Nvidia’s next flagship AI system has hit a manufacturing wall. The Kyber rack, a server cabinet designed to house the company’s 2027 Rubin Ultra chips, has been delayed by more than 12 months to 2028, according to the research firm SemiAnalysis.
Kyber is not a chip but a cabinet that packs 144 of Nvidia’s most powerful GPUs into a single unit so they function as one giant computer, the kind of density AI labs need to train and run frontier models. It mounts chips in vertical trays to boost density and cut latency. Nvidia did not respond to a request for comment, so the details, first posted by SemiAnalysis on July 6, should be treated as analyst reporting rather than official guidance.
The bottleneck is an unglamorous component: a specialized multi-layer circuit board called the PCB midplane, which Nvidia also calls the orthogonal backplane, that connects the trays without cables. At 78 layers, it is among the most complex boards ever attempted for a commercial computer, and SemiAnalysis said it remains too difficult to build reliably at scale, citing problems with signal integrity, power delivery and thermal design.
The trouble does not stop there. A larger system, the NVL576, which links eight racks with optical connections, is also likely delayed or limited to low volumes. And a stopgap called NVL72x2, which bolted two current-generation racks together, was scrapped after cloud providers rejected it as awkward and costly to run. That, SemiAnalysis said, leaves Nvidia with no proven way to expand the scale of its most powerful Rubin Ultra clusters in 2027.
The problem is compounded by a separate setback. SemiAnalysis reported that Nvidia recently downgraded the Rubin Ultra chip itself from a quad-chip to a dual-chip design over manufacturing concerns, roughly halving its per-package compute, while fully production-ready optical switching is not expected until the following Feynman generation. Together, the two issues suggest Nvidia’s blistering pace of shipping a new architecture every year is colliding with the physical limits of manufacturing. Missing a rung on that annual ladder is rare for the company.
The timing hands Nvidia’s rivals a rare opening at the very high end. AMD and Google, whose in-house chips are already winning business from top AI labs, could exploit a period in which the largest cluster buildable with Rubin Ultra in 2027 may be no bigger than one built with the prior Rubin generation.
For the hyperscalers pouring hundreds of billions of dollars into AI infrastructure, the practical worry is capacity timing: racks tied to Rubin Ultra roadmaps may arrive later or in smaller volumes than procurement plans assumed. The report also fed jitters in the AI hardware trade, with Asian technology and circuit-board stocks sliding on the news.
Keeping It in Perspective
The delay does not dent Nvidia’s near-term dominance. Its current-generation Rubin systems are in full production and begin shipping this fall to eight cloud partners, including Amazon Web Services, Microsoft Azure and Google Cloud, and none of that is affected.
SemiAnalysis even projects Nvidia’s data-center compute revenue will run 20% above Wall Street consensus in the second half of its fiscal 2027, and Nvidia shares were roughly flat on the report. The caveats matter too: the reporting does not identify which supplier is struggling with the board, confirm the Rubin Ultra silicon’s broader schedule, or say what Nvidia will offer customers to fill the 2027 gap. For now it is a story about execution risk at the bleeding edge, not a crack in Nvidia’s core business.
