Source: Wall Street News
A report from SemiAnalysis, a leading AI industry chain analysis firm, pointing to delays in two core technology paths for AI data centers, triggered a major shock in the optical communication sector on June 10th, and sparked heated debate in the investment and industry communities about future technology routes and investment opportunities.
The report suggests that shipments of Nvidia's 800VDC power architecture will be delayed until 2028, while mass production of CPO (co-packaged optics) may be delayed until 2028 or even 2029. The simultaneous downward revision of these two expectations caught the market off guard.
Following the news, US-listed optical communication stocks generally plummeted. AAOI Optoelectronics (AAOI) fell by as much as 17% in a single day, Lumentum dropped by about 8%, and companies such as HIMX, Navitas Semiconductor Corp, and Wolfspeed, which were specifically mentioned in the report as having adopted a cautious stance, also experienced significant pressure.
Alongside the release of the SemiAnalysi report, interviews with Nvidia executives were also published. According to veteran semiconductor and technology investment journalist Tae Kim, Gilad Shainer, Senior Vice President of Nvidia's Networking Business, expressed a completely opposite optimistic view on the prospects of CPO at Computex 2026, stating that "CPO is the most exciting technology right now" and that mass shipments would begin in the second half of the year. This sparked a heated debate on social media regarding the CPO timeline.
It is worth noting that many market observers pointed out that the postponement of CPO does not mean that the demand for optical interconnects has disappeared, but is more likely to redirect the flow of capital back to the traditional pluggable optical module and NPO (near-package optics) track - this logic led some investors to look for opportunities to sell at a loss during the panic selling.
SemiAnalysis Report Highlights: Both Major Technology Paths Delayed
In this research note sent to institutional clients, SemiAnalysis presented two key judgments that have a profound impact on the market.
The 800VDC power architecture has been postponed until after 2028.
The report points out that Nvidia's original plan to widely adopt the single-ended 800VDC power supply design has significantly delayed its shipping window. Hyperscale cloud vendors are currently more inclined to continue using mature low-voltage solutions or gradually transition to 400VDC, rather than rushing to switch to 800VDC.
The report argues that the marginal efficiency gains of 800VDC under current grid conditions are insufficient to support its system complexity. In contrast, 400VDC products are expected to begin ramping up production in the second quarter of 2026, with significant growth anticipated in 2027.
CPO production is lagging far behind market expectations.
The report states that CPO shipments in 2027 will be significantly lower than previously aggressive forecasts, and mass production may be delayed until 2028 or even 2029. The main bottlenecks are concentrated in three areas:
The yield rate of optical engine connections (around 95% in an optimistic scenario, but the CPO production driven by a single ASIC is still extremely limited), the difficulty of ASIC integration, and the overall cost economy.
Shipments of scale-out CPO switches face the risk of being revised downwards, and shipments of Sidecar, which relies on new platforms such as Rubin Ultra/Kyber, will also be delayed until the 2028 window.
At the individual stock level, SemiAnalysis maintains a relatively positive view on Amphenol, Vertiv, and Legrand, while taking a cautious stance on Lumentum, Himax Technologies, Navitas Semiconductor, and Wolfspeed.
However, the report itself also acknowledges that CPO as an important direction for future data center network architecture has not been denied. The core reason for the delay is that the engineering challenges have not been fully overcome , rather than the disappearance of demand.
At the same time, the report also pointed out that some NPO (Near Package Optics) projects may be accelerated.
Nvidia executives publicly contradict Tae Kim, whose interview sparks attention.
At the same time that the SemiAnalysis report was widely circulated in institutional circles, Tae Kim, a senior semiconductor and technology investment journalist, published a transcript of a one-on-one interview with Gilad Shainer, Senior Vice President of Nvidia's Networking Business, during Computex on his Substack column. The content of the interview stood in stark contrast to SemiAnalysis's assessment.
In an interview, Shainer stated, "The most exciting thing today is co-packaging optics, which is at the forefront of technological advancement."
He further revealed that Nvidia is ready to begin shipping, and its partner Lambda has confirmed in a blog post that it has secured CPO switches. The company will accelerate the rollout of CPO in the second half of the year, extending it from scale-out to scale-up scenarios. "If it were up to me, I would want to adopt CPO wherever optical networks are used."
In his article, Tae Kim added that Shainer's overall demeanor and body language during the interview demonstrated a high level of enthusiasm for both the near-term and long-term growth of CPOs. He stated that this statement "seems to directly contradict SemiAnalysis's narrative."
This contrast has plunged the market into a chaotic information war. User @qinbafrank on the social media platform X pointed out that Bernstein had already clearly stated in its mid-May report that cloud vendors would not sacrifice system reliability for energy efficiency, and that no cloud vendor currently plans to deploy CPO on a large scale in 2026-2027. "If you had read Bernstein's report in detail, you wouldn't be surprised by SemiAnalysis today."
