Equity research

BOCOM International: Zhongji Innolight

AI Infrastructure & Market Forecasts

> Role of Communication Networks: Zhongji Innolight is positioned to benefit as a leading global manufacturer of optical modules due to the rising importance of communication networks in AI infrastructure.

> Chip Market Forecast: The global AI communication network chip market size is forecasted at USD 107.5bn (2026E) and USD 154.9bn (2027E), driven by upstream giants NVIDIA and Broadcom.

Optical Modules in AI Data Centers

> GPU-to-Transceiver Ratio: Based on NVIDIA Blackwell Ultra NVL72 systems, 1 GPU corresponds to ~6.1 transceiver modules.

> Shipment Estimates: High-bandwidth (>=400G) optical module shipments in data centers are expected to reach 81.1m units (2026E) and 146.7m units (2027E).

1.6T Optical Modules: Expected to reach 28.4m units (2026E) and 76.3m units (2027E), serving as the primary driver for Zhongji Innolight's earnings growth.

Technology Position & Market Share

> Technology Timeline: CPO may scale up in 2H27 or later, while NPO could begin volume ramp in 2026 as the primary interface technology form.

> Global Ranking: According to Frost & Sullivan, Zhongji Innolight's data communication optical module market share was 23.1% in 2025, ranking first globally.

$Applied Optoelectronics(AAOI.US) $NVIDIA(NVDA.US) $Broadcom(AVGO.US) $Marvell Tech(MRVL.US)

Needham: "DRAM/HBM will remain a real bottleneck for AI."

> DRAM/HBM Bottleneck: DRAM and High Bandwidth Memory (HBM) are projected to remain significant bottlenecks for AI hardware scaling.

> Rapid Bit Growth: Total DRAM bit growth (millions of GB) is expected to expand dramatically through 2028, driven heavily by increasing HBM adoption.

> Surging HBM Share: HBM bits as a percentage of total DRAM are forecasted to grow steeply, scaling past 15% by 2028.

> Wafer Capacity Shift: DRAM wafer capacity dedicated to HBM (measured in kWSPM) is projected to accelerate sharply, with expectations that nearly 50% of total DRAM wafers will go into HBM by 2028.

$Micron Tech(MU.US) $DRAM $EWY

UBS: Open-Source Models Momentum

AI Demand & Enterprise Adoption

> Broadening Demand: AI demand remains exceptionally strong and is expanding rather than slowing down, despite heightened scrutiny around return on investment (ROI).

> Transition to Production: Enterprises are shifting from early testing and proof-of-concepts into full production and large-scale deployment.

> Incremental Categories of Spend: Companies like Perplexity, AlphaSense, ElevenLabs, Sierra, and Cursor reported exponential growth (e.g., Perplexity experiencing a 3x increase in ARR year-to-date) driven by new, non-cannibalistic categories of AI spending such as voice, digital coworkers, agents, and coding.

Pricing Shifts & Cost Optimization

> Consumption-Based Pricing: Vendors like Microsoft (with Copilot) are transitioning certain clients from traditional seat-based licensing to consumption-based pricing.

> Focus on Token Costs: This shift has intensified enterprise focus on maximizing ROI per token, making model-harness optimization, smart routing, and workload-specific model selection crucial components.

> Reallocation, Not Reduction: Demand is not shrinking; rather, it is being strategically reallocated toward the most cost-effective models and workloads.

Open-Source Momentum & Implications for NVIDIA

> Closing Performance Gap: Open-source models have significantly narrowed the performance gap against frontier solutions over the past year.

> Explosive Token Utilization: Open-source adoption has surged from sub-1% last year to 1% at the start of 2026, multiplied several times over by mid-2026, and is projected to potentially represent the majority of tokens used over time.

> Favorable for NVIDIA: This trend is viewed as highly positive for NVIDIA due to its software leadership (such as the Nemotron family) and the fact that most open-source models are trained and fine-tuned on NVIDIA hardware, resulting in superior inference performance.

> Expanded Infrastructure Demand: While open-source adoption pressures frontier-only economics, it ultimately expands total inference demand by turning lower-cost models into viable solutions for a broader range of mature workflows.

Fast-Inference Infrastructure

> Optimal Use Cases: Cerebras (CBRS) enterprise customer feedback indicates that wafer-scale engine (WSE) technology is ideally suited for latency-sensitive search and one-shot retrieval workloads.

> Long-Running Workloads: Long-running agentic workflows continue to prioritize model quality, tooling, and orchestration over raw inference speed.

> Market Share: NVIDIA's view that fast inference accounts for roughly 10-20% of the overall inference market remains intact, though use cases are expected to expand as token costs decline and architectural advancements (like disaggregated solutions and stacked-memory SRAM architectures) improve.

$NVIDIA(NVDA.US) $Cerebras(CBRS.US)

J.P. Morgan: AI Server Market

"The AI Server market is estimated to expand to $356bn in 2026 from $195 bn in 2025, implying +83% y/y growth, while the long-term CAGR from 2026 through 2030 is expected to track at +37%. With respect to customer types, Hyperscalers are expected to track to a CAGR of +28% from 2026 through 2030, while Enterprise and Rest of Cloud are expected to drive faster growth at +47% and +49% CAGRs, respectively, over the same period."

