SemiAnalysis report triggers sell-off in U.S. optoelectronics sector

Written by Xiao Bing, Chaoxiang Research

On June 9, the U.S. stock market's optical communications sector plunged collectively.

Applied Optoelectronics (AAOI) plunged 14%, Coherent (COHR) dropped 11%, Lumentum (LITE) fell 8%, Ciena (CIEN) declined 7%, Corning (GLW) slid 9%, and Marvell (MRVL) dropped 9%. In a single day, the entire photonic electronics supply chain was hammered.

The trigger was a report sent by SemiAnalysis to institutional clients, titled bluntly: “Power Outage: 800V DC Deployment Delayed, Co-Packaged Optics (CPO) Progress Slowed.” The report expressed a short-term bearish outlook on two of this year’s most crowded market themes: 800V DC high-voltage power delivery and co-packaged optics, with the core assessment being that the deployment timelines for both are significantly slower than market expectations.

Gilad Shainer, Senior Vice President of Nvidia’s networking division, clearly stated in an interview at the same Computex that there are no delays in CPO shipments. Nvidia recently announced in early June that its Spectrum-X Ethernet Photonics has entered mass production, with CoreWeave, Lambda, Meta, Microsoft, and Oracle among the first adopters.

On one side, the most influential chip company in the semiconductor industry says “no problem”; on the other, a research institute says everything is delayed. The market chooses to believe the latter.

This report is highly information-dense, covering both the 800VDC and CPO lines; below are the key arguments.

800VDC high-voltage DC: Nvidia's flagship solution faces widespread cooling

SemiAnalysis points out that there are two independent technical pathways for 800VDC, with vastly different fates.

Nvidia’s promoted single-ended 800VDC solution, originally slated for widespread adoption by 2027, has now been delayed to 2028 and beyond, due to a specific reason: the baseline Rubin hardware does not require 800VDC power and continues to use the 50V solution. 800VDC becomes essential only for next-generation, high-power hardware such as Rubin Ultra and Feynman, whose design specifications for Rubin Ultra won’t be finalized until late this year. Cloud providers generally agree that converting grid voltage from 350–450V up to 800V and then stepping it down to 50V is highly inefficient. The industry consensus is shifting toward high-voltage transmission with centralized step-down conversion.

In contrast, the ±400VDC solution has seen no impact on its timeline. This architecture is being driven independently by cloud providers, primarily paired with proprietary ASICs, with orders expected to begin by the end of 2026 and mass production targeted for Q1 2027. The report emphasizes that the two approaches are independent, and the ±400VDC power distribution modules will also be compatible with Nvidia hardware in the future.

SemiAnalysis's qualitative assessment of the 800VDC delay is "delayed, not abandoned": when the power consumption of a single computing module exceeds the 15-kilowatt threshold, the efficiency advantages of native high-voltage DC distribution will become evident, making 800VDC the ultimate solution for high-power hardware.

CPO co-packaged optics: Yield arithmetic shatters optimistic expectations

The other half of the report focuses on CPO, providing assessments along both the scale-out and scale-up pathways.

The primary challenge facing scale-out CPO is yield. The report provides a set of key figures: under optimistic assumptions, the yield for a single COUPE optical engine is 95%. Since the Nvidia Spectrum 6 CPO switch requires integrating 32 optical engines per chip, the overall system yield is only 19.4%. More critically, the optical engines are directly soldered onto the substrate, making repair impossible once faulty. To achieve scalable profitability, the individual yield must reach 99.5%, so that the system yield for 32 engines combined reaches 85%.

The report also revealed a previously unknown technical issue: the Nvidia second-generation optical engine Spectrum 6 CPO switch has an onboard system insertion loss exceeding 3.5 dB, having exhausted all optical channel tolerance, resulting in performance even weaker than the previous generation. Nvidia and TSMC have not yet identified the root cause and are redesigning the assembly process.

The CPO scale-up assessment is more aggressive: while the market generally expects large-scale production between 2027 and 2028, SemiAnalysis delays this timeline to 2029. The rationale is that key projects from Amazon Web Services, AMD, and Feynman will not be fully deployed until 2029, at which point the optical engine technology with embedded interposers will be mature. Although a limited number of NVL576 units may ship in 2027–2028, they will only be used for switch interconnects without accompanying GPUs, resulting in limited volume.

The conclusion of the report directly identifies market structure risks: a large concentration of capital going long on optoelectronics and power semiconductor stocks, while shorting large platform companies such as Nvidia and Broadcom. Relevant individual stocks have risen to historical highs, and risk appetite has reached its peak. If the timing of implementation falls short of expectations, capital positioned at these highs will exit en masse. However, SemiAnalysis emphasizes that it remains fundamentally bullish on both sectors, merely anticipating a short-term delay in timing.

To understand why this report has such a powerful impact, you must first understand the organization SemiAnalysis itself.

