Equity research

Global Market Size for Co-Packaged Optics (CPO)

> Market Growth: The global market is projected to surge rapidly from 18.4 billion yen (~$115.7 million) in 2025, reaching 1.27 trillion yen (~$7.99 billion) by 2030, and expanding to 3.55 trillion yen (~$22.33 billion) by 2035 (forecast).

> Scale-Up (Within Rack): Representing scale-up applications makes up 68% of the market (~$5.43 billion) in 2030 and increases to 67% (~$14.96 billion) by 2035.

> Scale-Out (Switch Layer): Representing scale-out applications accounts for 32% of the market (~$2.56 billion) in 2030 and 33% by 2035 (~$7.37 billion).

(h/t @semi_journal_jp)

We need more FAUs for CPO.

"Based on the estimated shipment of approximately 20,000 Spectrum-X units this year, the optical engines required would need 600,000 to 700,000 FAUs (Fiber Optic Active Units), and this number is expected to reach millions by 2027."

"Currently, the supply chain capacity cannot meet this demand."

$NVIDIA(NVDA.US)

Goldman Sachs: WFE

Overall Market Growth

> Significant Upward Revisions: Goldman Sachs has revised its total WFE forecasts upward, projecting $150B in 2026 (a 6% increase over old estimates) scaling up significantly to $281B by 2028 (+35% vs. old estimates).

> Strong Year-over-Year (YoY) Momentum: Total WFE year-over-year growth is expected to peak at +45% in 2027, following a robust +36% in 2026.

Segment Breakdown (DRAM & Foundry/Logic Lead)

> Foundry/Logic Drive: Foundry and Logic segments account for the largest share, with 2026 spending estimated at $92B (+9% revision) growing rapidly to $162B by 2028.

> DRAM Surge: DRAM is forecasted for explosive growth, moving from $48B in 2026 to $97B by 2028, reflecting persistent high demand with YoY growth rates staying above 35%.

> NAND Lagging: In contrast, NAND is experiencing downward near-term revisions (-4% to -$13% compared to older estimates), though it is still expected to recover toward $22B by 2028.

Regional Trends (China vs. Ex-China)

> Sustained China Demand: China continues to command a robust appetite for WFE investments, expected to contribute $44.3B (29% of total WFE) in 2026 and peaking around $60.7B by 2028.

> Ex-China Expansion: The rest of the world (Ex-China) accounts for the lion's share of spending, projected to accelerate from $106.0B in 2026 up to $220.2B (78% of the market) by 2028.

(Cred @knockouttrader)

$Applied Materials(AMAT.US) $Lam Research(LRCX.US) $ACM Research(ACMR.US)

Edgewater: Feedback suggests Nvidia likely signed a multi-year supply agreement for DRAM/HBM with SK Hynix and Micron. CY27 HBM pricing is not expected to be settled until late 2H26.

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

Zhongji Innolight's Interim Report covering optics:

> Silicon photonics penetration in the global datacom optical interconnect market is expected to rise further, from about 38% in 2025 to about 73% in 2030.

> Optical interconnects accounted for about 5% of global AI capex in 2025, and are expected to reach about 10% over the long term. The global datacom optical interconnect market is projected to grow from $19.4 billion in 2025 to $98.6 billion in 2030, a compound annual growth rate of 33.8% from 2026 to 2030.

Regarding CPO:

> CPO is described as currently viewed as one of the most highly integrated optical packaging solutions, assembling the switch ASIC chip and optical engine together on the same substrate.

> The report states CPO is still in the development stage, facing challenges around technology and ecosystem maturity, manufacturing consistency, cost, and maintainability.

> Over the long term, CPO will coexist with pluggable optical modules and other connectivity solutions (rather than fully replacing them).

$Lumentum(LITE.US) $Coherent Corp.(COHR.US) $Applied Optoelectronics(AAOI.US) $Broadcom(AVGO.US) $NVIDIA(NVDA.US)

Goldman Sachs: Optical Networking Supply Chain

Too detailed to fit into just one pic...