AI Server Market Growth

> Rapid Expansion: The total AI server market is projected to skyrocket from $14.7 billion in 2022 to $1.24 trillion by 2030.

> Hyperscale Dominance: Hyperscalers are driving the vast majority of the demand, growing from $8.09 billion in 2022 to an estimated $611.6 billion in 2030.

> Massive Peak Growth: The market saw its highest year-over-year percentage growth in 2024 at 211%, with steady, strong growth projected through the rest of the decade.

Total Server Market Share

> Nvidia's Surge: Nvidia's total market share expanded significantly from 18% in 2023 to 33% in 2025.

> Competitor Stability: Major hardware vendors like Dell and Super Micro maintained or moderately increased their standing, with Dell holding 14% and Super Micro at 9% by 2025.

AI Server Market Share by Vendor

> Nvidia Leadership: Nvidia maintains dominance in dedicated AI servers, commanding 47% of the market in 2025 (down slightly from a peak of 58% in 2024 due to rising competition).

> Gaining Competitors: Dell and Super Micro are prominent players in the AI server space, each holding an 11% market share as of 2025.

GF Securities: Taking Actions to Manage Memory Cost

Management of AI Memory Costs & Spec Cuts

> Nvidia VR200 NVL72 Adjustments: Nvidia is cutting the default SOCAMM configuration in VR200 NVL72 racks by half (shipping 96GB SOCAMM modules instead of 192GB) to address LPDDR5X supply constraints and cost optimization.

> Impact on Memory Volume: CPU-side LPDDR5X memory per rack drops from ~54–55TB down to ~28TB, while GPU-side HBM4 capacity remains unchanged at ~20.7TB per rack.

> Vera CPU Rack Outlook: Nvidia's Vera CPU racks are also expected to adopt 96GB SOCAMM per CPU (768GB total per CPU instead of the 1.5TB spec sheet), though deliveries may be delayed due to stacking and density changes with Samsung as the sole supplier.

> General Servers Lowering Specs: General server DDR5 specs are expected to be reduced by roughly 50% per CPU, with 96GB/64GB becoming the new mainstream via new RDIMM and/or MRDIMM.

Financial & BOM Impact

> VR200 Cost Reduction: Reducing LPDDR5X capacity effectively lowers LPDDR5X costs to $293k (under an extreme scenario down to one-fourth capacity) or $586k (4Q26E version), preventing costs from escalating to $1.2m.

> Controlling Total BOM: Without capacity adjustments, memory and storage costs would spike to $2.1m (29% of the total BOM); cutting capacity successfully manages memory costs back down to the ~20% level.

Market Shipments & Outlook

> Deployment Priority: The specification cuts allow Nvidia to prioritize faster system deployment and higher rack availability rather than delaying deliveries.

> General Server Recovery: General server shipments are projected to rise mildly QoQ in 2Q–3Q26, followed by a strong pick-up of 20% to 30% QoQ in 4Q26.

$NVIDIA(NVDA.US) $Micron Tech(MU.US) $DRAM $EWY

Nomura Research: TSMC CoWoS

TSMC CoWoS Capacity Expansion Trend

> Aggressive Upward Revision: According to Nomura estimates, TSMC has turned significantly more aggressive on its CoWoS capacity expansion compared to previous targets (Dec 2025 baseline).

> Growth Trajectory: Quarterly capacity is projected to scale up dramatically from around 200 thousand pieces (kpcs) in late 2025/early 2026 to near 600 kpcs per quarter by late 2027.

> Divergence from Older Projections: While previous estimates flattened out near 330 kpcs per quarter through 2026 and 2027, current forecasts show continuous sequential expansion starting from mid-2026 onwards.

CoWoS Output Breakdown (Volume Growth)

> Total Volume Surge: Total output volume is scaling multifold, expanding rapidly from 2023 levels to an estimated peak approaching 2,000 kpcs annually by 2027F.

> NVIDIA Dominance in Volume: NVIDIA remains the single largest consumer of TSMC's CoWoS capacity by a wide margin, scaling from a minority share in 2023 to over 1,000 kpcs by 2027F.

> Diversification of Hyperscalers: Volume is increasingly supporting customized accelerators and ASICs from major cloud service providers, notably Google and AWS, alongside AMD+Xilinx and Meta.

CoWoS Output Allocation Share (%)

> NVIDIA Share Stabilization: NVIDIA’s relative share of total CoWoS allocation stabilizes in the 55% to 58% range from 2025F to 2027F, after peaking earlier relative to its initial 2023 baseline.

> Google's Expanding Footprint: Google captures the second-largest share of allocation, maintaining a steady slice around 24% to 26% of total capacity through the forecast window.

> Other Players: AMD+Xilinx, AWS, Meta, and other networking/FPGA applications split the remaining allocation, with hyperscaler custom silicon taking up a stable overall proportion of advanced packaging lines.

$Taiwan Semiconductor(TSM.US) $Alphabet(GOOGL.US) $AMD(AMD.US) $NVIDIA(NVDA.US) $Meta Platforms(META.US) $Amazon(AMZN.US)

Nomura Securities: GPU Substrate

Substrate Size Trends for Key GPUs

> Overall Trend: The substrate size for key GPUs is growing over successive generations.

> Early Generations (Pascal, Volta, Ampere, Hopper): Maintained relatively stable substrate sizes, ranging from approximately 3,000 to just over 3,000 sq. mm.

