Nvidia Kyber delayed to 2028, winners and losers in the supply chain.

Written by: Chaoxiang Research

On July 6, semiconductor research firm SemiAnalysis released a report: Jensen Huang’s hand-demonstrated Kyber NVL144 cabinet at GTC 2026 suffered a major setback after only three months of its debut. Its deployment timeline was delayed by more than 12 months, pushing it to 2028. The supporting NVL72x2 back-to-back cabinet architecture was directly canceled, and the NVLink expansion domain of Rubin Ultra was therefore compressed.

This news has become the single biggest source of negative catalysts for the technology sector today.

Taken in isolation, this is a product postponement. If you stretch the timeline to 30 days, you can see a clear retreat curve.

On June 10, SemiAnalysis issued a report to institutional clients, noting that large-scale shipments of NVIDIA’s native 800VDC power architecture would be delayed to after 2028, and that CPO (co-packaged optics) mass production could be postponed to 2028 and even 2029. The U.S. optical communications sector took an immediate hit: AAOI fell as much as 17% at its deepest point in a single day, Lumentum dropped by about 8%, and Himax, Navitas, and Wolfspeed were all under concurrent pressure.

On June 30, SemiAnalysis spoke again: the four-chip version of Rubin Ultra, which was trumpeted at GTC 2026, was canceled due to manufacturing execution risks, and it was switched to a dual-chip design. The actual compute and memory bandwidth were about half of the original. Behind this is that TSMC’s CoWoS-L advanced packaging has hit physical limits at the four-chip, multi-reticle scale. The successor CoPoS will not be able to reach mass production until the end of 2028 at the earliest.

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

What is Kyber, and why is it so hard?

Kyber is the next-generation cabinet architecture designed by NVIDIA for Rubin Ultra and the subsequent Feynman generation. Its core move is to rotate the compute tray 90 degrees and insert it vertically into the cabinet—like books on a shelf—and then replace tens of thousands of copper-cable connections inside the cabinet with an orthogonal backplane PCB. According to the original GTC 2025 specifications, a single cabinet’s power consumption is as high as 600 kilowatts, supported by a new 800VDC power system.

This backplane is the hardest PCB to manufacture in the entire industry. According to a Jefferies research note, it requires 78 layers of M9-grade material. Meanwhile, based on industry observers’ teardown of the GTC 2026 showcase models, the connector pin count on a single mid-plane exceeds 10,000, and the total number of NVLink pins for the entire cabinet exceeds 87,000. Any bending of any single pin could cause the entire board to be scrapped. Only two or three suppliers worldwide have the capability to mass-produce at this specification.

Jefferies had actually warned as early as June 22: the Kyber backplane PCB solution would most likely be delayed to 2028, and in the worst case could be completely canceled. Based on that, it lowered its global AI PCB market size forecasts for 2027 and 2028 by 5% and 11%, respectively, and lowered CCL (copper-clad laminate) by 8% and 16%. On June 23, combined with rumors such as “NVIDIA asked PCB manufacturers to cut prices by 10%” that were later debunked, panic selling hit the A-share and Hong Kong-listed PCB sectors. Today’s SemiAnalysis report effectively puts a confirmation stamp on that warning.

And the cancellation of NVL72x2 means NVIDIA has abandoned the transitional approach of expanding the NVLink domain by splicing two cabinets back-to-back. In 2027, Rubin Ultra will likely retreat to the mature Oberon architecture (i.e., the NVL72 form factor), with its expansion capability returning to the framework of the previous generation.

Jensen Huang has emphasized externally that NVIDIA is the first technology company in history to publish a four-generation product roadmap all at once. The original intention behind revealing it early was to give the supply chain sufficient preparation time: data center site selection, power system retrofits, and liquid cooling solutions all require upfront investment measured in years.

The side effects have become concentrated and visible this year: the roadmap itself has turned into a tradable asset. The market prices the entire industrial chain according to the PPT timeline—optical modules modeled by CPO penetration, PCB valuations based on backplane ramp-up pace, and power-supply manufacturers’ production schedules based on the 800VDC switching timeline. As physical laws reclaim these commitments one by one, each revision corresponds to a sector-level repricing. Over the past 30 days, the three sectors—optical communications, PCB, and power supplies—have taken turns going through this process.

Winners and losers along the industrial chain

What the postponement changes is the rhythm—and, along the way, it also rewrites the list of winners.

Copper-cable suppliers receive a “reprieve.” With the lifecycle of the Oberon architecture extended, the copper-cable demand that was originally planned to be replaced by backplane PCBs is preserved. In the framework of SemiAnalysis, connector makers such as Amphenol are listed as relatively advantaged beneficiaries, and Vertiv and Legrand also receive a somewhat positive assessment.

The logic for upstream materials is the toughest. The tight supply of fiberglass fabric and CCL has nothing to do with Kyber. It is driven by demand across the whole industry. Copper-clad laminates have already risen four times within half a year. Kyber’s delay only changes the demand structure, not the total demand, and pricing power still remains in the hands of material suppliers.

The PCB manufacturing side bears the most direct pressure. Among those stuck in the middle, the situation is worst for manufacturers in the mid-tier: high-end players have deep technical know-how and customer stickiness, allowing them to follow specification upgrades; low-end capacity has a cost advantage; and mid-tier firms are squeezed from both directions, accelerating elimination in a knockout competition.

The overall time window for the optical communications and CPO industrial chain has shifted backward. Sidecar shipments that rely on the Rubin Ultra and Kyber platforms are pushed to the 2028 window. SemiAnalysis maintains a cautious stance toward Lumentum, Himax, Navitas, and Wolfspeed, while also pointing out that some NPO (near-packaged optics) projects may accelerate in the opposite direction.

The bigger narrative shock lands on NVIDIA itself. With Rubin Ultra’s specifications essentially cut by half combined with Kyber’s postponement, some analysts interpret this as a signal that NVIDIA’s performance moat is starting to erode; AMD and the TPU ecosystem are identified as potential marginal beneficiaries.

Although this conclusion currently lacks sufficient evidence, the fact that it is entering mainstream discussion is itself a change.

Going forward, watch for two confirmation signals: whether NVIDIA positively addresses the Kyber and Rubin Ultra timelines during its next earnings call; and whether order guidance from Taiwanese ODMs and PCB manufacturers shows structural adjustments.

At the same time, it is also necessary to note that the postponement and cancellation information discussed in this article comes from third-party channels such as SemiAnalysis and Jefferies. NVIDIA has not officially confirmed it. Supply-chain information has always been subject to repeated revisions. NVIDIA’s network business executives previously also gave optimistic remarks about the CPO timeline that ran contrary to third-party research. Until official wording is in place, treat this as a higher-probability scenario—not as a settled fact.