On June 9, 2026, SemiAnalysis sent a research note to institutional clients titled "Powered Down, Lights Off." By the end of the trading session, AAOI had dropped 14%, COHR 11%, GLW 9%, MRVL 9%, LITE 8%, and CIEN 7%. The entire optical networking sector got mauled in a single day.
But if you actually read the report, it never says Co-Packaged Optics (CPO) is the wrong direction. It says the ramp is slower than the market expected. The market heard "slower" as "dead."
What did the report actually say? Are the counterarguments convincing? And the more practical question: if CPO really is delayed, is Near-Packaged Optics (NPO) the better bet?
What the SemiAnalysis Report Says
The report covers two technology threads: 800VDC power distribution and CPO optical integration. Bundled together as an "AI infrastructure delay" narrative, they triggered a sector-wide selloff.
800VDC: NVIDIA's Single-End Approach Slips; Cloud-Driven ±400VDC Is Fine
800VDC is not one architecture but two.
NVIDIA's single-end 800VDC was originally targeted for 2027 adoption. The report pushes it to 2028 and beyond. The reasoning is straightforward: Rubin base hardware does not need 800VDC and still runs on 50V. Only higher-power chips like Rubin Ultra and Feynman genuinely require 800VDC, and Rubin Ultra's design won't be finalized until the second half of 2026. Worse, converting grid power from 350-450V up to 800V and then back down to 50V is energy-inefficient. Cloud providers are unhappy with this conversion chain.
The other track is ±400VDC, driven by hyperscale cloud providers for their custom silicon. This track is completely unaffected. Mass production begins in late 2026, with volume in Q1 2027. Its power distribution modules will also be compatible with NVIDIA hardware in the future.
The report frames 800VDC as "delayed but not dead." When individual compute modules exceed 15kW power draw, native high-voltage DC distribution's efficiency advantage becomes compelling. It's a timing question, not a feasibility question.
CPO: Yield Arithmetic and the Repairability Problem
The report splits its CPO analysis into scale-out (switch-side) and scale-up (GPU-side).
For scale-out, the core obstacle is yield. NVIDIA's Spectrum 6 CPO switch integrates 32 COUPE optical engines on a single chip. The report assumes 95% per-engine attach yield (already optimistic), which produces a system yield of 0.95 raised to the 32nd power: approximately 19%. To reach an 85% system yield break-even, per-engine yield must hit 99.5%.
Beyond yield, there's a repairability crisis. The optical engines are soldered directly onto the substrate, co-packaged with ASICs worth tens of thousands of dollars. When a laser degrades or a fiber is damaged, the entire board must be pulled and sent back for repair. Traditional pluggable modules can be swapped in 30 seconds. CPO failures could mean week-long downtime. Cloud providers are deeply concerned about this operational model.
There's also a technical detail: the report discloses that Spectrum 6's second-generation optical engine has an onboard insertion loss exceeding 3.5dB, consuming the entire optical channel tolerance. Performance is actually worse than the first generation. NVIDIA and TSMC have not yet identified the root cause and are redesigning the assembly process.
For scale-up, the report is more aggressive. The market expected mass production in 2027-2028; the report pushes it to 2029. AWS, AMD's Feynman, and other core projects cluster around 2029. Embedded interposer optical engine technology won't mature until then. Some NVL576 units will ship in 2027-2028, but only for switch interconnect, not GPU-facing.
The Report Also Names NPO as an Accelerator
The report isn't purely bearish. It explicitly notes that some NPO (Near-Packaged Optics) projects may accelerate. NPO positions the optical engine near the ASIC but not inside the same package, keeping it independently replaceable. Implementation barriers are significantly lower.
The report maintains positive views on Amphenol, Vertiv, and Legrand. It also identifies Chinese optical module makers Zhongji Innolight and Eoptolink as beneficiaries of CPO delays. It takes a cautious stance on Lumentum, Himax, Navitas, and Wolfspeed.