> Blackwell / Blackwell Ultra: Shows a significant increase, rising to approximately 6,000 sq. mm.

> Rubin & Rubin Ultra: Jump further to approximately 8,000 sq. mm.

> Feynman: Reaches the highest projected size on the chart, with a tentative assumption of an 80x115 mm substrate size, translating to roughly 9,200 sq. mm.

Morgan Stanley: Soitec

Financial Performance & Estimates

> F1Q27 Beat: Revenue came in at €113 million, beating street estimates by 6% and marking a 23% year-over-year growth ( outpacing the company's 15% guide).

> Strong Guidance: F2Q27 revenue growth is guided at more than 30% year-over-year, blowing past the 4% consensus expectations.

> Significant Earnings Upgrades: Morgan Stanley raised its revenue estimates for FY27–FY29 by 8–19% and upgraded EPS for FY28 and FY29 by 43% and 32% respectively, projecting almost €8 of EPS by FY29. 💡

Core Investment Drivers (Photonics-SOI)

> The Growth Engine: The stellar performance and guidance are heavily driven by Photonics-SOI demand for high-speed optical interconnects in AI data centers.

> Doubling Revenue: Management explicitly noted that Photonics-SOI FY27 revenue is expected to more than double, acting as the catalyst the market had been anticipating.

> Secular Alignment: Soitec’s stellar outlook mirrors broader industry tailwinds from major players scaling up silicon photonics (SiPho) capacity—such as Tower Semiconductor, STMicroelectronics, and TSMC. Photonics is expected to scale from roughly 8% of Soitec's revenue in FY25 to about one-third in the current fiscal year.

Scenario Analysis (Price Target Cases)

> Bull Case (€300.00): Assumes a stronger recovery in RF-SOI alongside more aggressive, accelerated growth in Photonics-SOI, paving the way for €10 in earnings power.

> Base Case (€200.00): Assumes multi-year growth in Photonics-SOI and normalization in RF-SOI, applying a 25x multiple to CY28 (FY29) EPS.

> Bear Case (€80.00): Assumes a slower RF-SOI recovery and weaker Photonics growth, applying a 20x multiple to discounted bear-case earnings.

Morgan Stanley: Intel CapEx Spend

Core Forecasts and Capex Revisions

> WFE and Capex Projections: The 2027 WFE forecast is set at $202 billion (a 31% year-over-year increase), which incorporates a $25 billion capex assumption specifically for Intel, though analysts see potential for further upside.

> Upward Revisions: Intel previously revised its 2026 capex outlook to over $20 billion and signaled that 2027 capex will be "significantly above 2026 levels".

> CPU Demand Drivers: Producing 5 million incremental Xeon 6 units requires an estimated $5.3 billion+ in WFE. Expanding logic wafer capacity to capture two-thirds of incremental CPU units (22.4 million units through 2030) demands a minimum of $23.8 billion for Xeon 6 and $27.9 billion for Xeon 7.

Key Equipment Beneficiaries

> Outside the US:

Tokyo Electron (TEL): Historically counts Intel as a major customer (ranging from over 10% up to 20.4% in F3/20).

Lasertec: Has seen Intel represent 10% or more of its customer base historically, peaking at 31.6% in F6/22.

> Within the US:

KLA Corp (KLAC): Highlighted as an outsized US beneficiary. Although Intel has not been a 10% customer since F6/15, Intel is heavily leaning on KLA to improve manufacturing yields and margins to succeed as a viable foundry, positioning KLA for potential massive revenue growth from Intel between 2025 and 2027.

$KLA $Intel(INTC.US)

Major news on CXMT & YMTC all at once.

1. CXMT plans expansion that may more than double monthly wafer output

2. CXMT inks five-year ByteDance supply agreement valued over $7 billion.

3. Trump administration split on additional limits for CXMT, YMTC

4. YMTC executives pushing internally for trillion-yuan IPO valuation target

5. CXMT, YMTC in some cases charge more for memory chips than Samsung, SK Hynix

6. Apple sought assurances CCXMT won't be added to U.S. entity list.

A Senior Executive at Supply Chain Logistics Company on fiber cable:

"We're seeing a huge restriction in fiber cable availability, which that's a very critical part of the network and builds just cannot happen if that fiber is not available. That's for sure creating some concern out there."

Market Demand and Supply Overview

> Seasonal Demand: Demand for fiber buildouts is strong, following the typical seasonal peak that runs from late March/April through the end of the summer.

> The Primary Bottleneck: Bulk fiber cable is severely restricted due to high demand from the data center and hyperscale environment (such as Amazon, Google, and Meta), leaving less capacity for traditional telecom builds.

> Other Components: Aside from fiber cable, other products and miscellaneous items (such as conduit and hand tools) are experiencing no supply or lead time issues.

Impact Across Telecom Tiers

> Tier 1 Providers Protected: Major service providers like AT&T, Verizon, Brightspeed, and Lumen are largely insulated from shortages because they have direct contractual allocation programs with major manufacturers.

> Tier 2 & Tier 3 Providers Impacted: Smaller regional providers (such as Cincinnati Bell or Consolidated Communications) are facing significant challenges, with an estimated 25% of their builds currently impacted or pushed out due to a lack of available fiber.

