In-Depth Review of the CPO Delay Controversy: Why a Single SemiAnalysis Report Slammed US Optical Communications in a Single Day, and Should We Really Believe the 'Delay'?

On June 9, 2026, a research report released exclusively to paying institutional clients triggered a single-day collective plunge in the US optical communications sector. The author was neither Goldman Sachs nor Morgan Stanley, but SemiAnalysis, an independent research firm that started as a newsletter. Using this report as an entry point, this article systematically dissects whether the large-scale deployment of Co-Packaged Optics (CPO) has indeed been delayed, why the market collectively bought into a highly controversial static yield model, and how investors should establish an analytical framework free from emotional influence when facing events like an "authoritative report triggering a market crash."

1. Event Recap: How an Institutional-Exclusive Report Wiped Out Tens of Billions in Market Value

Let us first reconstruct the pricing facts of that day. On June 9, 2026, US optical communications stocks suffered a collective single-day decline: Applied Optoelectronics (AAOI) plunged about 17%, POET Technologies (POET) dropped about 12%, Coherent (COHR) fell about 11%, Lumentum (LITE) declined about 8%, Marvell (MRVL) shed about 7.6%, and Ciena (CIEN) also recorded a slide of around 7%. In addition, other optical interconnect chain players such as Fabrinet (FN) and Credo (CRDO) were also under pressure.

Before diving into the technical debate, one detail is worth noting beforehand: under the exact same negative news, AAOI plummeted about 17% in a single day, while Lumentum, which has highly overlapping technology, fell only about 8%—a nearly twofold difference in decline. This structural divergence in itself reveals the nature of the capital behind that day's panic selling: it resembled a shakeout of weak hands rather than a calm repricing of fundamentals. We will revisit this clue in the "Reflexivity" section at the end of this article.

The core thesis of the report boils down to a single statement: the large-scale commercialization of CPO needs to be delayed. This leads to the two main questions this article aims to answer: first, how can an independent research firm that started as a newsletter wipe out tens of billions of dollars in market value from US optical communication stocks in a single day with just one report? Second, should we believe its definitive claim that the large-scale deployment of CPO will be delayed?

2. Technical Background: What is CPO, and Why Does It Affect the Entire Optical Communications Value Chain?

To evaluate the weight of this "delay," one must first understand CPO's position in AI data centers. Traditional data centers rely on pluggable optical transceivers to perform electro-optical conversion between switches. As AI clusters expand from thousands of GPUs to hundreds of thousands or even millions, the power consumption, latency, and reliability of interconnects are becoming new system bottlenecks. The idea behind Co-Packaged Optics is to package silicon photonic engines directly with the switch chip (ASIC) on the same substrate, shrinking signal paths from "inch-level" to "millimeter-level" and eliminating bulky DSP re-timers, thereby achieving a quantum leap in power efficiency and density.

This is not just a theoretical concept. According to NVIDIA's official announcement at GTC 2025 (March 18, 2025), the company introduced two co-packaged optics switching platforms: Spectrum-X Photonics for Ethernet and Quantum-X Photonics for InfiniBand, with port speeds reaching 1.6 Tb/s. NVIDIA's official comparison data showed that, compared to traditional pluggable solutions, CPO can achieve an approximately 3.5x improvement in energy efficiency, a 63x enhancement in signal integrity, 10x network resilience, and a 1.3x faster deployment speed, while requiring about three-quarters fewer lasers (officially stated as "4x fewer lasers," with some subsequent materials revising the energy efficiency gain upward to about 5x). Founder Jensen Huang positioned this as a crucial step to integrate silicon photonics directly into the switch, breaking old limits of hyperscale networking, and viewed it as a vital link to million-GPU AI factories.

