Nvidia disputes report that its Kyber AI rack system faces a 12-month delay to 2028

Research firm SemiAnalysis reported on July 6, 2026 that Nvidia's next-generation Kyber rack-scale architecture, designed to house the 2027 Vera Rubin Ultra chips, has slipped by more than a year to 2028 because of manufacturing problems with a key circuit board. Nvidia publicly rejected the report, saying its roadmap is intact, but did not address the specific technical claims.

What's new

SemiAnalysis said the Kyber NVL144 rack — which mounts GPUs vertically in compute trays rather than horizontally to improve density and reduce latency — has been delayed because its PCB midplane, a specialized 78-layer circuit board connecting the system's modules, is proving extremely difficult to manufacture at scale. According to the research firm, the board's trace widths and spacings are 25 micrometers or less, needed to support SerDes signal integrity at rates exceeding 448G, making it one of the most complex PCBs ever designed for a commercial computing product.

The report also said NVL576, a larger system that links eight racks together via optical connections, is likely to be delayed or limited to small production volumes as a result of the same manufacturing constraints. A stopgap plan to bolt together two of Nvidia's current-generation racks to approximate the missing capacity has reportedly been scrapped after cloud service providers pushed back on its design and the operational burden it would create.

Asked about the report, Nvidia said simply: "Our roadmap is intact." The company did not respond to further questions about whether that statement referred to its original Kyber timeline or an already-revised one, leaving the specific dispute over the 2028 date unresolved.

Context

Kyber is the rack-scale platform Nvidia has been previewing since GTC 2025 as the successor to its current Oberon design, intended to pair with the Vera Rubin Ultra chips slated for 2027. Nvidia has leaned heavily on an near-annual cadence of new chip and rack platforms — Blackwell, Vera Rubin, and now Kyber — to stay ahead of AMD, Google, and in-house silicon efforts from its largest customers. SemiAnalysis has also reported this year on manufacturing snags affecting Nvidia's Blackwell-generation shipments, so this fits a pattern the firm has been tracking of execution risk in Nvidia's rack-scale roadmap even as chip-level shipments continue.

Why it matters

A one-year slip in rack-scale infrastructure is a bigger deal than it sounds: hyperscalers plan data center buildouts years in advance around Nvidia's stated cadence, and a delay to the interconnect and power infrastructure around Rubin Ultra could ripple through capacity planning at every major cloud provider betting on that chip generation. SemiAnalysis frames the missing scale-up capacity as leaving Nvidia with "no proven solution to expand the scale-up world size for Rubin Ultra," which could open a rare technical window for AMD or Google's in-house TPUs at the high end of the market. Nvidia's terse, non-specific denial — asserting the roadmap is intact without confirming or refuting the PCB manufacturing details — leaves customers with limited information to plan around, and is likely to keep the dispute in the spotlight until Nvidia either publishes a firm Kyber ship date or SemiAnalysis's account goes unchallenged on the specifics.