Manufacturing Shifts (Corning and CommScope)

> Corning's Hyperscaler Focus: Corning has locked up mega-deals with hyperscalers like Amazon Web Services, Google, and Meta, turning its focus toward data center customers.

> CommScope Supply Disruption: Corning cut off its long-standing partnership of supplying raw glass to its competitor CommScope to satisfy its own customer demands, heavily disrupting CommScope and causing lead times to stretch out by months.

> Capacity Timeline: Industry participants are advised that it will take 1.5 to 2 years for new U.S. manufacturing capacity to come online to meet the surge in demand.

Active Electronics Bottlenecks

> Secondary Concern: Active network equipment requiring memory chips—sourced from vendors like Adtran, Nokia, and Calix (including ONTs, OLTs, CPE, and servers)—represents a growing but secondary bottleneck.

> Cost and Lead Times: These active components are facing double-digit cost increases and doubled lead times, though they are not currently causing widespread project cancellations.

$Corning(GLW.US)

Former Strategic Sales & Business Development Manager at Intel:

"Google TPUs are trying to beat that performance per token with 30-40% less power requirement compared to Nvidia, so that's the main point... the performance per token will increase probably by 40x to 50x in two years."

> Packaging Bottlenecks & Reticle Size Limits: TSMC's Chip-on-Wafer-on-Substrate (CoWoS) holds proven yields but is structurally capped at 6.5x to 7.5x reticle sizes. This creates a bottleneck as AI accelerators demand larger dies, more computing cores, and expanded high-bandwidth memory (HBM) stacking.

> Intel's EMIB-T vs. TSMC: Intel's EMIB-T (Embedded Multi-die Interconnect Bridge with Through-Silicon Via) offers projected yields above 90% and is estimated to be roughly 30% lower in cost compared to CoWoS. However, it remains commercially unproven against the 95% threshold required by major customers like Google. Furthermore, Intel faces packaging capacity constraints compared to TSMC's massive expansion.

> Cost Breakdown: Advanced packaging costs typically run under $5,000 to $7,000, representing a fraction of a total system-on-chip (SoC) cost that can reach around $60,000 for advanced 3.5D setups.

> Google's TPU Roadmap: Google is pursuing an aggressive trajectory targeting 30% to 40% lower power than Nvidia alongside a 40x to 50x increase in performance-per-token within two years. This push is driving faster dual-track adoption of 3.5D packaging solutions using small outline integrated circuits and Intel's EMIB-T.

> Glass Substrates Timeline: Meaningful mass production of glass core substrates for high-volume AI accelerators is realistically expected to land around 2029. Consequently, near-term EMIB-T adoption relies on conventional organic substrates.

> Foundry Preferences (14A vs. 18A): If Intel's packaging solutions successfully mature, large hyperscalers are more likely to lean toward Intel's 14A node rather than 18A. This preference is driven by 14A's power and area metrics being comparable to TSMC's N2P process, coupled with potential cost savings and relief from TSMC's allocation bottlenecks.

$Alphabet(GOOGL.US) $Taiwan Semiconductor(TSM.US) $Intel(INTC.US)

Former Senior Director, SoC Design & Development, Data Centre & AI Group at Intel:

"USD 120bn server CPU TAM forecast by 2030 is considered conservative, with physical AI, edge inference and agentic orchestration potentially driving 30-40% upward revision within one year."

> Market Split & AMD's Rise: The x86 server CPU market is currently split roughly 60/40 between x86 and Arm. Within x86, AMD commands nearly 50% of the server segment following years of historical Intel dominance.

> Agentic AI & CPU-to-GPU Ratios: Agentic AI workloads and mixture-of-experts models are dramatically reversing the historical training-phase GPU-to-CPU ratio (which favored GPUs heavily at 4:1 up to 8:1) back towards a 1:1 or 1:2 ratio, with expectations to hit 1:8 in favor of CPUs within a year due to intensive reasoning and orchestration demands.

> Conservative TAM Forecasts: The USD 120bn server CPU Total Addressable Market (TAM) forecast by 2030 is viewed as conservative. Physical AI, edge inference, and agentic orchestration could potentially drive a 30-40% upward revision within a year.

> Competitive Landscape (Venice vs. Diamond Rapids): AMD's Venice architecture (reaching 256 cores) provides it with an unofficial 12-month head start over Intel's Diamond Rapids, making AMD breaching a 50% x86 server share by the end of 2027 a realistic possibility.

> Intel's Structural Advantage: Despite operating as two distinct business entities, Intel’s tight internal coupling between design and manufacturing provides an inherent 20-30% architectural advantage once foundry execution and the 18A node stabilize.

> Arm's Core Demand Impact: Arm's CEO, Rene Haas, noted that Agentic AI is scaling CPU core demand up to 4x higher, pushing up to 120 million cores per gigawatt.

> Supply Constraints: Current severe shortages—with AMD sold out through the end of 2026 and lead times hitting 30-plus weeks, while Intel underships true demand by roughly 20%—are tied to the entire integrated ecosystem of advanced front-end logic wafer capacity and tied-in packaging capacity rather than memory.

> Volume vs. ASP Strategy: About 66% of AMD's server CPU growth comes from unit volume increases rather than Average Selling Price (ASP) expansion, contrasting with Intel's historical high-ASP strategy.

