Nvidia delays Rubin Ultra-based Kyber servers from 2027 to 2028 over manufacturing problems

Nvidia has delayed its Kyber server architecture, built on Rubin Ultra accelerators, from 2027 to 2028, CNBC reported, citing analyst firm SemiAnalysis. The delay is attributed to difficulties manufacturing one of the system's key components.

What we know about the delay

• Source: CNBC, citing SemiAnalysis analysts
• The Kyber architecture is moving from 2027 to 2028
• Kyber is built on Rubin Ultra accelerators
• Cause of the delay: difficulties manufacturing one of the system's key components

Rubin Ultra is one of Nvidia's upcoming generations of AI accelerators. On the company's annual roadmap, it follows the Blackwell, Blackwell Ultra, and Vera Rubin architectures. Kyber is the name of the rack-scale server platform that ties these chips together into a single computing system for data centers that train and serve large AI models.

Since 2024, Nvidia has publicly committed to shipping a new generation of AI infrastructure roughly once a year. That cadence was once considered nearly impossible for hardware as complex as liquid-cooled rack systems with tens of thousands of chip-to-chip connections. This unprecedented pace helped the company dominate the market for chips used to train large language models, but it also depends on the entire supply chain running without interruption, from the contract fabs that make the dies to the companies that assemble the server racks.

Why the delay matters for the market

In recent years, Nvidia has stuck to a tight annual refresh cycle for its top AI chips and racks, and that pace has helped it keep its grip on the market for accelerators used to train and run inference on large language models. Any shift in that schedule directly affects hyperscalers such as Microsoft, Google, Amazon, and Meta, which plan data center hardware purchases years in advance and size their capacity around specific launch dates for new chip generations.

Manufacturing problems with key components for cutting-edge AI chips have been common in recent years. Bottlenecks have appeared in advanced chip packaging, HBM memory supply, and contract manufacturing capacity, and they hit not only Nvidia but the entire semiconductor supply chain serving the AI industry.

How this fits Nvidia's strategy

Nvidia publicly builds its market position around the promise of a predictable annual architecture refresh. That approach helps keep customers from switching to rivals such as AMD or to hyperscalers' in-house chips, because buyers know in advance when the next generation will arrive and can plan capital spending around it. Pushing back even a single server platform by a year raises questions about whether that pace is sustainable and forces major customers to revisit their own timelines for bringing new data centers online.

For SemiAnalysis, which first reported the delay, slips like this also serve as an indicator of the overall health of the advanced semiconductor supply chain, above all advanced packaging capacity and next-generation memory production, where demand has consistently outstripped supply in recent years.

What this means

A one-year delay to a flagship server platform signals that even the leader in AI chips faces physical limits on how fast it can scale production. That could slow the pace at which the largest cloud providers build out compute capacity for new generations of AI models.