$NVIDIA(NVDA.US) $Tower Semicon(TSEM.US) $Broadcom(AVGO.US) $Marvell Tech(MRVL.US) $Lumentum(LITE.US) $Coherent Corp.(COHR.US) $AXT(AXTI.US)

Meritz Securities: CSPs flooding requests to secure HBM supply for 2028.

$Alphabet(GOOGL.US) $Microsoft(MSFT.US) $Meta Platforms(META.US) $Amazon(AMZN.US) $DRAM $Micron Tech(MU.US) $EWY

According to the Fed minutes, expectation of a rate cut is in early 2028.

"The median respondent to the Desk survey, by contrast, expected no change in the policy rate this year or the next but expected a rate cut in early 2028."

GF Securities: CPO/NPO Update

NVIDIA Spectrum-X CPO Scale-Out

Production Status: Spectrum-X Co-Packaged Optics (CPO) switches for scale-out networking have officially entered mass production.

Volume Forecasts:

2026E: 15,000 units

2027E: 100,000 units

Supply Chain Ramp: TSMC has expanded CPO inspection equipment capacity—specifically adding Insertion 2/3 capacity—alongside supply chain progress in Fiber Array Units (FAU), shuffle boxes, and system assembly.

Rubin Ultra Scale-Up Architecture & Optical Engines

Tray Design Options: Rubin Ultra is slated to adopt a 9-18-9 tray design, though mechanical challenges associated with the 0.75U height profile could drive a potential reversion to a 10-9-8 layout.

Design Impact: Architectural shifts in tray layout will have no impact on the optical engine (OE).

NVIDIA Platform OE Shipments:

2027E: 6 million units

2028E: 19 million units

Total Industry-Wide OE Shipments:

2027E: 11 million units

2028E: 40 million units

Amazon Trainium 4 & NPO Deployment

AWS Consumption Projections:

2H27E: 5 million OE units

2028E: 12 million OE units (primarily 6.4T specifications)

Configuration & Packaging: Trainium 4 is projected to feature 3 distinct configurations, with two configurations expected to adopt Near-Package Optics (NPO).

$Amazon(AMZN.US) $NVIDIA(NVDA.US)

WSJ: Nine top tech companies had some $3 trillion of off-balance-sheet commitments mostly related to AI.

A lot of spending going into this buildout...

$Meta Platforms(META.US) $Amazon(AMZN.US) $Microsoft(MSFT.US) $Alphabet(GOOGL.US)

J.P. Morgan: DRAM & NAND bit demand vs bit supply

DRAM bit supply/demand to be worse in 2027 before balancing in 2028, same for NAND.

Note: this is for bits. Does not mean shortage is solved in 2028.

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

Impressive...but at the same time hard to believe that by 2029, the market share for GPUs will be ~21% compared to ~79% for ASICs. Also implying a 17.5% market share for Nvidia lol.

I don't buy that a single bit.

$NVIDIA(NVDA.US) $AMD(AMD.US) $Intel(INTC.US) $Alphabet(GOOGL.US) $Amazon(AMZN.US) $Microsoft(MSFT.US)

Mizuho: InP & Memory

Supply Chain & Gating Factors

> InP Laser Supply: Remains tight into 2027E as multiple suppliers work to bring capacity online.

> NPO & CPO Dynamics: NPO continues to act as a near-term bridge for CPO, which is expected to ramp in 2H27/28E.

> CRDO Position: Believed to be the leader in Active Electrical Cables (AEC), with market ramping at a 98% 2024–29E CAGR, driven by additional upside from ZFO, ALC/MicroLED, and SiPho ramps.

> LITE Commentary: 1.6T ramps feature 200G accounting for >25% of revenues, with CPO delayed to 2H27. CRDO benefits from this alongside Neocloud strength for AEC.

