Nvidia's Rack Roadmap Hits A Delay

SemiAnalysis reported that Nvidia's Kyber NVL144 rack architecture has been delayed to 2028, citing PCB midplane manufacturability challenges. The thread also said CPO-NVSwitch would not arrive until the Feynman generation. A Jefferies-linked market note separately said the absence of Nvidia Kyber/backplane PCB in 2027 was becoming highly likely, implying Rubin Ultra would stick with Oberon/NVL72-style structure. (Secondary: https://x.com/wallstengine/status/2074059082408034356)

The AI buildout is not just a money problem. It is a physical-infrastructure problem. If Nvidia's next-scale rack architecture is delayed, the market has to reprice how quickly frontier compute can keep scaling. That complicates the "spend more, get more intelligence" story. It also gives AMD and Google TPU architectures a wider competitive opening. For the thesis, this is a compute-layer bottleneck that reinforces the monetization gap: spending keeps rising while the technical path gets harder.