Full Breakdown
Nvidia's Kyber Rack System Delayed to 2028 Amid Manufacturing Hurdles
7/6/2026, 8:31:36 PM
Core Event: Kyber Rack Architecture Pushed to 2028
Nvidia’s Kyber NVL144 rack, intended to hold 144 Rubin Ultra GPUs, has been pushed from a 2027 launch to 2028. SemiAnalysis cites manufacturability problems with the system’s multi-layer PCB midplane as the cause. The larger NVL576 configuration linking eight racks is also expected to be delayed.
Background and Timeline
Nvidia has pursued an annual AI-hardware cadence for three years. CEO Jensen Huang unveiled Kyber at GTC on March 16 2026. SemiAnalysis announced the midplane issue and a 2028 target on July 5 2026. Meanwhile, current Rubin systems entered full production and began shipping in fall 2026 to eight cloud partners, including AWS, Azure and Google Cloud.
Technical Specs and Data
The NVL144 cabinet integrates 144 GPUs via an all-copper NVLink interconnect. A larger NVL576 variant would connect eight cabinets optically. SemiAnalysis notes the NVL72x2 back-to-back rack, meant as an interim scaling bridge, has been cancelled. Nvidia expects data-center compute revenue to be about 20 % above consensus in H2 FY2027.
Official Statements & Responses
Nvidia declined comment to CNBC about the delay. It confirmed that current Rubin systems are in full production and will ship to the eight cloud partners. No official response was given on the cancelled NVL72x2 design.
Criticism & Opposition
Cloud providers and hyperscalers rejected the backup plan of bolting two current racks together, calling it odd and an operational burden. SemiAnalysis cites this pushback as a key reason for cancelling the interim solution.
Impact on AI Infrastructure and Crypto Mining
The delay narrows scaling options for top AI labs, potentially opening market share to rivals like AMD and Google, whose in-house chips already attract customers. Nvidia GPUs also power a sizable share of crypto mining; a disrupted roadmap could affect GPU supply and pricing for miners.
Conflicting Reports & Gaps
All sources agree on a 2028 launch, but Nvidia’s silence leaves the exact manufacturing timeline and mitigation plans unclear. Specific midplane fabrication challenges at TSMC are not disclosed.
Verbatim Quotes
- “Kyber NVL144 rack architecture has been delayed to 2028 as the PCB midplane remains challenging from a manufacturability standpoint,” — SemiAnalysis
- “It has since been cancelled due to heavy pushback from CSPs [cloud service providers] and hyperscalers over its odd design and heavy operational burden,” — SemiAnalysis
- “no proven solution to expand the scale-up world size for Rubin Ultra,” — SemiAnalysis
- “We're seeing the limits of what's physically possible,” — Supply chain executive (background)
Outlook: What Lies Ahead
Rubin Ultra’s progress now depends on co-packaged optics, which embed optical components directly into chip packages. Delays could rewrite the AI-hardware scaling roadmap and push hyperscalers toward alternative vendors.