Memory & Storage Pricing (NAND/DRAM)

> ASP Spikes: NAND/DRAM Average Selling Prices (ASPs) are projected up 406%/421% y/y in 2026E.

> Growth Moderation: Growth slows down in 2027E, with NAND expected to be flat and DRAM up 52% y/y.

AI DC Short-Reach Copper Cable Split

> Market Expansion: Total units grow steadily from 8M in 2024 to 23M by 2029.

> AEC Adoption: AEC units increase from 0M (2024) to 12M (2027) and 12M (2029), with total AEC penetration (including optical %) scaling to over 40% by 2029.

Estimated Wholesale 1.6T Cable Prices ($)

AEC: $225

ALC: $400

EML-based Transceiver: $700

CPO: $150

$Credo Tech(CRDO.US) $Lumentum(LITE.US) $Coherent Corp.(COHR.US) $Micron Tech(MU.US) $DDRAM $Sandisk(SNDK.US) $EWY

BofA: AI Compute

> Total Market Size: The total Data Center Systems TAM is projected to reach $2.2Tn in 2030E.

> AI vs. Non-AI Split:

AI Data Center Systems TAM accounts for the vast majority at $1.8Tn.

Non-AI DC Systems TAM accounts for $384bn (with Trad'l Cloud CPU representing ~$30bn of that value).

> AI Servers & Infrastructure Breakdown: Within the $1.8Tn AI Data Center TAM, AI Servers make up $1.4Tn, alongside $310bn for AI Networking and $85bn for Storage.

> AI Server Component Breakdown ($1.4Tn Total):

AI Accelerators: $1.1Tn

HBM (High-Bandwidth Memory): $277bn

AI CPUs: $180bn total value, which splits roughly 50/50 between AI Cluster Compute/Head Nodes ($90bn) and AI Agentic Standalone Nodes ($90bn).

Other: $53bn

$NVIDIA(NVDA.US) $AMD(AMD.US) $Alphabet - C(GOOG.US) $Amazon(AMZN.US) $Meta Platforms(META.US) $Dell Tech(DELL.US) $Super Micro Computer(SMCI.US)

Missed out on the space? No problem. Here is the recording.

We covered:

> Transition from copper to optical

> CPO scale-up timeline

> InP dependency on China

> The potential bottleneck in the InP supply chain

J.P. Morgan on memory LTAs: "A relatively misunderstood element within LTA is the resilience of pricing. Majority of the LTA discussion is currently centered around server memory. CSP and AI related memory demand is likely to be the majority (over 70% of total bit demand and 85+% of total revenue) of the contract. Server memory solutions carry a significant price premium over non-server applications and we believe the LTA mix increase will be pricing/margin accretive from a mid-to-long term horizon."

$Micron Tech(MU.US) $DRAM $EWY $Sandisk(SNDK.US) $SK Hynix(SKHY.US)

Memory Stanley: Memory

> Nature of Inflation: Shifted from viewing memory inflation as cyclical to recognizing it as structural.

> Demand Dynamics: Moved from the belief that higher prices cannibalize demand to realizing that higher prices and supply scarcity are actually accelerating demand.

> OEM Pricing Power: Changed from assuming OEMs absorb higher costs to seeing that OEMs are successfully passing through pricing and margin stacking.

> Supply Outlook: Replaced the expectation that supply would eventually catch up with the outlook that supply will remain constrained for years.

> Market Drivers: Expanded from a simple refresh cycle view to a combination of refresh, capacity expansion, and pull-forward demand.

> Overall Market Impact: Shifted from viewing memory inflation as universally bearish for hardware to understanding that memory inflation creates winners and losers.