$Intel(INTC.US) $AMD(AMD.US)

Former Senior Finance Role at Nebius Group says,

"The problem of CoreWeave is their business model is sign a contract and then allocate data centres, then allocate GPUs and buy the GPUs they don't have and then start project management. The whole cycle within CoreWeave is about 24 months. The majority of their revenue, maybe 95% or more is still Hopper revenue."

"There are neoclouds like Nebius who do it a bit different. We were building data centre capacity, buying GPUs without knowing to whom we're going to sell it to. Coreweave only start working when they have a signed contract and then they're going to look for the data centre capacity and the GPUs, which is fully hedged. It's like without risk, you have always a coverage by a five-year contract. The disadvantage is you can't monetise fast."

> Revenue per Gigawatt Trends & Waves: Revenue per gigawatt does not decline in a straight line; instead, it moves in cycles tied to GPU generations. Prices spike at the launch of a new GPU generation (like Blackwell or Vera Rubin) and then gradually decrease over time within multi-year contracts. CoreWeave’s revenue per gigawatt has decreased over recent quarters because older contracts signed during peak Hopper pricing are rolling alongside newly added capacity at lower rates.

> Business Model Differences (CoreWeave vs. Nebius):

CoreWeave follows a risk-hedged model where they sign multi-year contracts (typically five years) with major customers first, and then procure data center space and Nvidia GPUs.

Nebius often builds out data center capacity and secures hardware proactively before locking down every final end customer, allowing them to monetize faster and cater to urgent capacity needs or mid-sized enterprises.

> Long-Term Outlook on Hyperscalers: Nebius anticipates that hyperscalers (like Microsoft and Meta) will only utilize neoclouds as stopgaps for five to eight years until they fully internalize their own infrastructure. Consequently, Nebius is positioning itself as a "mini hyperscaler" targeting enterprises, European entities seeking data sovereignty away from U.S. jurisdiction, and diverse corporate clients.

> Memory Pricing & Unit Economics Impact: Memory prices (such as high-bandwidth memory) cannot typically be locked in when striking multi-year customer agreements. However, memory accounts for a relatively small percentage of overall unit economics (e.g., roughly 2% to 3% historically), meaning even a 2x-3x price spike only minorly impacts contribution margins. Neoclouds offset these headwinds through advanced natural air and water cooling (saving up to 30% on power costs) or by slightly extending GPU depreciation timelines (e.g., from 5 to 5.5 or 6 years).

> Data Center Outfitting Timelines: For a standard 50MW greenfield powered shell, outfitting and bringing servers online takes about two months, while larger 300MW sites with roughly 90,000 GPUs take 3 to 4 months. While moving from Hopper to Blackwell requires similar configuration windows (around 6 to 10 weeks), upcoming Vera Rubin chips are vastly more powerful and thermally intense, meaning older colocation sites will struggle to handle them without expensive retrofits.

> Financing & Profitability Realities: CoreWeave utilizes heavy debt financing (such as recent deals at a 9.5% interest rate), creating a large interest expense burden. While CoreWeave anticipates operating margins jumping significantly as new capacity comes online, the expert is skeptical of rapid jumps (such as 1% to 16% in the short term) given high capital costs and market impatience regarding profitability.

> GPU Utilization Targets: Neoclouds actively target high capacity utilization rates exceeding 90% (often 94% to 95%) on external customer data centers to maintain competitive pricing against hyperscalers, who typically settle for 75% to 80%.

$Nebius(NBIS.US) $Coreweave(CRWV.US)

Texas Instruments $Texas Instruments(TXN.US) doubled DC sales YoY and grew gross margins by 330 bps QoQ. FCF exploded with cut back in CapEx, but this is one of the few companies that was ahead of the curve with all their CapEx spend to get capacity ready.

They mentioned in the CC that power semis content will only grow as more DCs get built and come online.

They had a great report despite the AH reaction, and expect data center sales to be strong for the foreseeable future.

Power semis look good while many are ignoring them for other names.

$Infineon Technologies(IFNNY.US) $ON Semiconductor(ON.US) $STMicroelectronics NV(STM.US) $Analog Devices(ADI.US)

BofA: Anthropic & AMD Deal

Key Takeaways from the Anthropic Partnership

> Partnership Details: AMD announced a partnership with Anthropic to deploy up to 2GW of AMD MI450 Series AI GPUs starting in the first half of 2027.

> Financial Commitments: AMD committed to a $5 billion strategic equity investment and will use Anthropic's Claude in research and development (R&D) projects.

> Expanded Engagements: This deal expands AMD's MI engagements to four major customers: Anthropic, OpenAI (likely via Microsoft cloud), Oracle, and Meta.

> Financial Upside: The engagement indicates at least a $10 billion to $15 billion upside to current consensus expectations of $35 billion to $40 billion for CY27 MI sales.

Analyst Event & Product Outlook

> Advancing AI Event: Pre-announcements involving Microsoft and Anthropic serve as incremental positives ahead of AMD's analyst event.

> Server CPU Roadmap: The event is expected to spotlight the server CPU roadmap (specifically upcoming Venice CPUs) and how they compete against NVIDIA's Vera in the agentic AI era.

