Nvidia's Kyber NVL144 AI rack slips to 2028 on PCB snag

Nvidia's most talked-about next-generation rack, the Kyber NVL144, has reportedly slipped by more than a year, from a 2027 launch alongside the Rubin Ultra chip family to a 2028 target. The claim comes from research firm SemiAnalysis and was picked up by CNBC, which notes that Nvidia demoed the system on stage at GTC just three months ago.

The reason SemiAnalysis gives is refreshingly specific for this kind of delay story. The bottleneck is the rack's PCB midplane, the multi-layer printed board that carries every signal between the modules stacked inside a Kyber cabinet. That board, per the report, "remains challenging from a manufacturability standpoint." Kyber was Nvidia's answer to tying together 144 of its biggest GPUs as one training-scale domain, so if the midplane cannot be built at yield, there is not much to ship.

The knock-on effects are the interesting part. Nvidia had a fallback design, the NVL72x2 back-to-back rack, that would have paired two current-generation racks to buy time. SemiAnalysis, in a post on X, says that fallback was "cancelled due to heavy pushback from CSPs and hyperscalers over its odd design and heavy operational burden," which is unusually blunt language about hyperscaler behaviour. The larger NVL576 configuration, which uses optical links between eight racks, is also likely delayed or limited to small volumes. That leaves Nvidia with "no proven solution to expand the scale-up world size for Rubin Ultra," which is the polite way of saying the biggest cluster you can build with Rubin Ultra in 2027 might be the same size you can build with Rubin.

The honest caveat is that this is an analyst report, not an Nvidia disclosure. CNBC says Nvidia did not respond to its request for comment, so the 12-month figure, the NVL72x2 cancellation, and the scope of the NVL576 slip should be taken as reported, not settled. The reporting also does not say which supplier is struggling with the midplane, whether the Rubin Ultra silicon itself is on track, or what Nvidia will actually sell customers in the 2027 gap.

The forward-looking read, and the one Nvidia's competitors will find most interesting, is that SemiAnalysis explicitly frames this as "a rare technical opening at the high end of the market" for AMD and Google. Rack-scale interconnect has been the moat, not the die. If that moat is thinner for a year, cloud buyers who kept an AMD or TPU option warm suddenly have leverage they did not have last quarter.