$Dell Tech(DELL.US) $Hewlett Packard Enterprise(HPE.US) $Super Micro Computer(SMCI.US) $Micron Tech(MU.US) $EWY $DRAM $SK Hynix(SKHY.US)

J.P. Morgan: AI Accelerators

> Rapid Overall Market Expansion: Total annual AI accelerator unit shipments are projected to grow dramatically from 3.3 million units in 2023 to 23.3 million units by 2027E.

> NVIDIA's Continued Dominance: NVIDIA GPUs lead the market significantly throughout the period, growing from 1.7 million units in 2023 to an estimated 9.9 million units in 2027E, though their share of total units shifts as competition scales.

> Surge in Custom/Alternative Accelerators (ASIC/XPUs): Alternative hardware lines—including Google TPUs, AWS Trainium/Inferentia, Meta MTIA, and Microsoft MAIA—are scaling rapidly. By 2027E, custom ASICs/XPUs are projected to make up 53% of total unit shipments, surpassing traditional GPUs (47%).

> Significant Short-Term Growth Spikes: The market saw triple-digit year-over-year growth in 2024 at 128%, with strong continued expansion expected through 2026E (62%) and 2027E (43%).

$NVIDIA(NVDA.US) $Marvell Tech(MRVL.US) $Amazon(AMZN.US) $Alphabet(GOOGL.US) $Meta Platforms(META.US) $Microsoft(MSFT.US) $Broadcom(AVGO.US)

BofA: Nvidia's $500B Financing

> NVDA Underwrites Value, Not Debt: NVDA has signed Memorandums of Understanding (MOUs) with six top financiers (Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR) to mobilize over $500 billion in third-party capital via independent platforms. The financial burden sits with the consortium rather than NVDA's balance sheet.

> GPUs as an Investable Asset Class: With $500 billion of capital treating compute as an investable asset class, residual values must hold. NVDA supplies fungible and transferable compute across operators, and CUDA extends its useful life to keep resell/rental rates high and depreciation curves benign.

> Extending the AI Buildout Runway: Funding—rather than demand—has been the primary bottleneck. A $500 billion pool allows non-investment-grade buyers (such as labs, neoclouds, and sovereigns) to secure hardware at attractive rates, de-risking offtake and supporting long-term AI systems Total Addressable Market (TAM) paths.

> Key Debates and Risks Monitored: Notable cautions remain, including the fact that MOUs are not deployed capital (requiring real end-customers paying real money), potential power and regulatory pushback, input-cost inflation, and the opacity/complexity that new financing structures might introduce to AI buildouts.

Morgan Stanley: Semiconductor Sales

"Our CY26 memory forecast at $981bn is higher than our prior forecast and up meaningfully from $223bn last year. For CY27, we are assuming 29% growth y/y to $2.22 trillion, mostly due to memory dynamics."

$Micron Tech(MU.US) $DRAM $EWY $Sandisk(SNDK.US) $SK Hynix(SKHY.US)

Sumit Sadana, Chief Business Officer of Micron, at KeyBanc Technology Leadership Forum:

> "All of the efforts we are doing, we still don't have line of sight when supply is going to meet demand. Demand continues to escalate at a very rapid pace over time."

> "Very early stages of how this demand is going to proliferate."

> "HBM is growing significantly."

> "To produce 100 bits of HBM, we have to reduce 300 bits of DDR supply, 3:1 ratio. Closer to 4:1 for HBM4."

> "The number one constraint our customers tell us they have is DRAM."

> "We expect very robust business and financial performance for a very long time."

> "Aggregate demand across all of the segments is at much higher level than it has been."

> "Customers are telling us that despite the fact the prices are at a very high level, they are eager to get more supply because they are not able to meet their own business case requirements."

> "In data center, we are not able to meet any more than half of the demand our customers have."

> "Multi year signals on demand, not just 2026."

> "We have visibility on the engineering side that stretches out to roadmaps of our customers that are beyond 2030 timeframe."

> "Increased level of anxiety of our customers to secure supply."

> "We have signed the most number of SCAs across our industry."

> "Since our earnings, we have signed more SCA agreements with our customers."

