Nvidia's next-gen AI rack system delayed to 2028 on manufacturing snags
CNBC
Nvidia's next-gen Kyber server rack is slipping a year, to 2028, over circuit board manufacturing problems. Nvidia says its roadmap's fine, but rivals AMD and Google just got a longer runway at the high end.
Based on reporting by CNBC — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Nvidia's annual cadence of bigger, denser, more powerful AI hardware has been the backbone of its dominance. So when research firm SemiAnalysis says the company's next flagship rack system, Kyber, is slipping more than a year — from 2027 to 2028 — it's not a small thing, even if Nvidia insists otherwise.
Kyber was supposed to be the physical home for Rubin Ultra, Nvidia's 2027 chip generation, cramming 144 top-tier GPUs into a single cabinet with vertically mounted compute trays instead of horizontal ones. The vertical layout was meant to squeeze more chips closer together and cut the latency between them, which matters enormously when you're trying to make 144 processors behave like one giant brain. According to SemiAnalysis, the holdup isn't exotic chip physics or software — it's a printed circuit board. Specifically, the multi-layer PCB midplane that stitches all those modules together has turned out to be brutally hard to manufacture at scale.
Making things worse, Nvidia reportedly had a fallback plan: just bolt two current-generation racks together to approximate Kyber's power. Cloud providers hated it. SemiAnalysis says hyperscalers pushed back hard on the design's clunkiness and the operational headache of running it, and the idea got scrapped entirely. That leaves Nvidia, per the report, without a clear path to scale up Rubin Ultra's rack-level performance — a gap that could give AMD and Google's homegrown TPUs a genuine opening at the very top of the AI hardware market, right as both are already chipping away at Nvidia's grip on major AI labs.
Nvidia's response was blunt: the roadmap is intact, full stop. And the stock barely blinked, ticking up about 1% Monday. Analysts like Paul Triolo argue the bigger constraint on U.S. AI buildouts isn't chip availability anyway, it's electricity — data centers are increasingly bottlenecked by power, not silicon. If that's true, a Kyber delay might matter less than it sounds, since the grid wasn't going to be ready for it in 2027 regardless. Meanwhile Nvidia's current Rubin systems are shipping this fall to AWS, Microsoft Azure and Google Cloud, and SemiAnalysis actually expects Nvidia's data-center revenue to beat Wall Street's second-half fiscal 2027 estimates by 20%.
There's a wider story here too, about China. Triolo's take is that Huawei and other domestic chipmakers aren't really racing to catch Nvidia anymore — they're building a parallel stack on their own terms. The real question by 2030 won't be who copied whom, but how good China's alternative actually gets.
My take — AI-written commentary, not fact-checked reporting
A one-year slip on a PCB isn't the end of Nvidia's empire, but it's a tidy reminder that hardware has physical limits software roadmaps love to ignore. I'd watch AMD and Google's TPU wins less as a threat to Nvidia's dominance and more as proof that 'good enough at the top end' is becoming a real strategy, not just a consolation prize. And the China framing is the smartest line in this whole story — stop measuring them against Nvidia's yardstick.
Read more about this at: CNBC