The underlying manufacturing of this solution falls on TSMC. TSMC Chairman C.C. Wei emphasized in the same event that their silicon photonics solution combines advanced node processes with SoIC 3D stacking capabilities. Specifically, the Quantum-X800 switch ASIC uses TSMC's 4N process, integrating approximately 107 billion transistors with a throughput of 28.8 Tb/s. Its silicon photonic engine is based on TSMC's COUPE (Compact Universal Photonic Engine) platform, utilizing 200G micro-ring modulators (MRM) and continuous-wave distributed feedback (CW-DFB) lasers as light sources. It is worth noting that the repeatedly mentioned "COUPE" is precisely the name of TSMC's photonic engine packaging platform—understanding this provides a concrete engineering target for the "yield equation" discussed later.

Equally critical is the supply chain list. According to NVIDIA's official statements, the co-inventors and collaborators of its CPO platform include TSMC, Coherent, Corning, Foxconn, Lumentum, and SENKO. In the parallel pluggable optical module camp, industrial partners named by NVIDIA include Coherent, Eoptolink, Fabrinet, and Innolight. In other words, many of the stocks that tumbled in a single day are precisely the core members NVIDIA integrated into its optical ecosystem—which explains why a report on the adoption timeline could emotionally shake up the entire value chain. It should be added that CPO is not a narrative unique to NVIDIA; chipmakers like Broadcom are also positioned in this race, and the industry has always had divided opinions on "when and in what form volume shipments will scale."

3. Report Claims: Three Downward-Revised Timelines and a Yield Equation That Triggered the Panic

Returning to the report itself, titled *Powered Down, Lights Off*. It revised three technical timelines downward: first, CPO shipments for scale-out networking in 2027 will be significantly lower than the aggressive market expectations; second, the timeframe for full-scale mass production will be pushed back to 2028 or even 2029; third, the 800VDC power delivery architecture originally planned by NVIDIA will also be delayed until 2028.

It must be objectively pointed out that the firm did not blindly turn bearish across the board. The report clearly stated that ±400VDC technology remains on track and will see volume ramp-up in 2026; even some Near-Packaged Optics (NPO) projects might accelerate due to this delay. This "qualified bearishness" is precisely what distinguishes it from pure doom-mongering.

The report offered three core arguments: manufacturing yields, system integration complexity, and overall cost-efficiency. What panicked the market most was a mathematical equation regarding yield. The report assumed a 95% yield for assembly of a single optical engine. Under the new switching architecture it described, a single switch ASIC needs to package 32 optical engines (COUPE). Multiplying 0.95 by itself 32 times (0.95 to the 32nd power) yields a system-level yield of only about 19.4%. This means for every five units built on the production line, only about one is qualified. More critically, these optical engines are directly soldered onto the substrate with no rework path—if any single engine exhibits defects after coupling, the entire expensive substrate must be scrapped.

In addition to yield, the report presented empirical evidence of system integration issues: the relevant switching platforms reportedly showed insertion loss exceeding 3.5 dB in system-level testing, almost entirely exhausting the tolerable limit of the optical channel budget. Moreover, they were not alone in taking a cautious stance. Morgan Stanley (Greater China Semiconductor Team) partially echoed this in a June 10 report: it agreed that near-term CPO scaling is tracking below expectations, projecting 2027 global optical engine shipments of only 6 million to 7 million units, far below the prior market consensus of 20 million to 30 million units. However, it must be emphasized that Morgan Stanley did not fully align with the bear case—it explicitly refuted the judgment that "800V mass production would be delayed to 2028," noting that supply chain checks showed 800VDC racks remain on track to progress in the second half of 2026 as planned, and maintained an "Overweight" rating on CPO, expecting real explosive growth to kick off starting in 2028 with the long-term thesis intact. In other words, even within the bearish camp, there is clear disagreement over the scope and extent of the "delay."

4. Three Timelines: Unpacking the "Delay," Both Bulls and Bears Have a Point

Facing these seemingly irrefutable engineering metrics, a more analytical approach is to dissect the "delay" into three distinct timelines for separate examination, rather than accepting or dismissing them wholesale.

The first layer is small-scale validation and introduction in the second half of 2026. On this layer, there is no real conflict between bulls and bears, and even original manufacturers have stepped up to push back. On the day the report was released, NVIDIA immediately issued a clarification: Gilad Shainer, Senior Vice President of Networking, stated publicly during GTC/Computex in Taipei that CPO is currently the most exciting technology and explicitly noted that related products are already shipping and will begin ramping up in the second half of the year, directly contradicting the report's tone. To clarify, the official "ramp-up" refers to small-scale physical deployment and validation for a few tier-one cloud customers, a step that has not been delayed.

