Nvidia's Kyber Rack for Rubin Delayed to 2028

Nvidia has delayed the launch of its Kyber NVL144 rack, originally planned for 2027, to 2028, according to a report from SemiAnalysis picked up by Tom's Hardware. The delay, of more than 12 months, is due to manufacturing issues in the orthogonal PCB backplane, a 78-layer board that connects the 144 Rubin Ultra packages without cables. Additionally, Nvidia has canceled the NVL72x2 rack, a temporary solution that linked two Oberon racks to achieve Kyber's density, due to opposition from major customers.

What happened?

The Kyber delay, confirmed by a SemiAnalysis thread on X (formerly Twitter) on July 5, 2026, represents a significant setback in Nvidia's roadmap, just three months after Jensen Huang showcased the orthogonal backplane on stage at GTC in March 2026. According to Tom's Hardware, the core issue lies in the backplane's manufacturing: a printed circuit board (PCB) with 78 layers, nearly one square meter, combining three 26-layer sections laminated together. Copper traces have a spacing of 25 micrometers or less, and impedance must be maintained within a 5% tolerance for the 448 Gb/s NVLink signals to function correctly. This technical complexity has exceeded current manufacturing capabilities, causing the delay.

Furthermore, Nvidia has canceled the NVL72x2 rack, an interim solution that linked two Oberon racks to offer similar density to Kyber. According to SemiAnalysis, major customers (hyperscalers like Microsoft, Amazon, Google) opposed this solution because it complicated operation and thermal management, preferring to wait for the final Kyber design. The cancellation leaves Nvidia without an alternative for scaling Rubin Ultra in 2027, limiting the growth of large-scale AI clusters.

This is not the first setback in Nvidia's roadmap. Rubin Ultra already underwent a redesign, from four chiplets to two, to simplify manufacturing. However, the Kyber delay is more severe because it directly affects packaging and interconnection, areas where Nvidia had bet on copper instead of optics. The company plans to adopt optical interconnects (optical NVLink) in the Feynman generation, expected in 2028, but until then relies on the copper backplane, which now faces manufacturing challenges.

Why is this important?

Kyber is key to Nvidia's strategy for large-scale AI: it doubles the GPU density of current racks (144 vs 72) and uses a copper backplane for 448 Gb/s NVLink links, eliminating thousands of cables. According to Tom's Hardware, the orthogonal backplane replaces the cable harnesses of previous racks with a rigid board that carries the entire copper NVLink fabric. This reduces latency and cabling complexity but introduces challenges in signal integrity, power delivery, and thermal management. Without Kyber or NVL72x2, Rubin Ultra's scaling domain in 2027 will be limited to what Oberon already offers (72 GPUs per rack), slowing the growth of AI clusters. Hyperscalers that planned to double their per-rack capacity will need to adjust their plans, potentially slowing the deployment of large language models and other intensive AI workloads.

Historically, Nvidia has maintained its AI leadership through rapid hardware iterations (Hopper, Blackwell, Rubin). However, this delay recalls previous issues, such as delays in HBM3E memory modules or packaging challenges in Blackwell. The difference is that Kyber is a critical infrastructure component, not just a chip, and its delay affects the entire Rubin Ultra platform. Additionally, the cancellation of NVL72x2 shows that customers reject interim solutions that complicate operations, which may indicate market maturity demanding final, robust solutions.

Market consequences

Hyperscalers that expected to double their per-rack capacity in 2027 will have to settle for Oberon or wait until 2028. This could benefit competitors like AMD or Intel, which have windows of opportunity. AMD, with its Instinct MI400 platform and 144-GPU rack, could capitalize on the delay if it scales in time. Intel, with its Falcon Shores architecture, could also offer alternatives. Additionally, the Kyber delay postpones the adoption of Rubin Ultra, which already underwent a redesign from four to two chiplets. Nvidia's reliance on complex PCB backplanes and the lack of a mature optical alternative until the Feynman generation add uncertainty. The large-scale AI market will need to adjust its expansion plans, potentially slowing sector growth in 2027.

According to SemiAnalysis estimates, Rubin Ultra's scaling capacity is halved until 2028, which could affect delivery timelines for next-generation AI models. However, Nvidia still dominates the market with its current products (Hopper, Blackwell), which are unaffected. The short-term impact is limited, but long-term confidence in Nvidia's roadmap could be affected if delays recur.

What readers should know

The delay does not affect current products (Hopper, Blackwell), but it does impact the 2027 roadmap. Nvidia is betting on the copper backplane to reduce costs and latency, but manufacturing a 78-layer board with 5% impedance tolerances is extremely difficult. According to Tom's Hardware, the backplane has an area close to one square meter and requires precision lamination and drilling processes that are beyond current PCB manufacturers' capabilities. The cancellation of NVL72x2 suggests customers prefer to wait for the final solution, indicating market maturity demanding stability. The large-scale AI market will need to adjust its expansion plans, but Nvidia remains the undisputed leader. Investors and analysts should closely monitor advances in backplane manufacturing and Nvidia's optical alternatives for the Feynman generation.