> CoWoS Supply & Sales Potential: Latest TSMC Chip-on-Wafer-on-Substrate (CoWoS) supply estimates point to roughly 1.4 million MI450 Series GPU units in CY27, which could surpass $40 billion in sales assuming a $30k average selling price (ASP).

$AMD(AMD.US)

Manager at a Global Distributor on optical component demand:

"Overall optical components market exceeded our expectation because we were not projecting such a high growth rate for us. Customers were accelerating transition from 400G to 800G."

Vendor Growth Metrics:

"Between Coherent and Lumentum, Coherent performed the strongest."

> Coherent: Saw estimated demand growth increase by 8% quarter-on-quarter and 25% year-on-year, driven by 800G optical components, high-performance lasers, and AI cluster interconnected solutions. Coherent dominated in meeting increased optical bandwidth requirements per GPU cluster, high-fiber connectivity requirements. Overall performance was supported by increased networking investments from hyperscalers.

> Lumentum: Experienced growth in the range of 5% to 7% quarter-on-quarter, or 18% to 22% year-on-year, bolstered by cloud and AI demand growing at 25% to 35% year-on-year.

Speed Segment Breakdown:

> 400G: Remains flat or exhibits low-single-digit growth year-on-year, serving as a volume product for traditional cloud and enterprise deployments.

> 800G: Acting as the core growth engine for AI infrastructure, seeing 50% to 100% year-on-year growth.

> 1.6T: Moving from early evaluation into initial procurement or production validation testing (PVT) phases by hyperscalers, with a 30% quarter-on-quarter demand growth from a small base and a commercial ramp-up expected by 2027.

> 3.2T: Confined to prototype and roadmap discussions with fewer than 10 global customers, contributing less than 5% of current optical revenue.

Pricing & Margins:

> Average selling prices (ASPs) for 800G AI optic modules saw a 10% increase year-on-year driven by product mix and tight supply.

> Quarter-on-quarter ASP growth for 800G is tracking at 2.5% to 5%, while high optical fibers are up 1% to 4%.

> Distributor margins are tightening on mature optical products and telecom optics due to increased competition and price erosion, while high-speed AI optics maintain stability.

Competitive Dynamics & Contracts:

> Chinese vendors (such as InnoLight, Eoptolink, and Accelink) are aggressively pricing and winning share in cost-sensitive, high-volume segments like 400G and pluggable optics, primarily displacing Lumentum within China.

> Coherent typically engages hyperscalers on 1- to 3-year multi-year contracts for Al optical products, while Lumentum primarily secures 1- to 2-year strategic agreements (such as for EML lasers).

$Coherent Corp.(COHR.US) $Lumentum(LITE.US)

Former Senior Post Silicon Validation Engineer at NVIDIA on CPO:

"Eventually, CPO is where the technology is converging and going to converge. Whoever comes with a reliable CPO product first, they will see a huge demand. There is absolutely no doubt about that. All these NPO and LPO are temporary compromise, for the lack of a better word."

Current State & Commercial Timeline

> Experimental Phase: CPO technology is currently in pilot and experimental stages, with no commercial products available yet.

> Timeline for Adoption: Initial CPO products are tentatively expected around 2028, though unexpected delays could occur due to major, unresolved engineering hurdles.

> Demand Dynamics: Once CPO is successfully and reliably commercialized, demand will experience an explosive surge, limited only by production supply rather than market interest.

Key Technical Hurdles

> Thermal Management: Compute chips generate kilowatts of heat, whereas optical engines are exceptionally sensitive to temperature changes (even minor shifts can alter laser wavelengths).

> Glass Substrates & Via Drilling: Transitioning from traditional organic substrates to glass is necessary for scale, but drilling microscopic, precise vias into glass without causing micro-cracking remains a massive challenge.

> Laser Integration: Internal laser setups face severe thermal interference from compute chips, while external laser setups struggle with high optical loss when trying to focus light onto extremely thin substrate layers.

Competitive & Technological Landscape

> Interim Solutions: Technologies like NPO (Near-Packaged Optics) and LPO (Linear Drive Pluggable Optics) are viewed as temporary compromises; once CPO matures, the industry will fully converge on it, causing demand for traditional pluggables to decline.

> Key Players: Major industry stakeholders—including TSMC, Samsung, Intel, and Amkor—are heavily investing in packaging R&D, with TSMC viewed as a dominant pioneer via solutions like CoWoS and CoPoS.

> Laser & Component Suppliers: Lumentum is positioned well ahead of competitors in the optical landscape, supported by strategic partnerships and investments (such as NVIDIA's capital backing).

Bill of Materials (BOM) Breakdown

> Packaging: Forms the largest share of the CPO BOM (~50% to 55%) and represents the hardest problem to commoditize.

> Light Sources / Lasers: Accounts for roughly one-third of the BOM.

> Cooling Solutions: Makes up the remaining portion, playing an equally vital role due to severe thermal constraints.

Optical Circuit Switching (OCS)

> Role in Future Infrastructure: Once signals are converted to light via CPO, OCS will become critical to route traffic dynamically across large clusters of GPUs without re-converting signals back to electrical (which destroys power-saving benefits).

> Current Limitations: Existing OCS solutions lack packet-decoding capabilities and rely on rigid, round-robin scheduling schemes, meaning optical switching technology must heavily evolve alongside CPO.

