Written by: Xiao Bing, Trend Research

On June 9, the US stock market for optical communications plummeted.

Applied Optoelectronics (AAOI) dropped 14%, Coherent (COHR) fell 11%, Lumentum (LITE) decreased 8%, Ciena (CIEN) went down 7%, Corning (GLW) fell 9%, Marvell (MRVL) dropped 9%. In one day, the entire optoelectronics industry chain was pressed to the ground.

The trigger was a report sent by SemiAnalysis to institutional clients, with a straightforward title: "Power Outage Halt: 800 Volt DC Plan Delayed, Co-packaged Optics (CPO) Process Postponed." The report is also bearish in the short term on two of this year's most crowded market segments: 800VDC high-voltage direct current power supply and co-packaged optics, with the core judgment being that the pace of implementation for both is significantly slower than the market expected.

Nvidia's Senior Vice President of Networking, Gilad Shainer, clearly stated in an interview at Computex that there is no delay in CPO shipments. Nvidia just announced in early June that Spectrum-X Ethernet Photonics has already entered mass production, with CoreWeave, Lambda, Meta, Microsoft, and Oracle being the first adopters.

On one side is the most authoritative chip company in the semiconductor industry saying "no problem," and on the other side is a research institution claiming a comprehensive delay, and the market chose to believe the latter.

What did the report say?

This report has a high density of information, covering both 800VDC and CPO, with the following key points.

800VDC High Voltage Direct Current: Nvidia's promoted solution has collectively cooled off

SemiAnalysis pointed out that 800VDC has two independent technical routes with drastically different fates.

The single-ended 800VDC scheme strongly promoted by Nvidia, originally expected to be widespread by 2027, has been postponed to 2028 and beyond, with specific reasons: The basic version of Rubin hardware does not require 800VDC power supply and continues to use the 50-volt scheme. 800VDC only becomes a necessity on next-generation high-power hardware like Rubin Ultra and Feynman, and the design scheme for Rubin Ultra won't be finalized until the second half of this year. Cloud vendors generally believe that the efficiency of the conversion path that steps up power from the grid's 350-450 volts to 800 volts and then steps it down to 50 volts is extremely low. The industry consensus is shifting: High voltage transmission, centralized step-down.

In contrast, the ±400VDC program has seen no impact on its progress. This architecture is driven autonomously by cloud vendors, mainly supporting self-developed ASICs, expected to start taking orders by the end of 2026 and enter mass production in the first quarter of 2027. The report emphasizes that the two routes are independent, and the ±400VDC distribution module can also be compatible with Nvidia hardware in the future.

SemiAnalysis qualifies the delay of 800VDC as "progress delayed, solution not abandoned": When the power consumption of a single computing module exceeds the critical value of 15 kilowatts, the efficiency advantages of native high-voltage DC distribution will be prominent, and 800VDC remains the ultimate solution for high-power hardware.

CPO Co-packaged Optics: Yield calculations shatter optimistic expectations

The other half of the report focuses on CPO, also providing judgments based on scale-out (mass deployment) and scale-up (capacity upgrade) lines.

The primary obstacle facing scale-out CPO is yield rates. The report provides a key set of data: The mounting yield rate for a single COUPE optical engine is 95% under optimistic assumptions, but the Nvidia Spectrum 6 CPO switch requires integrating 32 optical engines, resulting in a combined system yield of only 19.4%. More critically, the optical engines are directly soldered onto the substrate and cannot be repaired if they fail. To achieve scalable profitability, the yield rate for a single unit needs to reach 99.5%, so that the system yield for 32 units can reach 85%.

The report further disclosed a previously unknown technical issue in the market: Nvidia's second-generation optical engine in the Spectrum 6 CPO switch has an onboard system insertion loss exceeding 3.5 decibels, exhausting all optical channel tolerances, with performance being even weaker than the previous generation. Nvidia and TSMC have yet to identify the cause of the fault and are redesigning the assembly process.

The judgment for scale-up CPO is even more aggressive: The market generally expects large-scale production in 2027-2028, while SemiAnalysis has pushed this timeline back to 2029. The reason is that core projects like Amazon Cloud, AMD, and Feynman won’t materialize until 2029, and the optical engine technology embedded in the intermediary layer will only mature by then. While there may be a small quantity of NVL576 models shipped in 2027-2028, they will only be for switch interconnections, not supporting GPUs, with limited volume.