The second layer is the full scale-out ramp-up in 2027. On this layer, SemiAnalysis's warnings do hold merit. The market had previously embraced an overly optimistic narrative that 2026 would be the first year of commercial CPO, and that it would be fully deployed across hyperscale data centers within two years. Such an overly compressed adoption timeline was indeed too aggressive, and shipment expectations for 2027 needed a downward revision—which is the same layer of concern targeted by Morgan Stanley's estimate of "only 6 million to 7 million units in 2027."

The third layer is the full scale-up architectural transition after 2028, when CPO is loaded directly onto next-generation platforms like Rubin Ultra and Kyber, similar to the Sidecar solution. The window originally projected for this process was always around 2028, rather than something expected to happen in 2026 or 2027. It is worth adding that NVIDIA's next-generation Rubin platform will increase the number of GPUs per rack from 72 to 144 or even 576, raising the intra-rack interconnect bandwidth to hundreds of terabytes per second. This quantum leap in density is exactly what transforms CPO from a "nice-to-have" to a "must-have." Therefore, taking this reasonable technical transition window and exaggerating it as the "death" of CPO's overall technical roadmap is clearly an overinterpretation driven by market panic. Looking at the three layers together, the conclusion is far clearer: what was reasonably corrected is only the middle layer's aggressive 2027 curve, not the entire technical pathway.

5. Two Counterarguments: Methodological Flaws in the Yield Equation and Hard Metrics from Upstream Laser Orders

To balance the perspective, we must introduce two heavyweight counter-arguments—which marks the fundamental difference between rigorous analysis and mere regurgitation of reports.

The first counter-argument points to the methodological flaw within the yield equation itself. Raising 0.95 to the 32nd power essentially freezes the initial yield at a specific, pessimistic point in time, assuming it will never improve during future mass production. This calculation completely ignores the semiconductor industry's core mechanisms of chip binning, grading, redundant designs, and the steep learning curve that rises rapidly alongside volume production. In fact, continuous yield optimization throughout mass production is the norm, not the exception, in semiconductor manufacturing. In response, opposing research firm Global Semi Research published a dedicated article titled *Co-Packaged Optics Is Not Delayed* to directly refute the claim, arguing that the report mistook a conservative engineering model for a static, non-evolving conclusion.

The second counter-argument comes from hard upstream capacity data, which is the most difficult metric to reconcile with a "delay to 2029." Global Semi Research pointed out in the same rebuttal that no manufacturer would dramatically raise its high-power laser order guidance from around 40 million units to approximately 100 million units—and book out Lumentum's entire final assembly line—for a technology that is allegedly "significantly below prior expectations." It should be clarified that these two figures are claims made by Global Semi Research and, as of publication, have not been officially confirmed by NVIDIA or Lumentum; they should be treated as strong supporting evidence from the opposing side rather than established official figures. However, orders represent capital expenditure commitments, which are far more binding than verbal guidance. In the commercial logic of the industry, almost no tech giant would spend real cash years in advance to fully book an entire upstream laser production line for a technology that would not deploy until 2029.

Synthesizing these two counterarguments yields a conclusion free from emotional distortion: if "delayed large-scale commercialization" refers to a downward revision of the overly aggressive 2027 ramp curve that the market had priced in, that is indeed an ongoing industrial reality. However, if the report is interpreted as meaning that the entire CPO technical roadmap will slip completely to 2029, the evidence of major players aggressively securing laser capacity via orders clearly does not support such a bearish conclusion.

6. SemiAnalysis and Dylan Patel: From Beekeeper to the Intelligence Firm Wall Street Fears Most

Since the 19% yield equation has obvious methodological flaws—whether judged by the public statements of chipmakers' executives or the hard data of "100 million lasers fully booked upstream"—a deeper question arises: why did the market collectively buy into a highly controversial static mathematical model, wiping out tens of billions of dollars in market value in a single day?