$Marvell Tech(MRVL.US) $Amkor Tech(AMKR.US) $NVIDIA(NVDA.US) $Taiwan Semiconductor(TSM.US) $Marvell Tech(MRVL.US) $Lumentum(LITE.US) $Intel(INTC.US) $Applied Optoelectronics(AAOI.US)

Former manager at Google on data center interconnect market and spending:

"Your neoclouds, their spend for DCI is double because they don't have that infrastructure in there yet. They're brand new. Some of these guys have only been around five years. They don't have a network between all their buildings yet. They're building that out. Their spend is double."

> 2026 Total Addressable Market (TAM): Estimated at USD 10 billion to USD 15 billion in total DCI orders for 2026, encompassing coherent optical modules, line systems, switches, and software at the system level.

> Long-Term Growth: Most estimates project a Compound Annual Growth Rate (CAGR) of 20% to 30% over the next 3 to 5 years, with the low end around 15% to 20% and the high end around 30% to 35%. The expert suggests the industry could scale up to a USD 30 billion TAM by 2030.

> DCI vs. Scale-Across: The expert highlights that traditional DCI was for redundancy and backup, but AI has created a new category called "scale-across". Scale-across connects multiple buildings within a mega-site campus (e.g., five 200MW buildings roughly 2,000 meters apart) and represents about 50% of the 100GW of new capacity coming online by 2030. True DCI connects data centers hundreds of kilometers apart (or in the 10km to 40km optimal range for AI clusters) due to power constraints and model training scale.

Executive at POET on $Applied Optoelectronics(AAOI.US):

"AOI having their own laser source is a huge differentiator...The laser supply will continue to be constrained. The future networks, the laser power has to increase significantly."

> 800G optics are positioned to remain a dominant mainstream workhorse for the next 12 to 24 months (with a solid 2 to 4-year total lifecycle), giving AOI a durable, multi-year runway before 1.6T fully takes over.

> Owning in-house indium phosphide laser fabrication serves as a vital competitive differentiator and moat, insulating AOI from industry-wide component shortages and positioning them ideally for future high-power architectures like ELSFP (External Laser Small Form-Factor Pluggable).

> Supply constraint in transceivers expected to last several more years.

According to a senior engineer at Inphi:

> Hyperscalers prefer to stay on 800G as long as possible to maximize existing infrastructure investments unless specific inference workloads (like long-context or domain-specific enterprise models) necessitate a faster migration to 1.6T.

>While some anticipate 3.2T deployment by late 2028, a true majority market field deployment is more realistically expected closer to 2030.

> Regarding Credo, the expert notes that the company operates in a similar space to Marvell's scale-up segment and achieved major success by dominating the active copper cable market with a well-timed solution after nearly failing in the past. Many of Credo's engineers previously worked within the expert's group at Marvell, and they are recognized as having strong engineering capabilities. Credo initially capitalized because Marvell's product management and roadmap focus were directed toward optical connectivity rather than that specific copper segment. However, the expert believes that Marvell's next-generation offerings position them on par or ahead in the ongoing race, even though Credo will not easily relinquish its dominant position.

$Marvell Tech(MRVL.US) $Credo Tech(CRDO.US)

BofA: China Semiconductors

> Shift to System-Level Supernodes: China's leading AI accelerator and infrastructure vendors are shifting focus from standalone chip performance to system-level integration through broad product launches of rack-scale supernodes ranging from 64 cards to over 1,000 cards (e.g., Huawei and Biren architectures). This strategy helps offset individual chip performance gaps via higher density, faster interconnects, and parallel computing.

> PCle Switch Demand Growth: Most supernode compute trays feature a high XPU-to-switch ratio of 2:1 (two PCIe switches for every four GPUs), driving meaningful demand for PCIe switches to handle increased density and complex I/O configurations within compute trays.

> Expansion of Edge AI and AIoT: Edge AI is expanding past smartphones and PCs into broader AIoT categories like AI glasses, toys, audio products, vehicle boxes, industrial terminals, and robots. Architectures combining host processors with edge AI coprocessors (such as Rockchip's dual-chip design) are raising semiconductor content value per device.

> Implications of Kimi K3: Moonshot AI's launch of Kimi K3 validates sustained infrastructure demand with its 2.8-trillion-parameter Mixture of Experts (MoE) architecture, native multimodal capabilities, and 1M-token context window.

> Increased Hardware & Memory Demands from K3: K3's scale increases high bandwidth memory (HBM) capacity needs (requiring about 1.4TB of raw model storage before caches and buffers), while its jump to 896 experts and 16 active experts per token places a heavy premium on memory bandwidth, efficient expert routing, and scale-up interconnect efficiency.

