Nvidia's next flagship AI rack, once slated for 2027, won't ship until 2028, and the cause is a circuit board most investors have never heard of.
Nvidia's Kyber NVL144 rack architecture has slipped by more than 12 months, landing in 2028 instead of alongside the Vera Rubin Ultra chips it was built to house next year. According to a report from SemiAnalysis, relayed by CNBC on July 6, the culprit isn't a chip at all. It's the PCB midplane, a 78-layer circuit board that has to connect dozens of compute trays inside a single rack without failing.
That sounds like a small thing. It isn't. Kyber's whole pitch was mounting GPU compute trays vertically instead of horizontally, packing them tighter and cutting the distance signals have to travel between chips. Nvidia unveiled the design at GTC 2026 as the physical shell for Vera Rubin Ultra, the chip line the company has been selling hyperscalers on for next year. Building a board with 78 layers that can carry that density of high speed signals without shorting or warping under heat turns out to be a harder manufacturing problem than Nvidia's roadmap assumed.
Kyber isn't the only casualty. SemiAnalysis also reported that the NVL576, which links eight Oberon racks together using co-packaged optics between NVSwitches, is likely to be delayed or held to low volume because of the same class of manufacturing problems in CPO. Nvidia has reportedly canceled the NVL72x2 back to back rack design outright.
Nvidia has not confirmed any of it. The company declined to comment when CNBC asked, which is its own kind of answer. When a company doesn't deny a report like this, procurement teams at hyperscalers tend to take it at face value and start adjusting their own timelines quietly.
The market didn't wait for confirmation either. Bloomberg reported that Asian tech stocks and U.S. optical component makers sold off on the news, with AAOI dropping 17% and Lumentum down 8% as investors priced in slower near term demand for the optical gear Kyber and NVL576 would have required.
This is the kind of delay rivals wait years for. Nvidia has spent the past three years setting an annual cadence that nobody else in the industry has matched, and a slip like this is the clearest evidence yet that the pace is running into physical limits rather than just competitive ones. AMD and Google are the two most obvious names to benefit, according to CNBC's framing of the report. Google already ships its own TPUs at scale for internal workloads, and AMD's Instinct line is the nearest thing to a drop in alternative for hyperscalers who don't want to wait a year for a rack that isn't ready.
Custom silicon programs stand to gain the most breathing room. Amazon has been building its own Trainium chips for years specifically to reduce its Nvidia dependence, and Anthropic's compute arrangements tied to Google and Samsung's foundry capacity point at the same instinct across the industry: don't build your entire roadmap on one supplier's ability to hit its own dates. A 12 month slip on a flagship rack design is exactly the kind of event that validates that hedge.
None of this means Nvidia is in trouble. The company still controls the overwhelming majority of AI training compute, and a rack delay is not the same as a chip delay. Vera Rubin Ultra itself hasn't been reported as pushed back, only the specific rack it was designed to sit in. But for hyperscalers who built 2027 capacity plans around Kyber's density and power efficiency, this report means going back to their own vendors and asking hard questions about what actually fits in a rack that ships on time instead of one that doesn't.
Frankly, the more interesting number here isn't the delay length. It's that Nvidia would rather eat a public manufacturing setback than rush a board that could fail inside a customer's data center. That's a defensible engineering choice. It's a much harder one to defend to investors who priced in an unbroken run of annual leaps.
Also read: SK Hynix files to raise $28 billion in a Nasdaq listing built entirely on AI memory demand • Samsung's record AI chip bonus deal is tearing its own union apart • A federal judge just forced the Pentagon to give Alibaba back its lobbyists