> "These SCAs cover a long time horizon. Some customers that are smaller like automotive have 3 year SCAs. The SCAs that cover overwhelming amount of revenue under SCA is going to be 5 year type of terms through the end of 2030."

> "SCAs are binding commitments on purchases of volumes by year, by customer. Take or pay agreements with no contractual outs. Backed up by tremendous amounts of upfront cash like commitments."

> "SCAs are something LTAs never had. Some SCAs are going to be floating pricing, so pricing is going to be consistent with what the market prices at that time. Most of the volume under SCAs is going to have a ceiling price. The ceiling price at time of earnings with 16 SCAs is Q2 pricing. Since then, we have signed more SCAs. SCAs signed in the future is going to have ceiling pricing that is going to be consistent with whatever the market prices at the time the SCAs are signed in the future."

> "Some SCAs have no pricing mechanics and focus on market pricing."

> "The floor pricing is set at a level that provides a gross margin for us well above any prior peak in the cycles of the industry. These provide very high ROI."

> "SCAs have tremendous number of binding terms and no contractual way of backing out. More years can be added to the backend of the SCAs to extend them."

> "Huge amount of growth coming from robotics, which is going to be a ginormous growth driver."

> "This is not going to be the last time you see significant deficit of supply in this industry."

> "Physical AI in its infancy right now and has tremendous potential."

> "Lack of adequate data to train robots."

> "Each single humanoid robots is expected to have hundreds of GB of DRAM and TBs of SSD."

> "Customers struggling to get hands on adequate DRAM supply."

> "Availability of supply leading to despec among customers."

$Micron Tech(MU.US) $DRAM $EWY $Sandisk(SNDK.US) $SK Hynix(SKHY.US)

UBS: Micron & Memory

> BM Pricing & Premiums: Industry blended HBM ASP is up approximately 79% year-over-year, with HBM4E likely reaching >$30/GB.

> DRAM & NAND Contract Pricing:

For DDR, industry C3Q contract pricing (including LTAs) is modeled up 20% Q/Q, with MU specifically modeled up 20% Q/Q in C3Q and 11% Q/Q in C4Q.

For NAND, industry C3Q contract pricing is modeled up 20% Q/Q (down from 28% prior), and for MU specifically up 23% Q/Q (down from 35% prior).

Market Dynamics

> DRAM Tightness: HBM supply and demand have tightened beyond prior expectations, specifically for HBM4 and HBM4E, led by persistent supply tightness.

> HBM Consumption & Adjustments: C2027 HBM consumption is increasing from 58.7B Gb to 61.5B Gb, and MU HBM ASP is now modeled up 72% year-over-year (vs. +61% prior) with broadly unchanged bit shipments of ~11.65B Gb.

> NAND & SSD Strength: Stronger-than-expected server/storage SSD demand is offsetting weakness in PCs and smartphones, keeping the demand backdrop constructive.

> Competitor Impact: CXMT's DRAM bit supply share is expected to rise from ~7% in C2025 to ~9% in C2027, though capacity expansions are viewed as somewhat overstated due to yield and technology gaps. Meanwhile, YMTC is shifting incremental capacity toward DRAM over NAND, extending the NAND pricing cycle.

$Micron Tech(MU.US) $DRAM $EWY $SK Hynix(SKHY.US) $Sandisk(SNDK.US)

Nomura Securities: Global optical laser chip eco-system and competitive landscape

InP Wafer

AXT $AXT(AXTI.US)

Sumitomo Electric $5802.T

JX Advanced Metals $5016.T

Coherent $Coherent Corp.(COHR.US)

Lumentum $Lumentum(LITE.US)

Epi Wafer

IQE $IQE.L

Landmark Optoelectronics $3081.TW

VPEC $2455.TW

Coherent $Coherent Corp.(COHR.US)

Lumentum $Lumentum(LITE.US)

