SemiAnalysis report sparks debate over CPO timeline; optical stocks decline

TL;DR

This round of declines in optical stocks appears on the surface to be a cooling of the CPO narrative, but in reality reflects the market reassessing a more sensitive issue: whether the volume growth expected in 2027–2028 will mark a period of actual earnings realization or remain merely in the pilot validation phase.

CPO (Co-Packaged Optics) itself has not been dismissed. The demands for bandwidth, power consumption, and switching density in AI data centers continue to rise, and the physical limitations of copper cables and traditional pluggable optical modules have not disappeared. The issue is that previous price increases had already priced in an aggressive timeline: after NVIDIA pushed CPO into the commercialization window, components such as optical engines, lasers, silicon photonics, and switching chips were expected to rapidly enter large-scale shipment by 2027–2028.

The SemiAnalysis report dated June 9 directly challenged this pricing assumption. According to publicly shared insights, the report suggested that large-scale production of NVIDIA’s 800V DC and CPO may be delayed until around 2028–2029, while 400V DC ramp-up remains on track for 2026, with some NPO (near-package optics) projects potentially accelerating. Following the report’s release, market participants saw corrections of high single-digit to double-digit percentages in optical and related stocks such as AAOI, LITE, COHR, GLW, and MRVL. The market is not debating whether the CPO direction is valid—it’s focused on how quickly CPO can turn into orders.

But this is not a one-sided bearish case. “White-Haired God of Stocks” and AI supply chain analyst Serenity (@aleabitoreddit) subsequently countered SemiAnalysis, arguing that it overly relies on conservative engineering models and underestimates NVIDIA’s ability to compress hardware cycles. Based on his interpretation of signals from NVIDIA, Lumentum, Foxconn, and others, he emphasized that CPO is still on an upward trajectory for the second half of 2026, the second half of 2027, and 2028.

The value of this debate isn't about determining who "won," but about shifting the valuation anchor for the optical chain away from endgame narratives and back to timeline validation: CPO is coming, but its rate of adoption will determine how value is distributed among NPO, pluggable modules, light sources, and switching chips.

Over the past few months, the market has been buying into the optical chain, with the core focus not on current revenues, but on the shift in capital expenditures toward the next-generation network architecture for AI data centers.

As model training and inference clusters expand, communication pressure between GPUs, racks, and within data centers continues to rise. Networks are no longer just peripheral systems outside servers—they are increasingly becoming the efficiency bottleneck of AI factories. Higher bandwidth density and lower power consumption enable higher scalability limits for compute clusters, which is why CPO is now taking center stage.

The theoretical appeal of CPO is straightforward: by placing the optical engine as close as possible to the ASIC (application-specific switching chip), it shortens high-speed electrical signal paths, reducing power consumption, signal loss, and signal integrity challenges caused by serialization/deserialization circuits and copper traces. Compared to traditional pluggable modules, CPO offers superior potential for power efficiency and density in the era of high bandwidth.

The issue with the market is that it often trades “correct direction” as if it were “confirmed volume.” According to NVIDIA’s official press release, the Vera Rubin platform will introduce Spectrum-X Ethernet Photonics, with CPO switches already in production for horizontal scaling and cross-cluster deployment in AI factories. On June 3, Taiwanese media reported that NVIDIA’s network business executives stated that Spectrum-X CPO switches have begun shipping to select partners, with production capacity expected to expand in the second half of 2026.

These signals are sufficient to indicate that CPO is progressing, but they cannot be directly equated to risk-free fulfillment of production-scale orders. For capital markets, there is a significant valuation difference between “entering production,” “shipping to select partners,” “customer evaluation,” and “mass production.” The recent market correction triggered by the SemiAnalysis report essentially reflects the market beginning to distinguish between these different milestones.

SemiAnalysis is not saying that CPO has no future. Its core assessment is more like: CPO’s theoretical advantages are clear, but its large-scale implementation is slower than the market expects.

The issue isn't just that one or two components aren't ready—it's that CPO has concentrated the complexity previously spread across modules, circuit boards, and entire systems into a single, more deeply integrated system. The higher the integration, the more impressive the single-point performance, but the greater the pressure on manufacturing, testing, repair, and supply chain resilience.

The benefit of traditional pluggable modules is modularity: if a single optical module fails, it can be replaced, and switching between vendors is relatively straightforward. CPO, however, is different—the optical engine is positioned closer to the ASIC, even integrated into the same package. The power and density advantages stem from this tight coupling, but the repair radius also expands. When an optical component fails, it no longer affects just a easily replaceable module; it may impact the higher-value switching chip and the entire system.

SemiAnalysis’s previous CPO Book repeatedly emphasized serviceability, reliability, yield, and supply chain maturity. In the context of hyperscale cloud providers, performance is not the only metric. Large customers have high requirements for reliability and maintainability; if failure rates, repair processes, and replacement costs in production environments are uncontrolled, even the best power consumption models may be delayed in adoption.

InP lasers are also a point of contention. Laboratory-level runtime data can demonstrate technical feasibility, but it does not equate to long-term operation, mass manufacturing, on-site maintenance, and supply chain redundancy in large-scale data centers. For investors, these distinctions are critical: laboratory validation confirms direction, while field reliability determines scalability.

Within SemiAnalysis’s framework, NPO and plug-in modules are not a backward approach, but rather a more practical intermediate layer before engineering risks are fully resolved. While CPO is theoretically superior, its full-scale adoption would require more time, so the market must reprice these “less final but easier to mass-produce and maintain” solutions.

