Written by:潮向研究

On July 6, semiconductor research institution SemiAnalysis released a report: Jensen Huang personally showcased the Kyber NVL144 cabinet at GTC 2026, which encountered significant setbacks just three months after its debut, with the landing time delayed by more than 12 months, extended to 2028. The supporting NVL72x2 back-to-back cabinet architecture was directly canceled, resulting in the compression of the NVLink expansion territory for Rubin Ultra.

This news became the largest single bearish factor in the tech sector today.

Isolated, this is a product delay. Extending the timeline to 30 days reveals a clear retreat curve.

On June 10, SemiAnalysis issued a report to institutional clients, indicating that large-scale shipments of Nvidia's native 800VDC power supply architecture would be postponed until after 2028, and the mass production of CPO (Co-Packaged Optics) might be delayed until 2028 or even 2029. The U.S. stock market's optical communications sector reacted sharply, with AAOI experiencing a single-day drop of up to 17%, Lumentum falling about 8%, and Himax, Navitas, and Wolfspeed also coming under pressure.

On June 30, SemiAnalysis spoke again: The four-chip version of Rubin Ultra, which was prominently presented at GTC 2026, was canceled due to manufacturing execution risks and changed to a dual-chip design, with actual computing power and memory bandwidth being about half of the original version. Behind this is TSMC's CoWoS-L advanced packaging reaching physical limits in four-chip, multi-fold photomask scales, with the successor CoPoS not being able to achieve mass production until at least the end of 2028.

On July 6, it was Kyber's turn, three reports, three retreats.

What is Kyber and why is it so difficult?

Kyber is the next-generation cabinet architecture designed by Nvidia for Rubin Ultra and the subsequent Feynman generation. Its core action is to rotate the compute trays 90 degrees to be vertically inserted into the cabinet, like books on a bookshelf, and replace the thousands of copper cable connections inside the cabinet with an orthogonal backplane PCB. According to the initial specifications from GTC 2025, the power consumption of a single cabinet reaches as high as 600 kilowatts, supported by a completely new 800VDC power supply system.

This backplane is the most difficult PCB to manufacture in the entire industry. According to a Jefferies research report, it requires 78 layers of M9-level material; and according to industry observers who dismantled the display from GTC 2026, the number of connector pins on a single mid-board exceeds ten thousand, and the total number of NVLink pins in the entire cabinet exceeds 87,000. Any bending of a pin could lead to the scrapping of the entire board. There are only two or three suppliers worldwide capable of mass production at this specification.

Jefferies actually issued a warning on June 22: The Kyber backplane PCB solution is likely to be delayed until 2028, and in the worst case, completely canceled, resulting in a 5% and 11% downward revision of the global AI PCB market size forecasts for 2027 and 2028, respectively, and an 8% and 16% downward revision for CCL (Copper-Clad Laminate). On June 23, combined with rumors that "Nvidia asked PCB manufacturers to lower prices by 10%" and subsequently disproved, panic selling occurred in the A-shares and Hong Kong stock PCB sectors. Today's report from SemiAnalysis essentially confirmed this warning.

The cancellation of NVL72x2 means that Nvidia has abandoned the transitional plan of expanding the NVLink domain with two back-to-back cabinets. Rubin Ultra is likely to revert to the mature Oberon architecture (i.e., NVL72 form) in 2027, restoring the expansion capability to the framework of the previous generation.

Jensen Huang has emphasized that Nvidia is the first technology company in history to announce the product roadmap for four generations all at once. The original intention of the early announcement was to give the supply chain ample preparation time: data center site selection, power supply transformation, and liquid cooling solutions all require prior investments measured in years.

The side effects have concentrated this year: the roadmap itself has become a tradable asset. The market prices the entire industry chain according to the schedule on the PPT, with optical modules modeled based on CPO penetration rate, PCBs valued according to the pace of backplane ramp-up, and power supply manufacturers scheduling production according to the switching point to 800VDC. As physical laws gradually retract these promises, each correction corresponds to a sector-level repricing. Over the past 30 days, the optical communications, PCB, and power supply sectors have each gone through this process in succession.

Winners and losers in the supply chain

The delay changes the rhythm and inadvertently rewrites the list of winners.

Copper cable manufacturers have been granted a "reprieve." The lifecycle of the Oberon architecture has been extended, allowing the copper cable demand originally planned to be replaced by backplane PCBs to be retained. Connector manufacturers like Amphenol are listed as relatively beneficial in the SemiAnalysis framework, and Vertiv and Legrand also received a relatively positive assessment.

The logic of upstream materials is the hardest. The tight supply of fiberglass cloth and CCL is unrelated to Kyber; it is driven by demand across the industry, and copper-clad laminates have already seen four consecutive price increases within six months. The postponement of Kyber only changes the demand structure, not the total demand, and the pricing power remains in the hands of material suppliers.

The PCB manufacturing side bears the most direct pressure, with manufacturers caught in the middle facing the worst predicament: high-end players have technical depth and customer stickiness, allowing them to follow specification upgrades; low-end capacities have cost advantages; mid-range manufacturers are caught in the middle, facing accelerated elimination.

The overall time window of the optical communications and CPO supply chains has shifted. Shipments relying on the Rubin Ultra and Kyber platforms have been postponed to the 2028 window, with SemiAnalysis maintaining a cautious attitude towards Lumentum, Himax, Navitas, and Wolfspeed while indicating that some NPO (Near-Package Optics) projects may accelerate in reverse.

The larger narrative impact is on Nvidia itself. The halving of Rubin Ultra specifications combined with the postponement of Kyber has led some analysts to interpret it as a signal of wear in Nvidia's performance moat, with AMD and Google's TPU ecosystem being identified as potential marginal beneficiaries.

This conclusion currently lacks sufficient evidence, but it is beginning to enter mainstream discussion, which itself is a change.

Subsequent attention should be paid to two confirmation signals: whether Nvidia directly addresses the timelines of Kyber and Rubin Ultra in the next earnings call; whether order guidance from Taiwanese ODMs and PCB manufacturers shows structural adjustments.

At the same time, it should be noted that the postponement and cancellation information in this article comes from third-party sources such as SemiAnalysis and Jefferies, and Nvidia has not yet confirmed. Supply chain information has a history of potential revisions, and Nvidia's senior executives in network business have previously expressed optimistic statements about the CPO timeline that contradict third-party research. Until the official stance is established, treat it as a high-probability scenario rather than a foregone conclusion.