Dylan Patel, 29, founded SemiAnalysis in 2020 without a degree in semiconductors.

Over five years, this one-person Substack blog grew into one of the most influential information hubs for the global AI and semiconductor industries. Annual revenue surged from approximately $20 million in 2025 to an expected exceedance of $100 million in 2026, making it the top-ranked technology subscription on Substack with over 250,000 subscribers.

Patel’s influence has reached the highest levels of the industry. At the GTC conference in March 2026, Jensen Huang mentioned only two people by name throughout his entire keynote speech—one of them was Dylan Patel—and even projected SemiAnalysis’s InferenceX chip performance benchmark report onto the big screen, spending five minutes explaining it. AMD CEO Lisa Su scheduled a dedicated 90-minute in-person meeting.

SemiAnalysis's competitive advantage lies in translating highly specialized semiconductor supply chain analysis into language that investors can understand. It fills a void: traditional sell-side research is too slow and conservative, while tech media is too shallow and emotional. SemiAnalysis’s reports combine chip-level technical depth with sharp, action-oriented conclusions that directly inform trading decisions.

Tech company executives use it for competitive intelligence, and hedge funds use it as a basis for trading.

The CPO event on June 9 was almost a repeat of the Micron event five days earlier.

On June 5, SemiAnalysis released a report stating that Nvidia significantly reduced the modular memory capacity of its next-generation Vera Rubin servers from 55TB to 28TB. The market interpreted this as a sign of cooling demand for AI memory. Micron’s stock plunged 13% that day, marking its largest single-day drop since April 2025.

But on the same day, Jensen Huang publicly announced that Micron has been certified for Nvidia’s HBM4. Micron’s CFO also stepped forward to refute “inaccurate reports,” emphasizing that HBM4 is already in mass production and is shipping a quarter ahead of schedule.

SemiAnalysis later responded on X: The report was never intended to be bearish. Patel said the true conclusion of the report was that "Micron’s HBM delays are actually beneficial for Micron, because the margins on standard DDR are higher than those on HBM." He added, "Those claiming we’re bearish haven’t even read our full report. They don’t even have subscription access."

This response itself reveals the core issue: a report visible only to institutional clients was relayed by the media and reinterpreted by the market, reducing the conclusion to “Nvidia cut Micron’s orders,” triggering panic selling by retail investors. Some users directly questioned: you provide accurate information to institutional clients while leaving ordinary subscribers and retail investors confused—this effectively creates an asymmetric advantage for your wealthy clients.

If we turned back the clock three months, a similar report from SemiAnalysis would likely not have caused such significant damage—back then, the semiconductor sector was riding a steady upward trend, with bad news seen as buying opportunities and good news as reasons to add more positions.

But the current U.S. stock market is no longer the same U.S. stock market.

On June 5, the Philadelphia Semiconductor Index plunged nearly 8.5% in a single day, with a cumulative decline of over 10% over two days, wiping out $1.3 trillion in market value. The immediate trigger was Broadcom’s earnings guidance falling short of expectations, but the underlying cause was a perfect storm: stronger-than-expected May non-farm payroll data reignited expectations of Fed rate hikes; SpaceX launched its IPO at a $1.75 trillion valuation, triggering significant capital reallocation; the VIX index surged 24% within five days; and the Nasdaq recorded its largest single-day drop since April 2025.

In this environment, the market's reaction function has shifted from "buy on good news, ignore bad news" to "amplify bad news, question good news."

The CPO report from SemiAnalysis, if released at any other time, might have triggered only a 3%-5% sector adjustment. But in a market that has just experienced "Black Friday," with investors having just witnessed trillions of dollars vanish in real time, the same report’s impact was tripled.

The market has become skittish—the arrow hasn’t changed, but the bird is already frightened.

Second, over the past few months, substantial capital has flowed into photonics and power semiconductors, while shorting or underweighting platform companies like Nvidia and Broadcom. This bet rests on a simple logic: as AI infrastructure continues to spill over, suppliers of optical modules, power systems, and materials will keep benefiting. The sector itself is sound, but positions have reached extreme levels. AAOI has surged over 400% this year, LITE over 150%, and valuations are fully priced with perfect expectations of rapid CPO volume growth. There is virtually no margin for error.

This report by SemiAnalysis precisely identifies the weakest point in this trade; this crash is more accurately characterized as a decongestion of a crowded trade.

CPO has not been disproven, and 800VDC is not out of the picture. The physical limits of copper cabling at 600kW rack density remain unchanged; optical interconnects and high-voltage DC are still the right direction. The market simply moved too far ahead too soon. When a sector is pushed to high levels by capital, narratives, and valuations simultaneously, any signal of “not happening as fast” gets amplified into “is it even viable?”

The direction is still there, but the pace needs to be repriced. When a sector has risen too quickly and become too crowded, the only response to a slowdown signal is: get out first.