The answer lies in the mastermind behind the report: Dylan Patel, currently one of the most formidable intelligence operators in the semiconductor supply chain. His rise to prominence is quite legendary: holding no semiconductor-related degree, he worked as a rural beekeeper in his early years, later publishing in-depth analyses via an anonymous semiconductor Twitter account before founding SemiAnalysis in 2020. Step by step, he grew into a star analyst frequently cited by major AI labs, cloud giants, top hedge funds, and semiconductor leaders. His massive influence is evidenced by a key detail: at the GTC conference in March 2026, Jensen Huang named only two individuals throughout the entire event, one of whom was Dylan Patel, and displayed SemiAnalysis's InferenceMAX chip performance benchmark (reported as InferenceX by some media) on the main screen. AMD CEO Lisa Su also arranged a dedicated, roughly 90-minute face-to-face meeting with him.

In terms of scale, SemiAnalysis has long held the top spot on Substack's technology subscription rankings. Its total subscriptions across different metrics range from approximately 200,000 to over 280,000 (officially stated as "over 200,000," with third-party reports citing "over 250,000," and Substack's platform count showing nearly 290,000), most of whom are free readers, with a minority of paying annual subscribers. Its revenue illustrates its commercial scale even better: market estimates suggest its annual revenue has jumped from about $20 million in 2025 and is on track to break $100 million in 2026.

But what truly forms its moat is not subscription counts, but its intelligence-gathering methodology. Rather than reverse-engineering public financial reports, SemiAnalysis builds on solid physical supply chain intelligence: counting the construction progress of tech giants' data centers via satellite imagery, pinpointing the most critical capacity bottlenecks in wafer fabs, tracking underlying order shifts for laser components and optical transceivers, and even scanning code commits on GitHub to infer which technical paths companies are accelerating. It is this granular, "particle-level" intelligence network that makes institutions willing to pay premium fees. To locate this firm accurately, rather than calling it a traditional sell-side research house, it is better described as a supply chain intelligence company hiding behind the market.

Yet, highly detailed intelligence does not guarantee that macro-level projections about the future are correct. Before evaluating the credibility of this report, we must first understand its business model and clarify exactly who its target audience is.

7. Behind the Paywall: Whose Interests is This Report Serving?

SemiAnalysis's primary revenue comes from institutional-grade subscription data, customized advisory services, and direct sales of primary research to the top four cloud giants, leading AI labs, and major hedge funds. In contrast, public Substack subscription fees represent only a fraction of its income. In other words, its true clientele consists almost entirely of the core players who wield the greatest influence in this supply chain.

This leads to the most critical fact of this event: the CPO report that triggered the plunge in optical communication stocks, titled *Powered Down, Lights Off*, was a paid report restricted to institutional clients and not made public. Understanding this is essential. A transaction alert provided only to paying institutions, which inherently serves to guide position adjustments, has a very different writing motive than a purely neutral technical assessment. The purpose of institutional reports is often to provide an explicit trading catalyst or rebalancing excuse for capital already playing the game. Consequently, this report sparked discussions about "information asymmetry"—institutional clients could read the complete findings, while the public and retail investors received only a simplified, or even distorted, version through secondary sources, chasing rallies and dumping shares accordingly. Thus, when reading such research, the first question to ask is always: for whom was this report written?

It is worth mentioning that the firm also has clear preferences on public platforms—it once joked that it was running a long-term promotional campaign of sorts for Anthropic's Claude model. While this does not constitute an ethical flaw, it should be kept in mind as a background weight when referencing any of its judgments. It must be emphasized that discussing these interest structures is by no means indulging in conspiracy theories; understanding who paid and whom the report serves is a fundamental skill in reading any sell-side research, representing a rational exercise of due diligence rather than a malicious attack on the research firm.

8. Shattering the Myth: Accurate Scoops Do Not Equal Accurate Predictions

To evaluate this firm more rationally, we can deconstruct its past representative works into three distinct levels of capability.