$Micron Tech(MU.US) $DRAM $EWY

Yole Group: CPO Supply Chain

PIC design

Nvidia $NVIDIA(NVDA.US)

Broadcom $Broadcom(AVGO.US)

Marvell $Marvell Tech(MRVL.US)

Lightmatter

Celestial AI

Ayar Labs

Enosemi

ASIC & xPUs design

Nvidia $NVIDIA(NVDA.US)

Broadcom $Broadcom(AVGO.US)

Marvell $Marvell Tech(MRVL.US)

Intel $Intel(INTC.US)

AMD $AMD(AMD.US)

MediaTek $2454.TW

Laser

Lumentum $Lumentum(LITE.US)

Broadcom $Broadcom(AVGO.US)

Coherent $Coherent Corp.(COHR.US)

Sivers Semiconductors $SIVERS SEMICONDUCTORS AB(SIVEF.US)

Scintil Photonics

Scape Photonics

Quintiquent

Foundries

TSMC $Taiwan Semiconductor(TSM.US)

SilTerra - subsidiary of Dagang NeXchange $4456.Kl

NSTIC

IMEC

Leti

ITRI

GlobalFoundries $GlobalFoundries(GFS.US)

SOI wafer & Epi-wafer

Soitec $SOI.PA

Shin-Etsu Chemical $4063.T

LandMark Optoelectronics $3081.TWO

IQE $IQE.L

Shanghai Simgui Technology - subsidiary of National

Silicon Industry Group $688126.SS

Sumitomo Electric $5802.T

VPEC $2455.TW

AXT $AXT(AXTI.US)

Packaging / Assembling / Testing

ASE Holdings $Advanced Semiconductor Engineering(ASX.US)

Foxconn $2317.TW

Fabrinet $Fabrinet(FN.US)

SPIL - subsidiary of ASE Holdings $Advanced Semiconductor Engineering(ASX.US)

ShunSin Technology $6451.TW

Amkor Technology $Amkor Tech(AMKR.US)

Jabil $Jabil(JBL.US)

JCET $600584.SS

Teradyne $Teradyne(TER.US)

FormFactor $Formfactor(FORM.US)

EIC, Retimers, SerDes, PHY

Marvell Technology $Marvell Tech(MRVL.US)

Credo Technology $Credo Tech(CRDO.US)

MACOM Technology $Macom Tech(MTSI.US)

Semtech $Semtech(SMTC.US)

Alphawave Semi - part of Qualcomm $Qualcomm(QCOM.US)

Astera Labs $ALLAB

System / Equipment

Nvidia $NVIDIA(NVDA.US)

Micas Networks

Ruijie Networks $301165.SZ

Cisco Systems $Cisco(CSCO.US)

Inventec $2356.TW

Dell Technologies $Dell Tech(DELL.US)

Hewlett Packard Enterprise $Hewlett Packard Enterprise(HPE.US)

Wistron $3231.TW

Wiwynn $6669.TW

Connectors & Fibers

Molex

Samtec

Corning $Corning(GLW.US)

Yangtze Optical Fibre and Cable (YOFC) $6869.HK

US Conec

TE Connectivity $TE Connectivity(TEL.US)

Sumitomo Electric $5802.T

Hengtong Optic-Electric $600487.SS

Luxshare ICT $002475.SZ

End customers

AWS $Amazon(AMZN.US)

Microsoft Azure $Microsoft(MSFT.US)

Meta $Meta Platforms(META.US)

Tesla $Tesla(TSLA.US)

Tencent $Tencent(TCEHY.US)

ByteDance

OpenAI

xAI

Non-exhaustive list.

Supply shortfall is a big deal in optical components for CPO, says a Marvell engineer.

"Demand is enormous."

For instance, this year’s highest shipment volume is for 800G optical modules, with demand exceeding 80 million units, yet only about 60 million units can be produced, which represents the total available production capacity.

The demand for 1.6T modules this year may exceed 30 million units, but only slightly over 20 million units are likely to be delivered ultimately.

Next year, the demand for 800G optical modules is projected to be around 120 million-130 million units. The situation may improve slightly as all Chinese manufacturers are expanding production capacity and are expected to gradually catch up with demand.

However, he states that price increases for these modules are not expected year after year.

BofA: US Semis

> Chinese Open LLMs Drive Memory Demand: Recent open-source and open-weight model releases from China (such as Moonshot AI’s Kimi K3) do not threaten memory demand; rather, they reinforce a bullish thesis for memory and Micron Technology (MU).

> API Pricing vs. Hardware Requirements: Aggressive Chinese API pricing (up to 5-350x cheaper than Western models) reflects business-model choices and market-share strategies rather than lower hardware costs. For example, Kimi K3 still requires roughly 1.4 TB of High Bandwidth Memory (HBM) per serving instance.

> Non-Hardware Cost Advantages: Chinese model providers achieve cost advantages through architectural efficiency (such as Mixture of Experts [MoE] sparsity, MLA/hybrid attention, and native quantization) and infrastructure efficiencies (lower local electricity, labor, and land costs, alongside potential state-linked subsidies).

> Open Models Expand Memory Sockets: Unlike closed models where millions of users share a single centralized HBM pool, open-weight models are downloaded and self-hosted across thousands of separate enterprise, cloud, and sovereign endpoints, multiplying global memory and storage sockets.

> Scaling Outpaces Compression: While compression techniques and active-parameter sparsity reduce compute intensity, total model parameter sizes continue to scale up drastically (e.g., Kimi K3 reaching 2.8 trillion parameters), driving continuous demand for multi-tier memory hierarchies (HBM, DRAM, and NAND).

> Micron Technology (MU) Outlook: BofA reiterates its Buy rating on Micron with a $1,550 price objective, citing limited competitive threat from China's CXMT (which focuses on underserved consumer DRAM rather than advanced HBM3E/HBM4) and upcoming catalysts like the expiration of CHIPS Act buyback restrictions in late 2026.

$Micron Tech(MU.US) $DRAM $EWY