Nvidia Delays Kyber NVL144 Rack System to 2028
CNBC reported that Nvidia's Kyber NVL144 rack system has been pushed to 2028 after SemiAnalysis said the specialized PCB midplane remains difficult to manufacture. For infrastructure teams, the operational issue is capacity timing: racks tied to Rubin Ultra roadmaps may arrive later or in smaller volumes than procurement plans assumed. CNBC also reported SemiAnalysis' view that NVL576 could be delayed or limited to small volumes, while SemiAnalysis' X post said the NVL72x2 back-to-back rack architecture was canceled. Nvidia did not respond to CNBC's request for comment, so teams should treat the details as analyst-reported rather than official vendor guidance.
For practitioners, the important point is not just one delayed rack name. Rack-scale AI systems now couple GPU roadmaps, midplane manufacturing, optical links, server OEM capacity, and cloud procurement timelines, so a reported delay can move real project plans even before Nvidia confirms or denies the details.
What happened
CNBC reported on July 6, 2026 that research firm SemiAnalysis says Nvidia's Kyber NVL144 rack architecture has slipped to 2028 because the specialized PCB midplane remains difficult to manufacture. CNBC also reported SemiAnalysis' view that NVL576, the larger configuration intended to connect eight racks through optical links, is likely delayed or limited to small volumes. SemiAnalysis' X post, summarized by secondary outlets, also said the NVL72x2 back-to-back rack architecture was canceled. CNBC said Nvidia did not respond to a request for comment.
Technical context
The midplane matters because high-density rack systems depend on signal integrity, power delivery, thermal design, and manufacturable board layers all working at once. A delay at that layer can affect the system schedule even when the underlying GPU roadmap still looks intact. Separate Tom's Hardware coverage of SemiAnalysis' earlier Rubin Ultra reporting described similar manufacturability concerns around ambitious multi-chip designs, which makes this a broader execution-risk story rather than a simple date slip.
Industry context
CNBC framed the reported delay as a potential opening for competitors such as AMD and Google at the high end of AI infrastructure. That should be read as market analysis, not as evidence of Nvidia's internal plans. The safer conclusion is that buyers should expect more uncertainty around early high-density rack availability and should watch for OEM-level confirmation before locking schedules to analyst-reported timelines.
For practitioners
Teams planning large training or inference capacity should keep contingency buffers, validate delivery assumptions with vendors, and separate public roadmap announcements from purchase-order-ready availability. The story is especially relevant for organizations whose 2027 or 2028 plans assume a specific rack-scale topology, interconnect design, or cloud instance family.
What to watch
Watch for Nvidia roadmap updates, server OEM notices, cloud-provider capacity announcements, and follow-up SemiAnalysis reporting that confirms whether the midplane issue changes volumes, timing, or architecture. Public availability signals will matter more than social-post summaries for procurement decisions.
Key Points
- 1SemiAnalysis, via CNBC, says Kyber NVL144 slipped to 2028 because the PCB midplane remains difficult to manufacture.
- 2The report matters for buyers because rack-scale delays can shift capacity, cloud lead times, and training-plan assumptions.
- 3Nvidia has not publicly confirmed the delay, so procurement teams should track OEM notices and updated roadmap statements.
Scoring Rationale
This is a notable AI-infrastructure story because a reported Kyber NVL144 delay can affect capacity planning, cloud procurement assumptions, and high-end training timelines. The impact remains below major industry-shaking territory because the claims are analyst-reported, Nvidia has not publicly confirmed them, and the story does not change near-term model or software research directions.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