Online debate: Is the CPO postponement a negative factor or a misjudgment?
The market turmoil triggered by the report quickly spread to social media, with opinions diverging significantly regarding the investment logic surrounding the CPO extension.
Bearish view: Yield and reliability are the real bottlenecks.
SemiAnalysis' report emphasizes that in the CPO architecture, the optical engine and a large ASIC worth tens of thousands of dollars are co-packaged on the same substrate. Once the optical engine fails due to laser aging or fiber damage, the entire motherboard often needs to be disassembled and returned to the factory. The maintenance costs and downtime risks are far higher than those of traditional pluggable modules. This engineering challenge is considered a core obstacle to the large-scale deployment of CPO in the short term.
From a bullish perspective: The CPO delay is actually beneficial for pluggable modules and NPOs.
A user on the social media platform X, @TomSzczypka, analyzed the situation, stating, " If the CPO arrives late, the data still needs to be transmitted. AI clusters cannot wait two years. Hyperscale cloud providers will spend more time purchasing more pluggable modules and NPOs. The money won't disappear; it will just be transferred to other sources. "
He also pointed out that the fact that Xiangmao Optoelectronics' drop (17%) far exceeded that of Lumentum (8%) on that day indicates that the market's sell-off that day was not based on rational analysis, but rather on clearing out the weakest holdings.
User @michaelsikand stated that currently, no photonics company generates zero revenue from CPOs, and the current rapid growth comes from huge, unmet NPO opportunities. "The timeline may be delayed, but TAM will not."
There are also voices questioning the report's logic.
User @cherryPayment published a lengthy article pointing out internal contradictions in the SemiAnalysis report: on the one hand, the report claims that the supply chain is not ready in 2027, while on the other hand, it predicts that Celestial AI (acquired by Marvell) will reach a revenue operating rate of $1 billion by the end of 2028, and that Amazon has already signed a contract for Trainium 4. "You can't suddenly have $1 billion by the end of 2028, while the supply chain is not ready at all in 2027."
He also pointed out that SemiAnalysis targets procurement decision-makers at hyperscale cloud vendors, and its conclusion is "not to go all in now," rather than a judgment on the timing of investment in the capital market. "They analyze the deployment pace, not the timing of investment."
X platform user @Herman Jin criticized the timeliness of information from US investment research institutions, arguing that the delays in CPO and 800VDC were "just a matter of time," and that relevant information had been circulating in institutional circles for some time. SemiAnalysis' report was simply putting known information into writing.
Unexpected beneficiaries: copper connections and pluggable modules
Amid widespread market pressure, some analysts have turned their attention to potential beneficiaries of CPO extensions.
User @qinbafrank believes that more realistic revenue opportunities in 2026 are concentrated in areas such as 1.6T pluggable modules , LPO/NPO, light sources, testing, PCB, ABF, and CCL. "Light will not immediately eliminate copper, and copper will not always be available in all scenarios. Different solutions will be chosen for different distances and system levels."
Lumentum's CEO recently stated that non-Nvidia customers have shown a significant increase in interest in NPOs over the past two months.
User @RealNickMugalli analyzed that, under the conditions of 1.6T rate and 200G per channel, copper cable, even with retimer technology, has reached its physical limit. Optical solutions will become a mandatory option rather than an optional one within a reasonable distance. The potential market size of NPO may even exceed that of CPO.
The SemiAnalysis report also noted that some NPO projects may accelerate, and 400VDC products will begin to see mass production in the second quarter of 2026. The report maintains a relatively positive stance towards companies such as Amphenol and Vertiv, believing they will benefit from continued demand during the 400VDC transition period.
User @TomSzczypka cited industry chain data from this week to support the conclusion that the demand for AI infrastructure has not weakened:
Fujikura raised data center cable prices due to simultaneous orders from almost all US hyperscale cloud providers; King Slide rack rail revenue increased by 47% quarter-over-quarter; Google ordered 6 million TPUs from Intel; SK Hynix signed a multi-year storage cooperation agreement with Nvidia, etc.
"The real bottlenecks in AI lie in power, storage, and GPUs, and none of these three have gotten worse today."
Meanwhile, @tuolaji2024 posted on the social media platform X that the supply and demand tension of memory (HBM/DRAM), as the real physical bottleneck, is completely unaffected by this technology delay event.
Analysis suggests that, based on various perspectives, the market volatility triggered by this SemiAnalysis report reflects more of a recalibration of technology roadmap timelines than a fundamental reversal in overall demand for AI data centers.
