Nvidia Delays Kyber NVL144 Rack to 2028

This delay matters because rack-level design problems propagate into procurement, deployment timelines, and model-capacity planning for teams running large-scale training workloads. When a vendor-level scale-up path is uncertain, cloud providers and large labs re-evaluate cluster buildouts, spare-part inventories, and deployment automation assumptions.

According to research firm SemiAnalysis (reported by CNBC and Tom's Hardware), Nvidia's Kyber NVL144 rack architecture has been pushed into 2028 because the centralized circuit board that ties vertical compute trays together, described as the PCB midplane, "remains challenging from a manufacturability standpoint." CNBC reproduces SemiAnalysis' phrasing: "Kyber NVL144 rack architecture has been delayed to 2028 as the PCB midplane remains challenging from a manufacturability standpoint." SemiAnalysis also reported that the proposed interim approach, the NVL72x2 back-to-back rack architecture, was dropped after cloud providers pushed back on its operational complexity (reported by The Next Web and Tom's Hardware). SemiAnalysis further flagged that the larger NVL576 system, which would link eight racks over optical connections, is likely delayed or limited to small volumes (reported by CNBC and The Next Web).

Bloomberg and Yahoo Finance report that Asian PCB and supplier equities slid after the SemiAnalysis disclosure; Yahoo Finance cited specific moves including double-digit drops for several board suppliers (reported by Yahoo Finance citing Bloomberg). Tech press coverage (Techzine, The Next Web) notes that Nvidia's current Rubin systems remain in production and that Rubin shipments to major cloud partners are ongoing this year (reported by The Next Web and Techzine).

The public accounts single out the midplane as the failure point. Techzine reports the midplane is an extremely high-complexity board (they describe it as having 78 layers) and that such multilayer, high-density boards can suffer low yields and defect sensitivity during scale production. In analogous datacenter hardware efforts, high-layer-count boards and rigid orthogonal backplanes have historically elevated manufacturer qualification time and yield risk, increasing lead times even when the upstream silicon is ready.

Teams planning capacity for 2027-2028 should treat a single-board manufacturability failure as a systemic risk to rack-delivered density. Industry experience shows that when a single mechanical/electrical subsystem constrains volume, organizations often (a) accept lower-density interim designs, (b) shift workloads to existing-generation hardware, or (c) increase investment in cooling and floor-space to compensate. The reporting that the NVL72x2 stopgap faced customer pushback underscores a second practical constraint: cloud operators prioritize maintainability and operational simplicity, not only peak density.

Multiple outlets frame the delay as creating a temporary competitive window for rivals such as AMD and Google (reported by CNBC and The Next Web). For practitioners sourcing large-scale training capacity, this could broaden vendor choices for high-end jobs in 2027, but the long-term balance depends on when Kyber achieves manufacturable yields.

Observers will look for supplier commentary or manufacturing-volume guidance from PCB partners (Bloomberg/Yahoo tracked supplier stock moves), public statements from Nvidia on qualification milestones, and any incremental hardware revisions that reduce board-layer count or reorganize connectivity. Also monitor reported availability and volumes for NVL576 and alternative multi-rack configurations cited in coverage.

Reported sources: Coverage and claims above derive from SemiAnalysis' post (reported on X), CNBC, Tom's Hardware, The Next Web, Techzine, Bloomberg and Yahoo Finance, as cited in the reporting summarized here.