The conclusion of the report directly names market structure risks: a large amount of capital is concentrated in bullish positions on optoelectronics and power semiconductor targets while simultaneously shorting large platform companies like Nvidia and Broadcom. Related stocks have surged to historical highs, with risk appetite reaching a peak. If the pace of realization does not meet expectations, high-position capital will withdraw en masse. However, SemiAnalysis emphasizes that it remains bullish on both tracks in the long term, just judging that the short-term pace has been delayed.

SemiAnalysis: The "Super Influencer" of the Semiconductor Industry

To understand why this report can cause such significant damage, one must first understand SemiAnalysis itself.

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

In five years, this one-person Substack blog has transformed into one of the most influential information nodes in the global AI and semiconductor industry. Annual revenue skyrocketed from about 20 million USD in 2025 to a projected over 100 million USD in 2026, ranking first in technology-related subscriptions on Substack, with over 250,000 subscribers.

Patel's influence has penetrated the highest echelons of the industry. At the GTC conference in March 2026, Jensen Huang mentioned only two individuals during the entire keynote speech, one of whom was Dylan Patel, even projecting SemiAnalysis's InferenceX chip performance benchmark report onto the big screen and spending five minutes explaining it. AMD CEO Lisa Su has specifically arranged for a 90-minute face-to-face meeting.

SemiAnalysis's core competitiveness lies in translating highly professional semiconductor supply chain analysis into language that investors can understand. It fills a vacuum: traditional sell-side research is too slow and conservative, tech media is too superficial and emotional, while SemiAnalysis's reports have both the technical depth of chip-level analysis and sharp conclusions directly pointing to trading decisions.

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

A Bird Startled by the Bow

The CPO incident on June 9 was almost a replay of the Micron incident five days earlier.

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

But on the same day, Jensen Huang publicly announced that Micron has passed Nvidia's HBM4 certification. Micron's CFO also rebutted "inaccurate reports," emphasizing that HBM4 has already entered mass production, and the shipping schedule is even a quarter ahead of plan.

SemiAnalysis subsequently responded on X: the report did not imply any bearish sentiment. Patel stated that the true conclusion of the report is that "Micron's HBM delay is actually good for Micron because the profit margins of standard DDR are higher than those of HBM." He also added, "Those who say we are bearish haven't even read our complete report. They simply do not have subscription access."

This response itself exposed the core of the issue: a report that only institutional clients can see, after being relayed and reinterpreted by the media, was simplified to the conclusion of "Nvidia cut Micron's orders," triggering panic selling among retail investors. Some users directly questioned: you provide accurate information to institutional clients but confuse ordinary subscribers and retail investors, effectively creating asymmetric advantages for your wealthy clients.

If we turn back the clock three months, similar reports from SemiAnalysis would likely not have caused such damage; at that time, the semiconductor sector was thriving, bad news was seen as a buying opportunity, and good news was viewed as a reason to increase positions.

But the current US stock market is no longer the same.

On June 5, the Philadelphia semiconductor index plunged nearly 8.5% in a single day, accumulating a drop of over 10% in two days, evaporating 1.3 trillion USD 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 non-farm payroll data in May reignited interest rate hike expectations; SpaceX launched its IPO at a valuation of 1.75 trillion USD, with significant capital rotation effects; the VIX index surged 24% in five days; and the Nasdaq recorded its largest single-day decline since April 2025.

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

SemiAnalysis's CPO report issued at any other time would likely only trigger a 3%-5% sector adjustment. However, in a market that has just experienced "Black Friday" and where investors have just seen a trillion-dollar evaporation in panic, the damage of the same report was amplified threefold.

The market has turned into a bird startled by the bow; the bow and arrow remain unchanged, but the bird has been frightened.

Furthermore, in the past few months, significant capital has flowed into optoelectronics and power semiconductors, while simultaneously shorting or underweighting platform companies like Nvidia and Broadcom. This combination bets on a simple logic: AI infrastructure continues to spill over, benefiting those selling components such as optical modules, power supplies, and materials. The track itself is fine, but the positions have reached their limits. AAOI has surged over 400% this year, LITE over 150%, with valuations packed with perfect expectations for rapid CPO volume. The margin for error is almost zero.

SemiAnalysis's report precisely struck at the most vulnerable point of this trade; this drop can more accurately be qualified as a de-risking of crowded trades.

CPO has not been disproven, and 800VDC has not exited. The physical limits of copper cables at 600kW rack density are evident, and optical interconnects and high-voltage direct current remain the broad direction. The market simply ran too far ahead. When a track is pushed to a high position by funds, narrative, and valuations simultaneously, any signal of "not so fast" will be amplified into "is it not possible."

The direction is still there, but the pace needs to be repriced. Fast-rising, overcrowded tracks, when encountering signals of deceleration, only have one reaction: run first and ask questions later.