NVIDIA Kyber NVL144 Delayed by Over 12 Months to 2028

NVIDIA Kyber NVL144 delayed by over 12 months to 2028

Three months after Jensen Huang presented Kyber NVL144 at GTC, the project was pushed back more than twelve months.

The delay moves full availability to 2028. At the same time, plans for an NVL72x2 back-to-back rack design were canceled, tightening the scaling options for Rubin Ultra systems.

Announcement meets immediate revision

NVIDIA showed the Kyber NVL144 configuration during the GTC keynote as part of its next-generation roadmap. The system was positioned as a larger-scale successor to current NVL72 racks.

NVIDIA has not confirmed or denied the reported delay through earnings calls or official statements. SemiAnalysis first surfaced the revised 2028 target and the removal of the NVL72x2 rack pairing from active development. Reuters also noted supply-chain sources pointing to extended qualification timelines for the larger chassis. The shift directly affects customers who had begun capacity planning around the original Kyber schedule.

Capacity planning now under pressure

AI cloud providers and large enterprises book hardware two to three years ahead. A twelve-month slip compresses the window for Rubin Ultra deployments that rely on the wider Kyber chassis. Industry analyst Dylan Patel noted that “hyperscalers who penciled in Kyber racks for late-2027 clusters will now face either higher NVL72 counts or deferred capacity additions.” Bloomberg reported that several cloud operators have already begun reviewing alternative rack options to avoid gaps in 2027 AI training capacity.

Cancellation of the NVL72x2 back-to-back rack removes one path to higher node counts without redesigning power and cooling layouts. Operators must either stay with existing NVL72 clusters or wait for a different future architecture. In practice, this means teams that had reserved floor space and power feeds sized for Kyber now risk under-utilized liquid-cooling loops and stranded power capacity until a revised design ships.

Why the schedule moved

The injected update points to two technical constraints. First, the larger Kyber chassis requires new power-distribution and liquid-cooling components that have not yet reached volume production. Second, the cancellation of the NVL72x2 pairing reflects unresolved signal-integrity and mechanical issues when two full racks are placed back-to-back.

No official NVIDIA statement has confirmed or denied the specific dates. The information originates from supply-chain tracking cited in the SemiAnalysis thread.

What remains unverified

The twelve-month figure and the 2028 target have not been corroborated by NVIDIA earnings transcripts or partner briefings. Memory supply, board-level redesigns, and datacenter qualification timelines can all shift independently, so the exact length of the delay could change. The Verge has not yet reported independent confirmation of the revised timeline.

Customers tracking the program now face two open questions: whether partial Kyber shipments will appear before 2028 and what alternative rack topology NVIDIA will offer instead of NVL72x2.

Immediate market signals to monitor

Three concrete checkpoints will clarify the situation over the next six months. First, any mention of Kyber or Rubin Ultra rack configurations in the next NVIDIA earnings call will test the 2028 narrative. Second, published revisions to 2027 data-center build plans from major cloud providers will show whether they have already adjusted orders. Third, component lead-time reports for high-density liquid-cooling manifolds will indicate whether the physical constraints cited in the update are easing or persisting.

Until one of these signals appears, the Kyber NVL144 delay remains a reported claim rather than a confirmed program milestone.