Nvidia's next flagship AI server system just hit a wall, and the culprit is a single circuit board. The company's Kyber NVL144 rack architecture, designed to house 144 of its most powerful Rubin Ultra chips in a single cabinet, has been delayed by more than 12 months, pushing its launch from 2027 to 2028, according to research firm SemiAnalysis, as reported by CNBC on July 6.
The setback traces to a specialized 78-layer printed circuit board midplane that connects modules inside the rack. SemiAnalysis said the component "remains challenging from a manufacturability standpoint," meaning Nvidia can't yet produce it at the scale hyperscale customers require. Nvidia did not respond to CNBC's request for comment. Shares fell 1.4% to $194.83 on Monday, trimming the company's market capitalization to roughly $4.7 trillion.
A String of Design Setbacks
The Kyber delay isn't an isolated hiccup. Nvidia also canceled its NVL72x2 back-to-back rack design after cloud providers pushed back on its operational complexity, and the larger NVL576 configuration may also slip or ship in limited volumes. The company's Rubin Ultra chip has additionally been scaled back from a planned quad-chip design to a dual-chip variant. A key interconnect technology, CPO-NVSwitch, won't arrive until the generation after Rubin, called Feynman, leaving Nvidia without what SemiAnalysis called "a proven solution to expand the scale-up world size for Rubin Ultra."
Nvidia's current-generation systems remain unaffected. Its existing Oberon and Rubin racks are in full production and begin shipping this fall to major cloud partners including Amazon Web Services, Microsoft Azure, and Google Cloud.
Why This Matters for Business
This is exactly the kind of story that gets buried in technical detail but has real business consequences. I've advised companies for years on AI infrastructure decisions, and the lesson here is straightforward. Even the most dominant player in AI hardware runs into physical manufacturing limits. Companies planning multi-year AI compute strategies around a single vendor's roadmap need contingency built in, because roadmaps slip.
The delay also creates a genuine opening for AMD and Google, whose in-house TPU silicon is already winning business from top AI labs. SemiAnalysis projects Nvidia's data center compute revenue will still beat Wall Street estimates by 20% in the second half of fiscal 2027, suggesting current-generation demand remains strong even as the next generation slips. One portfolio manager described the stock weakness as mostly profit-taking rather than a signal that AI spending itself is slowing.
What to Watch Next
For businesses relying on cloud infrastructure or planning AI capacity for 2027 and 2028, this is a signal to build flexibility into procurement plans rather than betting everything on a single hardware generation arriving on schedule. Watch whether AMD or Google convert this timing gap into actual enterprise wins over the next two quarters. That will tell you whether Nvidia's moat is genuinely wobbling or whether this is a temporary stumble in an otherwise dominant position.