Serenity's rebuttal is not a denial of the engineering challenges facing CPO, but rather an argument that SemiAnalysis underestimated NVIDIA's organizational capabilities within the AI hardware cycle.

Her logic is clear: while conventional hardware adoption is often hindered by yield, reliability, and customer validation, NVIDIA is not a typical customer. It is both the architect of GPU cluster systems and a core driver of networking, switching, system integration, and supply chain cadence. When scaling AI factories is constrained by network power consumption and bandwidth, NVIDIA has both the strong economic incentive and industry influence to compress traditional adoption cycles.

The evidence cited by Serenity consists of two layers. The first layer comprises publicly verifiable company guidance, including NVIDIA’s official announcement regarding the production ramp-up of Spectrum-X Photonics, and Lumentum’s mention of CPO order and delivery timelines in its Q2 FY26 update. Lumentum stated that it has received incremental CPO orders in the hundreds of millions of dollars, with deliveries scheduled for the first half of 2027. Company materials also indicate that CPO-related business is expected to enter a broader scaling phase in the second half of 2026.

The second layer is her interpretation of supply chain signals, such as Foxconn delivering optical switches to NVIDIA ahead of schedule. However, the exact scale of these signals—whether they are test units or production-grade orders—still requires additional public information for confirmation.

This is also the key point of divergence between Serenity and SemiAnalysis: SemiAnalysis places greater faith in system engineering variables naturally extending the cycle, while Serenity believes NVIDIA’s supply chain execution capabilities will make this curve steeper.

These two assessments are not entirely contradictory. NVIDIA could enable CPO to enter production and customer validation earlier, and may also accelerate the adoption of CPO in certain horizontal scaling scenarios. However, this does not automatically mean that all AI data center networks will rapidly transition to CPO by 2027. Horizontal scaling, single-rack scaling, intra-rack, inter-rack, and reliability tolerances and cost models vary across different customers, and adoption timelines may also be phased.

Serenity is refuting the overly conservative conclusion that "CPO will be significantly delayed," not proving that "CPO is now fully risk-free." For the market, this is sufficient to support a rebound logic after overselling, but insufficient to directly rewrite the aggressive revenue curve for 2027–2028 as certain.

NPO became unexpectedly important in this controversy because it sits precisely between the two frameworks of SemiAnalysis and Serenity.

It is not the opposite of CPO, nor is it simply an extension of traditional pluggable modules. The core idea of NPO is to place the optical engine on a pluggable baseboard near the ASIC, shortening the electrical signal path to gain benefits in power consumption and density, while retaining better testability, replaceability, and supply chain resilience.

If SemiAnalysis’s conservative model is closer to reality, the deep packaging of CPO will slow down due to yield, repair, and reliability issues, making NPO a more viable long-term option. This allows hyperscale cloud providers to gradually build optical interconnection operational expertise without fully assuming the risks of CPO, while also giving existing optical module and optical engine suppliers a longer window of opportunity.

If Serenity’s assessment of NVIDIA’s execution is more accurate, NPO may not disappear at all. More likely, NPO, CPO, pluggable, and copper interconnects will coexist across different network layers. NVIDIA’s own roadmap also indicates that horizontal scaling of CPO can proceed first, while some single-rack scaling scenarios may still rely on copper cables or hybrid architectures through 2027–2028.

This means investors cannot price the optical supply chain using a simple “CPO wins, everyone else loses” framework. Different technology paths benefit different segments: CPO favors high-integration optical engines, laser sources, silicon photonics, and switching chip ecosystems. NPO and pluggable interface extenders, on the other hand, may allow existing optical module manufacturers, connectors, materials suppliers, and certain light source vendors to continue benefiting from order volume and gross margin support.

Previously, the market prematurely translated the technological endgame into a single-path performance trajectory. What has now been reopened is the valuation space for intermediate paths.

This debate won’t be resolved in the short term by a single report or a set of posts. SemiAnalysis reminds the market that the challenge with CPO lies in systems engineering. Serenity reminds the market that NVIDIA’s supply chain organization may alter the traditional pace of hardware adoption. The true divergence between the two will be validated by production-grade data from late 2026 through 2028.

What matters most next isn't whether there's been a sell-off, but the scale of the sell-off. Delivering to select partners, customer evaluation, initial production, scaling up, and large-scale deployment are entirely different stages. NVIDIA's subsequent descriptions of mass production for Spectrum-X / Quantum-X Photonics, along with statements from optical suppliers like Lumentum and Coherent regarding orders, capacity, and gross margins in their earnings reports, will be more significant than the wording of a single meeting.

It is also necessary to observe field reliability and maintenance data. If the CPO’s failure rate, replacement process, yield curve, and total cost of ownership in production environments demonstrate sufficient stability, SemiAnalysis’s conservative model will be adjusted. If these data remain confined to the laboratory or small-batch validation stages, the window for NPO and pluggable modules will continue to be pushed upward by the market.

The optical chain is no longer trading on the生死 of CPO, but on the slope of the timeline. The next validation point will be whether "entry into production" can transition into sustainable volume scaling, and how quickly this scaling ultimately manifests in orders, gross margins, and customer deployment metrics.

Although SemiAnalysis expressed concerns about CPO technology over the next two years, they still identified five favored semiconductor subsectors:
Copper / AEC / ACC;
Pluggable optical / DSP;
CPO test equipment;
Power gray space / UPS continuity;
Board-level VRM / Silicon-based Power / Passive Components

The specific assets are illustrated in the image below for your reference.