Four Lines of Counterattack
Within 48 hours of the report's release, at least four sets of rebuttals emerged.
Rebuttal 1: 0.95^32 Is Magic Math
GlobalSemiResearch published a lengthy SubStack rebuttal with a direct title: "Co-Packaged Optics Is Not Delayed. SemiAnalysis Is Just Wrong."
The core attack is concentrated on one formula. The 0.95^32 calculation freezes yield at a static snapshot, assuming no screening, no binning, no redundancy, and no yield learning curve. Real advanced packaging lines use known-good-die screening: defective units are culled before they reach the substrate. The scenario of 32 engines randomly drawn from a 95%-yield batch doesn't happen on a real production line.
The author calls the figure "not analysis but conjecture — a frightening number reverse-engineered to fit a bearish headline."
This rebuttal has merit but also a weakness. Learning curves are real, but how fast do they climb? If going from 95% to 99.5% takes 18 months, SemiAnalysis's "2028-2029" could actually be right.
Rebuttal 2: Capacity Signals Contradict the Conclusion
Multiple supply chain analysts pointed out a simple inconsistency: if CPO is truly delayed at scale, certain investments make no sense.
NVIDIA increased its laser orders from 40 million to 100 million units. Lumentum booked an entire final assembly production line. These are committed investments, not PowerPoint promises. Nobody expands capacity to this degree for a technology that won't ramp until 2029.
This argument is powerful but not conclusive. Ordering components to prepare for volume production is different from achieving volume production yields. NVIDIA may be building capacity ahead of the yield learning curve.
Rebuttal 3: NVIDIA Officially Denies Delays
Gilad Shainer, NVIDIA's SVP of Networking, stated in a Computex interview that CPO shipments have no delays. Spectrum-X Ethernet Photonics is already in mass production. CoreWeave, Lambda, Meta, Microsoft, and Oracle are first customers.
One side says "shipping"; the other says "delayed." But there's a subtle distinction: Shainer said "shipping," not "shipping at volume." The gap between first customer shipments and yield-profitable mass production can be over a year.
Rebuttal 4: The Market Misread the Report
PhotonCap's SubStack offered a different angle: the market got it wrong.
SemiAnalysis said "ramp is slower than expected," not "CPO is dead." Between first production and broad availability sits a packaging capacity ramp. SA pointed at the speed of that ramp. Moreover, AAOI, LITE, and COHR operate across pluggable modules, NPO, and CPO. Treating them as pure CPO plays and dumping them is a category error.
AAOI was up over 400% year-to-date. LITE was up over 150%. Their valuations priced in a perfect CPO ramp with zero room for error. SemiAnalysis didn't break the thesis; it broke the trade.
CPO vs NPO: Not Replacement, but Time Window
If CPO is genuinely delayed, is NPO better? It requires unpacking.
Physical Differences
NPO places the optical engine adjacent to the ASIC on the same high-performance substrate but outside the ASIC's package body. The electrical channel between them runs roughly 1-5cm.
CPO places the optical engine on the same package substrate as the ASIC. The physical distance approaches zero. Per OIF standards, optical engine-to-ASIC distance is within 50mm, with channel loss limited to 10dB.
This distance difference drives several key metrics. CPO has optimal power efficiency: NVIDIA's CPO solution consumes about 4-5W per 800G of bandwidth, compared to 16-17W for traditional 800G pluggable modules, a roughly 70% reduction. NPO sits between the two. CPO also has the highest density potential since optical engines live inside the package rather than consuming board space.
But NPO offers three things CPO cannot provide.
Repairability
NPO's optical engine is an independent module. While most NPO implementations are soldered to the board, its independence means fault isolation is easier. A failed optical engine means replacing the optical engine; the ASIC survives.
CPO offers no such option. 32 optical engines share a package with the ASIC. Any single engine failure can affect the entire package. This is the source of SemiAnalysis's "entire board goes back for repair" claim.
