A report from prominent AI industry analysis firm SemiAnalysis, which directly pointed to delays in two core AI data center technology paths, triggered significant volatility in the optical communication sector on June 10th, simultaneously igniting intense debate within the investment and industrial communities regarding future technology roadmaps and investment opportunities.
The report suggests that NVIDIA's 800VDC power architecture shipments will be postponed until 2028, while the scaled mass production of Co-Packaged Optics (CPO) may be delayed until 2028 or even 2029. The simultaneous downward revision of these two expectations caught the market off guard.
Following the news, the US optical communication sector broadly sold off sharply. Applied Optoelectronics (AAOI) plunged as much as 17% in a single day, Lumentum fell roughly 8%, and other companies mentioned in the report with a cautious outlook, such as Himax Technologies (HIMX), Navitas Semiconductor Corp, and Wolfspeed, also faced notable pressure.
Key Report Findings: Dual Technology Path Delays
The SemiAnalysis report, sent to institutional clients, presented two core judgments with significant market impact.
The 800VDC power architecture is pushed back to 2028 or later. The report states that the shipment window for NVIDIA's originally planned, large-scale adoption of a single-ended 800VDC power design has been significantly delayed. Hyperscale cloud providers currently prefer to continue using mature low-voltage solutions or gradually transition to 400VDC, rather than urgently switching to 800VDC.
The report argues that the marginal efficiency gains of 800VDC under current grid power conditions are insufficient to justify its system complexity. In contrast, 400VDC products are expected to begin ramping in Q2 2026, with significant growth in 2027.
The CPO mass production timeline lags far behind market expectations. The report indicates that 2027 CPO shipment volumes will be significantly lower than previous aggressive forecasts, with scaled mass production potentially delayed until 2028 or 2029. Key bottlenecks are concentrated in three areas: optical engine connection yield (optimistically around 95%, but CPO production per single ASIC remains extremely limited), ASIC integration difficulty, and overall cost economics.
Scale-out CPO switch shipments face downward revision risks, and Sidecar shipments dependent on new platforms like Rubin Ultra/Kyber are also delayed to the 2028 window.
At the individual stock level, SemiAnalysis maintains a relatively positive view on companies like Amphenol, Vertiv, and Legrand, while expressing caution towards Lumentum, Himax Technologies, Navitas Semiconductor, and Wolfspeed.
However, the report itself acknowledges that CPO, as an important future direction for data center network architecture, is not being negated. The core reason for the delay is that engineering challenges have not yet been fully overcome, not that demand has disappeared. Simultaneously, the report notes that some Near Package Optics (NPO) projects may accelerate.
NVIDIA Executive's Contradictory Viewpoint
Around the same time the SemiAnalysis report circulated widely among institutions, veteran semiconductor and technology investment journalist Tae Kim published a one-on-one interview transcript from Computex with NVIDIA's Senior Vice President of Networking, Gilad Shainer, in his Substack column. The content presented a stark contrast to SemiAnalysis's assessment.
Shainer stated in the interview, "The most exciting thing today is co-packaged optics. It is the leading edge of the technology frontier." He further revealed that NVIDIA is ready to begin shipments, with partner Lambda having published a blog confirming receipt of CPO switches, and that CPO volume ramp will accelerate in the second half of the year, extending from scale-out to scale-up scenarios. "If it was up to me, I would want CPO everywhere we use optical networking."
Tae Kim added in the article that Shainer's overall demeanor and body language during the interview showed high enthusiasm for both near-term and long-term CPO volume ramp, a statement that "seems to directly contradict the SemiAnalysis narrative."
This discrepancy plunged the market into an information battle. A user on social platform X pointed out that Bernstein had clearly stated in a mid-May report that cloud providers would not sacrifice system reliability for energy savings, and no cloud provider currently plans large-scale CPO deployment in 2026-2027, suggesting that a detailed read of that report would have prevented surprise at the SemiAnalysis update.
Market Debate: Is CPO Delay a Negative or a Mispricing?
The market volatility triggered by the report quickly spread to social media, where views on the investment logic surrounding the CPO delay diverged sharply.
Bearish View: Yield and Reliability are Real Bottlenecks. SemiAnalysis emphasized in its report that in the CPO architecture, the optical engine is co-packaged with a large ASIC worth tens of thousands of dollars on the same substrate. If the optical engine fails due to laser aging or fiber damage, the entire motherboard often needs to be removed and returned to the factory, with maintenance costs and downtime risks far exceeding those of traditional pluggable modules. This engineering challenge is seen as the core obstacle to large-scale CPO adoption in the near term.
Bullish Counters: CPO Delay Benefits Pluggable Modules and NPO. An analysis posted on X argued that if CPO arrives late, data still needs to be transmitted, AI clusters cannot wait two years, and hyperscalers will buy more pluggable modules and NPO for a longer period. The money does not disappear; it just changes pockets.
The same user noted that the fact Applied Optoelectronics's decline (17%) far exceeded Lumentum's (8%) indicated the day's selloff was not based on rational analysis but was a washout of the weakest holdings.
Another user stated that current revenue from CPO for any photonics company is zero, and the current high growth comes from the huge, unmet NPO opportunity, suggesting the timeline may be pushed back, but the Total Addressable Market (TAM) will not.
Skeptical voices questioning the report's logic also emerged. One user posted a lengthy critique pointing out internal contradictions in the SemiAnalysis report, arguing it was analyzing deployment pace, not investment timing for the capital markets.
Another X user criticized the timeliness of information from US investment research firms, stating that CPO and 800VDC delays were "a matter of time" and that related information had already been circulating in institutional circles, with the SemiAnalysis report merely formalizing known information on paper.
Potential Beneficiaries: Copper Interconnects and Pluggable Modules
Amid the broad market pressure, some analysts turned their attention to potential beneficiary directions from the CPO delay.
One user's analysis suggested that more realistic revenue opportunities in 2026 are concentrated in areas like 1.6T pluggable modules, LPO/NPO, light sources, testing, PCB, ABF, and CCL, noting that optics will not immediately eliminate copper, nor will copper forever hold all scenarios, with different solutions chosen for different distances and system levels.
The CEO of Lumentum recently stated that interest in NPO from non-NVIDIA customers has noticeably increased over the past two months.
Another analysis argued that at 1.6T rates and 200G per lane, copper cables have reached physical limits even with retimer technology, making optical solutions a mandatory option, not a choice, within reasonable distances, with NPO's potential market size possibly even exceeding CPO's.
SemiAnalysis also noted in its report that some NPO projects may accelerate, and 400VDC products will begin ramping in Q2 2026. For companies like Amphenol and Vertiv, the report maintains a relatively positive stance, believing they benefit from sustained demand during the 400VDC transition period.
Another user cited recent supply chain data as evidence that AI infrastructure demand has not weakened, pointing to price increases for data center cables, strong growth in rack slide revenue, a major TPU order, and a multi-year memory partnership as signs that the real bottlenecks for AI remain power, storage, and GPUs, none of which have worsened.
Simultaneously, a post on X noted that memory (HBM/DRAM), as the real physical bottleneck, faces a supply-demand tension completely unaffected by this technology delay event.
Analysis concludes that, synthesizing various viewpoints, the market volatility triggered by the SemiAnalysis report more reflects a recalibration of the technology roadmap timeline rather than a fundamental reversal of overall AI data center demand.
Comments