buy back a very large amount of stock, but the tone in our meetings around this issue (“gradually”, “over time” to returning 100% excess cash) make us wonder whether investors should keep expectations in check around the timing of any announcement. "On near-term business dynamics, we heard:  (1) MU's latest demand commentary was consistent with recent remarks at an investor conference, citing that demand has strengthened further since earnings. (2) on HBM ASPs, MU confirmed its intention to drive HBM gross margins higher, reflecting HBM's higher value as a product and the fact that existing HBM contracts were signed a long time ago, with pricing effectively lagging the increases in DDR pricing MU provided greater granularity around its goal of 50%+ of revenue under strategic customer agreements (SCAs). Specifically, it expects ~10% of the revenue under SCAs to be sold via take-or-pay agreements without fixed/pre-contracted pricing. Of the remaining 40%, an undisclosed portion will carry fixed pricing while another portion will be sold with bands"

Mizuho on $AVGO / $LITE / $CRDO / $NVDA etc. Key points: 1) strong VR200 NVL72 ramps in 2027E, ramping 4Q26E led by xAI and META, as NVDA potentially delayed 4-die transition to Feynman with VR-Ultra on 2-die configuration, 2) we see focus on driving performance with NPO/CPO connectivity, ramping Spectrum-X (positive for LITE), 3) we see HBM content/GPU from GB300 to VR200 flat at 288GB while VR-ultra moves to 256GB HBM4e, but 70-100% higher HBM4e pricing and Feynman potentially returning to HBM content growth, 4) 2027E CoWoS growing >75% with NVDA/AVGO, and volume CoPoS ramp in 2028E, 5) INTC EMIB-T tailwinds from MTK's TPU being the first high-volume customer 2028E, and 6) CSP RPO/backlog now at $2.3T (up 3.5x y/y) as AgenticAI/inference accelerates. We remain positive with potential upside to DELL/CRDO/NVDA/AVGO/LITE in 2027E.

Semianalysis: "Market chatter suggests $GOOG is working with $AMD on a TPU project in the v10 generation. AMD’s involvement would be the first real involvement in a custom AI ASIC project, despite having a custom silicon team. AMD has strong IP especially in advanced packaging and SoIC. Additionally, CPU IP could also be a draw given Google and its customers are pushing for TPUs with on-package CPU cores for RL workloads." $AVGO: "We are seeing lower TPU output in 2H26 than previous expectations...For 2027, we also see lower TPU output than our previous expectation of 6mn units of TPU 7 and TPU8i. This is driven by greater allocation towards other customers."