Nvidia postpones Kyber servers on Rubin Ultra from 2027 to 2028 due to manufacturing issues

Nvidia has postponed the release of its Kyber server architecture, built on Rubin Ultra accelerators, from 2027 to 2028 — according to CNBC, citing analytics company SemiAnalysis. The delay is attributed to manufacturing difficulties with one of the system's key components.

What is Known About the Postponement

• Information source — CNBC citing SemiAnalysis analysts
• Kyber architecture postponed from 2027 to 2028
• Kyber is built on Rubin Ultra accelerators
• Reason for delay — manufacturing difficulties with one of the system's key components

Rubin Ultra is one of the next generations of Nvidia AI accelerators, part of the company's annual roadmap following Blackwell, Blackwell Ultra, and Vera Rubin architectures. Kyber, in turn, is the name of a server rack-platform that consolidates these chips into a single computing system for data centers training and serving large AI models.

Since 2024, Nvidia has publicly adhered to a strategy of releasing a new generation of AI infrastructure approximately once per year — a pace previously considered nearly impossible for such complex equipment as rack systems with liquid cooling and tens of thousands of chip interconnections. This unprecedented update rhythm has allowed the company to maintain its dominant position in the market for chips used to train large language models, but it also requires seamless operation of the entire supply chain — from contract chip manufacturers to server rack assemblers.

Why This Postponement Matters for the Market

Nvidia has maintained a tight annual update cycle for its flagship AI chips and racks in recent years — this pace has allowed the company to maintain dominance in the market for accelerators for training and inference of large language models. Any shift in this schedule directly impacts the procurement plans of hyperscalers like Microsoft, Google, Amazon, and Meta, which plan data center equipment purchases years in advance and calculate capacity based on specific dates for new chip generations.

Manufacturing difficulties with key components for advanced AI chips have not been rare in recent years: bottlenecks arise in advanced chip packaging, HBM memory supply, and contract manufacturer capacity — affecting not only Nvidia but the entire semiconductor supply chain for the AI industry.

How This Fits Into Nvidia's Strategy

Nvidia publicly positions its market dominance around the promise of predictable annual architecture updates — this tactic allows the company to prevent customers from switching to competitors like AMD or hyperscalers' own developments, as clients know in advance when the next generation of chips will arrive and can plan capital expenditures accordingly. Postponing even one server platform by a year raises questions about the sustainability of this pace and forces major customers to reconsider their own schedules for bringing new data centers online.

For SemiAnalysis analysts, who first reported the delay, such postponements also serve as an indicator of the overall health of the advanced semiconductor supply chain — particularly the capacity for advanced chip packaging and next-generation memory production, demand for which has consistently outpaced supply in recent years.

What This Means

A one-year postponement of a flagship server platform signals that even the market leader in AI chips faces physical limits to the speed of manufacturing scale-up, and this may slow the pace at which the world's largest cloud providers expand computing capacity for new generations of AI models.