The first is its ability to reconstruct information and break exclusive scoops. In May 2023, the sensational internal Google memo titled "We Have No Moat, and Neither Does OpenAI" was first published by SemiAnalysis, which stated it had verified the document's authenticity; its author was later identified by Bloomberg as Google senior engineer Luke Sernau. Regarding the underlying architecture of GPT-4—including key details such as its use of the Mixture of Experts (MoE) model and its specific parameter size—it was also SemiAnalysis that first uncovered and detailed them. During the DeepSeek event that recently took the market by storm, rumors circulated that the model's training cost was only about $6 million. It was SemiAnalysis that wrote an article correcting this skewed figure to a more realistic capital structure. It pointed out that the total server capital expenditure behind DeepSeek was approximately $1.6 billion (with about $944 million in cluster operating costs; some early media outlets reported this as around $1.3 billion, but the order of magnitude is clear). It also revealed that DeepSeek holds about 50,000 Hopper-class GPUs, mixing H800s (around 10,000), H100s (around 10,000), and a large volume of China-specific H20s, rather than the "50,000 pure H100s" rumored externally. The $6 million represented merely the on-paper GPU compute cost for a single pre-training run of DeepSeek-V3, excluding R&D, infrastructure, and hardware total cost of ownership (TCO). These cases are sufficient to prove that the firm is indeed world-class at mining internal corporate intelligence and dissecting supply chains.

The second is its predictive capability regarding the future. For example, SemiAnalysis began publicly warning of a comprehensive rise in memory chip (DRAM/HBM) prices from the end of 2024, and subsequent market trends validated this projection. Its InferenceMAX quantization work on model deployment costs also established an important benchmark for AI inference economics. This kind of extrapolation of industry trends represents "prediction" in the true sense and must be evaluated separately from its "scoop-breaking" ability.

The third is the ability to directly move markets and influence stock prices. The plunge triggered by this CPO report is a prime example. Prior to this, its findings regarding the Vera Rubin platform and DRAM usage also sparked a similar "fire-drill-style" sell-off in the memory sector. In other words, every report this firm publishes now has the potential to become an independent market event.

From this, a key conclusion can be distilled: having access to exclusive scoops does not make one a prophet for future predictions. The most common mistake investors make is equating the firm's exclusive supply chain intelligence directly with a precise prophecy of technology adoption speed. In reality, when it comes to forecasting when a new technology will scale up in volume and how steep its growth slope will be, SemiAnalysis faces revisions and errors just like any other market analyst. The aforementioned static yield model of 0.95 to the 32nd power (yielding only about 19%) is the perfect counterexample: freezing a flawed yield and ignoring the learning curve is fundamentally different from the scoop-breaking ability displayed in digging up DeepSeek's true capital expenditures. "Being able to unearth scoops others cannot" and "being able to calculate future scaling slopes" are two distinct skill sets that must never be conflated.

9. Stock-Specific Divergence: The Same Report Is Not Bearish Across the Board

Looking closer, the report actually performed a refined differentiation of specific stocks, which further proves that it was not blindly bearish on the entire optical communications industry. In the report, SemiAnalysis explicitly named Innolight (300308.SZ) and Eoptolink (300502.SZ), the two Chinese optical module giants, as beneficiaries of this delay. The logic is straightforward: as the timing for large-scale CPO deployment shifts outward, it extends the demand window for traditional pluggable optical modules and NPO architectures. Technical delays do not mean demand vanishes into thin air; rather, demand is reallocated across the timeline.

Meanwhile, the report was relatively bullish on infrastructure-related plays, naming Amphenol, Vertiv, and Legrand, while taking a more cautious stance on Lumentum, Himax, Navitas, and Wolfspeed. Juxtaposing the beneficiaries with those under pressure reveals that the report was genuinely targeting the timing of a specific layer and particular names, rather than refuting the overall demand for optical communications.

10. The Truth of the Crash: Validated Technology, or a Believed Report?

Once the stock-specific divergence is laid out, the true power of institutional reports becomes apparent, raising a new question: the firm's influence has now grown so large that the report itself has become the cause of the stock declines. This is a real-world demonstration of the classic financial theory of "reflexivity," systematized by George Soros.