Sumitomo Electric $5802.T

Wafer Equipment

Aixtron $AIXA.DE

Veeco $Veeco Instruments(VECO.US)

Taiyo Nippon Sanso $4091.T

CW/EML Laser

Lumentum $Lumentum(LITE.US)

Coherent $Coherent Corp.(COHR.US)

Sumitomo Electric $5802.T

Mitsubishi Electric $6503.T

SiPho Foundry

Tower Semiconductor $Tower Semicon(TSEM.US)

GlobalFoundries $GlobalFoundries(GFS.US)

TSMC $Taiwan Semiconductor(TSM.US)

Chip Equipment

FormFactor $Formfactor(FORM.US)

Keysight Technologies $Keysight Tech(KEYS.US)

Anritsu $6754.T

UBS: Nvidia Vera Rubin BoM

Core Financial Breakdown (UBS BOM Analysis)

> Total Superchip Cost: $38,902 for the Vera Rubin core unit (1 Vera CPU + 2 Rubin GPUs).

> Total Memory Share: $24,297 (62%) goes entirely to memory semiconductors (HBM4 and SOCAMM2).

> CPU vs. GPU Memory Split:

SOCAMM2 (Vera CPU): $19,355 (49.8% of total cost) — a massive cost driver.

HBM4 (Rubin GPUs): $4,943 (12.7% of total cost).

> Memory Dominance in Costs: Memory semiconductors account for approximately 62% of the total cost of the Vera Rubin superchip. Out of a total cost of $38,902 per unit, memory components—specifically HBM4 and SOCAMM2—make up $24,297.

> Comparison to Previous Generations: In the previous Blackwell Ultra-based GB300, memory accounted for 53% of the cost; this has jumped to 62% for Vera Rubin. While the total cost of the superchip increased roughly 2.1 times compared to its predecessor, the cost specifically attributed to memory rose 2.5 times.

> Infrastructure Scale: The scale of the "Vera Rubin NVL72" system is immense. A single server rack contains 74.7 TB of DRAM. To put this in perspective, the LPDDR5X capacity in just one rack is roughly equivalent to the memory found in approximately 4,500 premium smartphones.

$NVIDIA(NVDA.US)

Third Bridge experts on CPO:

> Despite the hype, mass commercial adoption is not expected before the second half of 2027. Traditional pluggable optical modules are expected to remain the industry standard throughout 2026.

> CPO is likely to see early adoption in "scale-out" networking (connecting switches, GPUs, and CPUs) within distances under 500 meters. Deployment into GPUs themselves and "scale-up" chip-level systems is a harder engineering challenge, likely not seeing broad adoption until 2030–2032.

> Experts note that training workloads (which require extreme bandwidth) are better suited for CPO, while inference workloads are deterministic and remain well-supported by traditional pluggable modules.

> A primary hurdle for CPO is maintenance. With traditional pluggable modules, a failed component can be hot-swapped. With CPO, a failure in the optical subsystem might require replacing an entire board, leading to significantly higher costs and operational disruption.

> Nvidia and Broadcom are expected to dominate the CPO market, potentially controlling 70% of it by 2028. Nvidia is expected to capture around 50% market share through its leadership in AI infrastructure, while Broadcom could secure roughly 20% through continued expansion of its deployment footprint.

> Large cloud providers (like Google, Meta, AWS, Microsoft) are cautious about CPO because it risks "vendor lock-in." Pluggable modules allow these companies to mix and match components from various vendors, whereas CPO ecosystems are more restrictive.

> Through substantial investments in optical component companies like Coherent and Lumentum, Nvidia is positioning itself to control the entire AI value chain—from GPUs and networking switches to the optical modules themselves.

> Experts expect pluggable optical modules to remain the dominant solution for cloud giants such as Microsoft and Amazon throughout 2026. CPO adoption is expected to begin in earnest during 2027, reaching approximately 20% to 30% penetration by 2028.