In data center operations, this matters enormously. A large cluster has tens of thousands of switches. If each switch's mean time to repair (MTTR) goes from 30 seconds to one week, the availability model must be rewritten.
Thermal Isolation
NPO's optical engine is independently packaged and not directly exposed to the ASIC's thermal environment. Lasers are temperature-sensitive; thermal drift causes wavelength shifts that directly impact signal quality. CPO places optical engines next to ASICs drawing 500-1000W. Maintaining laser thermal stability in that environment is a genuine engineering challenge.
Supply Chain Value Distribution
This point is rarely discussed but crucial for understanding industry dynamics.
Under CPO, optical engines are packaged into switch chips under NVIDIA and TSMC's leadership. Optical module makers (Zhongji Innolight, Eoptolink, Coherent, Lumentum) are demoted from system integrators to component suppliers. They sell chips and components, not systems.
Under NPO, optical engines exist as independent modules. Optical module makers participate in the full chain: design, assembly, and delivery. Per supply chain technical leads, at equivalent data rates, NPO lets optical module makers capture roughly double the value content of CPO.
This means the entire optical module supply chain has strong economic incentives to push NPO. This isn't technology preference; it's survival.
Assessment: NPO Is the Pragmatic Choice for 2026-2028; CPO Is the Endgame for 2029+
The laws of physics dictate that CPO's architectural advantages are irreversible. Zero distance means minimum power, minimum latency, maximum density. As per-port speeds push from 1.6T toward 3.2T or even 6.4T, copper channels and near-field electrical signaling will hit physical walls. CPO is the only answer.
But in the 2026-2028 window, NPO is the more pragmatic path. Yield, repairability, thermal stability, and supply chain economics all point the same direction.
The more likely evolution is not "CPO replaces NPO" but coexistence across different scenarios. Scale-out (switch-to-switch networking) migrates to NPO first, then gradually transitions to CPO. Scale-up (GPU-to-GPU interconnect), which demands extreme latency and bandwidth density, may need CPO from the start, but on a later timeline.
My Read
SemiAnalysis's direction is correct. Yield is a real bottleneck. Repairability is a genuine pain point. The Spectrum 6 insertion loss regression, if accurate, is serious. But the "delayed to 2029" conclusion is too absolute.
The counterarguments' capacity signals are persuasive. NVIDIA adding laser orders to 100 million units and Lumentum booking an entire production line are real money behind real conviction. But "shipping" and "volume shipping at profitable yields" are different things with a wide gray zone between them.
The truth is probably in the middle. Small-volume shipments and yield ramping begin in 2027. Scale production completes between 2028 and 2029. Not the "overnight explosion in 2027" the market originally priced, and not the "nothing until 2029" SemiAnalysis suggested.
The real cause of the crash wasn't SemiAnalysis's persuasiveness. It was that optical stocks had run too far, too fast, with positions too crowded. AAOI up 400%+ year-to-date leaves zero margin for error. When a crowded trade meets a "not that fast" signal, panic turns a timing problem into a directional question. The direction holds; the tempo needs repricing.
The NPO judgment is more certain. At OFC 2026, Google and other leading cloud providers announced NPO deployment plans. Optical module makers have economic incentives to push it. Cloud providers have operational reasons to choose it. Implementation barriers are lower than CPO. In 2026-2028, NPO's certainty is far higher than CPO's. This isn't betting CPO will fail; it's acknowledging that engineering reality lets intermediate solutions live longer than expected. Much like 5G Non-Standalone (NSA): everyone knew Standalone (SA) was the endgame, but NSA remains dominant today.
Disclaimer: This article is based on the SemiAnalysis report "Powered Down, Lights Off" (June 9, 2026), rebuttals from GlobalSemiResearch and PhotonCap on SubStack, Chinese-language analyses from TechFlow and Futubull community, and public information from OFC 2026 and Computex 2026. Not investment advice. Data as of June 12, 2026.