On the surface, the collective plunge of US optical communications stocks looks as if the market validated SemiAnalysis's analysis. In reality, it was because the market chose to believe the report that investors collectively executed sell orders. When a research report itself becomes the catalyst for market movements, whether it is right or wrong degenerates in the short term into a circular argument—the stock plunge only proves the firm's massive psychological influence within institutional circles, rather than serving as ironclad proof that its technical judgment has been verified by facts.

At this point, we can tie back the clue dropped at the beginning of the article. Faced with the same negative news of a technical delay, AAOI plummeted about 17% in a single day, while Lumentum, which shares a highly similar technical profile, fell only about 8%. If this had been a fundamental-driven, rational repricing, the decline of these two highly similar targets should not have differed by a factor of two. Such capital flow indicates that the stocks suffering the worst hits that day were typically those with the loosest ownership structures, higher margin balances, or higher retail investor concentration. It resembled a shakeout designed to flush out weak hands using panic sentiment rather than a sober fundamental reassessment.

More importantly, looking at objective data, this event is clearly more of the latter, as global demand for AI compute has shown no signs of contraction. The combined capital expenditures of the four major cloud giants—Microsoft, Amazon, Google, and Meta—have risen to a range of approximately $650 billion to $725 billion for 2026 (approaching or even exceeding $700 billion by most metrics), representing a year-over-year increase of about 60% to 80% (about 1.6 to 1.8 times) compared to roughly $380 billion to $410 billion in 2025. At the corporate level, Amazon guided around $200 billion, Alphabet around $175 billion to $185 billion, Meta around $115 billion to $135 billion, and Microsoft's capital expenditure run-rate for fiscal year 2026 is equivalent to approximately $145 billion. The giants' demand for compute procurement remains robust; what has truly shifted is merely the market's expectation regarding the timing of technology adoption along a specific timeline.

11. Methodology: When an "Authoritative Report Triggers a Sell-off," Go Through These Two Layers of Thinking First

To condense the above breakdown into an actionable framework: when you next encounter a stock plunge triggered by an authoritative research report, there are two layers of analysis that must be completed in sequence before deploying capital.

The first layer is learning how to properly read the research. When a stock plummets, first ask yourself three questions: First, who was this report actually written for? Identify its incentive structure—is it an institutional-only trading report aimed at steering position adjustments, or a public academic discussion? Second, is the sharp drop a validation of bearish fundamentals, or is it a sentiment-driven repricing fueled by institutional influence? Third, what specific timeline does the "delay" claimed in the report actually refer to? Avoid mistaking the postponement of a highly complex architecture for the complete demise of a core technology path.

The second layer is how to adjust your actions accordingly, which focuses on the decision-making process rather than specific assets. High-quality research reports are meant to test your existing investment hypotheses, not to serve as actionable buy or sell signals. The correct approach is to ask: this report presents a counterargument I hadn't previously considered seriously; can my current investment thesis withstand the scrutiny of this counterargument? It is not to panic-sell the moment an authoritative institution mentions a "delay." When a report itself acts as a market-moving catalyst, trading in the direction of the report after retail circles have caught wind of the news and the stock has already crashed will often leave you holding the bag in a panic sell-off. This is because by the time the sell-off is publicly triggered, the negative information within the report has already been priced in almost instantaneously.

Therefore, the real question to ask at this moment is not "should I follow suit and sell?" but "has this sell-off triggered an irrational overreaction?" You must translate the previously discussed three-tier timeline into questions regarding your own portfolio and positions: is your original thesis for holding the asset built on the "adoption of a brand-new future architecture," or is it a bet on "the demand window for traditional pluggable optical modules?" Faced with the same CPO delay report, the underlying logic and market reactions of these two types of positions are often diametrically opposed. Finally, train yourself to think about "second-order questions": while most retail investors remain stuck in the linear thinking of "a delay equals bad news," more sophisticated investors are already asking—who stands to unexpectedly benefit from this technological delay? And who smuggled those overly aggressive, priced-to-perfection expectations into the stock's valuation in the first place?