NVIDIA is proving once again that even unlimited budgets must eventually bow to the laws of physics and the whims of manufacturing. The rapid hardware refresh cycle that Jensen Huang used to hook the market has hit a major bottleneck: the flagship Kyber NVL144 server system, unveiled with great fanfare at GTC just three months ago, has been officially delayed until 2028. According to analysts at SemiAnalysis, the root of the problem lies in manufacturing defects within the central printed circuit board (PCB midplane). For those who have already allocated billions in CapEx to expand data centers around this new capacity, the news means one thing: the rules of the game are changing on the fly.
Canceled Intermediate Solutions and Density Shortages
The revision of NVIDIA’s roadmap is more than just a delay; it is a total clear-out. As reported by SemiAnalysis, the company has completely scrapped the NVL72x2 architecture after cloud providers expressed skepticism regarding its awkward form factor and exorbitant operating costs. But the real blow was dealt to the Rubin Ultra chip: the version featuring four compute dies has been axed. We are left with only a dual-chip variant that delivers barely half of the promised performance.
The critical CPO-NVSwitch interconnect technology, capable of bridging dozens of chips into a single supercomputer, will not debut until the Feynman generation.
The absence of CPO-NVSwitch until the next cycle means NVIDIA currently lacks a viable method to scale Rubin Ultra into truly massive systems. To plug this gap, the company intends to aggressively push Oberon-Rubin racks, effectively extending the lifecycle of current form factors. In our view, this looks like an attempt to save face while engineers scramble to redesign the next-generation architecture.
Supplier Sell-off and a Window for Competitors
The market reacted instantly, hammering NVIDIA’s Asian supply chain. Shares of Japan’s Ibiden, for whom NVIDIA is a primary client, tumbled 10%, while Hong Kong-listed Kingboard Laminates dropped 18%. Samsung Electro-Mechanics and Taiwan’s Elite Material also recorded double-digit losses. This collapse serves as a stark indicator of the global supply chain’s fragility: any misstep in Santa Clara echoes across the entire industry. While this sell-off is partly driven by profit-taking, it highlights the inherent risks of vendor lock-in.
As NVIDIA pauses to regroup, AMD with its MI500X and Google with the TPUv8i Broadfly chips have a genuine opportunity to close the gap. For infrastructure leaders, this is a moment of truth: whether to remain loyal to the delayed Kyber architecture or begin diversifying portfolios toward alternative platforms. With ROI timelines directly tied to the availability of compute density, waiting until 2028 may be a luxury many businesses simply cannot afford.