Nvidia's Kyber NVL144 delay and Rubin Ultra cancellation: A fact-checked breakdown

Nvidia's product roadmap hit a significant snag in mid-2026 when reports emerged that its next-generation Kyber NVL144 AI server rack system had been delayed by more than a year, and the flagship four-die Rubin Ultra GPU was scrapped entirely. The news, broken by semiconductor research firm SemiAnalysis and covered by CNBC, Tom's Hardware, and Jefferies, sent Asian PCB-related stocks into a tailspin and opened a potential window for rivals AMD and Google. Below is a fact-checked breakdown of what happened, why it matters, and what remains uncertain.

Kyber NVL144 delayed to 2028 over PCB midplane manufacturing

What happened: Nvidia's Kyber rack-scale architecture — designed to house its Rubin Ultra GPUs — has been delayed by more than 12 months, pushing its arrival to 2028 . The news broke on July 6, 2026, when SemiAnalysis posted six tweets revealing the setback just three months after CEO Jensen Huang demonstrated Kyber NVL144 at GTC 2026 .

Root cause: The delay stems from manufacturing difficulties with a critical hardware component called the PCB midplane (also referred to by Nvidia as the "orthogonal backplane") . SemiAnalysis stated the midplane "remains challenging from a manufacturing feasibility standpoint" . Jefferies analysts, in a June 22 research note, confirmed that Kyber/backplane PCB deployment is likely delayed due to intra-rack connectivity bottlenecks, with 2027 remaining on the prior Oberon architecture .

Scope of impact: The larger NVL576 variant (connecting 8x Oberon racks via co-packaged optics, or CPO, on NVSwitches) may also be delayed or limited to small-batch production due to CPO challenges . Nvidia's CPO NVSwitch is not expected until the Feynman generation .

Supply chain revisions: Jefferies estimates the Kyber delay will reduce the global AI PCB market size by 5% in 2027 and 11% in 2028, and the CCL (Copper Clad Laminate) market by 8% and 16%, respectively .

The four-die Rubin Ultra is canceled, replaced by a smaller chip

What happened: The original four-die (quad-chiplet) Rubin Ultra GPU — announced at GTC 2026 — was canceled roughly three months later, in late June 2026 . SemiAnalysis reported the cancellation on June 30, and it was covered by Tom's Hardware, Igor's Lab, and others .

Replacement: Nvidia replaced the four-die design with a smaller, dual-die package that carries the same "Rubin Ultra" name but delivers roughly half the computing performance of the original . The revised chip is about half the physical size .

Reason: Manufacturing execution issues — specifically that the four-chiplet packaging hit a physical/technical wall that could not be resolved in time .

Roadmap uncertainty: With the quad-die design dead and Kyber delayed, Nvidia currently has "no proven solution to scale Rubin Ultra's scale-up domain," according to SemiAnalysis, which opens a window for competitors including AMD's MI500X and Google's TPUv8i Broadfly to surpass Nvidia in scale-up capability .

Market impact: Asian PCB stocks sell off sharply

The SemiAnalysis report triggered a sharp sell-off in Asian PCB-related stocks on July 6, 2026. Hong Kong-listed stocks were hit hardest:

• Kingboard Laminates (1888.HK) plunged over 18% (after earlier gaining nearly 9%)
• Circuit Fabology (9630.HK) fell over 14%
• Shengyi Technology (2476.HK) and Kingboard Holdings (0148.HK) each fell over 11%
• Guanghe Technology (1989.HK) dropped over 10%
• Han's CNC (3200.HK) fell over 7%

In Taiwan, shares of Taiwan Union Technology (2383) fell 10%, Nan Ya PCB (8046) dropped 9.7%, and Unimicron (3037) fell 7.4%. South Korea's Samsung Electro-Mechanics fell 11%; Japan's Ibiden slumped 9.9% .

A South Korea-based NH Investment & Securities equity head told reporters that "regional tech stocks are weak today, apparently affected by the SemiAnalysis report," noting the Kyber delay, the NVL72x2 cancellation, and the downgraded Rubin Ultra as key factors increasing uncertainty around Nvidia's scaling roadmap .

Who benefits: AMD, Google, and the longer-term AI capex view

AMD: SemiAnalysis explicitly identified AMD's upcoming MI500X as a direct beneficiary, stating the delays "leave room for competitors like AMD MI500X or TPUv8i Broadfly to surpass Rubin Ultra in scale-up capability" . The report also noted that Nvidia faces "structural market share erosion from Trainium, TPUs, and AMD" . AMD stock had surged 5.4% in August 2025 on earlier Rubin delay rumors , and Fubon Research had turned bullish on AMD citing Nvidia's redesign challenges .

Google: Google's TPUv8i Broadfly is named alongside AMD MI500X as a platform that could overtake Nvidia in scale-up capabilities during the gap created by the Rubin Ultra and Kyber delays . Google has already deployed rack-scale AI products of its own .

Broader AI infrastructure costs remain high: Despite the execution stumbles, Goldman Sachs maintained a bullish outlook on the AI infrastructure investment cycle extending through 2028, issuing Buy ratings on multiple stocks across the supply chain . Jefferies also maintained a Buy rating on Nvidia and raised its price target to $300 (42% upside from around $210.69), suggesting the delay does not fundamentally alter the long-term AI capex thesis .

However, a separate analysis by Dan Herbatschek, CEO of Ramsey Theory Group, highlights a burgeoning crisis in AI infrastructure costs, with enterprises reportedly underestimating these expenses by 30% or more . Morgan Stanley estimated the cooling system for a single Nvidia GB300 NVL72 rack carries a BOM cost of nearly $50,000, rising to an estimated $55,710 for the Vera Rubin NVL144 platform .

Key caution: Nvidia has not officially confirmed

The reports originate from SemiAnalysis, a respected semiconductor research firm, and have been picked up by CNBC, Tom's Hardware, and Jefferies. However, Nvidia has not officially commented or confirmed the Kyber delay or the Rubin Ultra cancellation. PCB manufacturers have previously pushed back against related rumors as "exaggerated and misinterpreted" . The situation remains unconfirmed by Nvidia's official communications.