Nvidia Kyber rack systems delayed to 2028 amid manufacturing challenges
Nvidia has delayed its Kyber rack-scale architecture for Rubin Ultra chips from 2027 to 2028 due to manufacturing challenges involving a key printed circuit board. The delay impacts Nvidia's ability to provide high-density compute solutions, potentially creating a market opening for competitors like AMD and Google.
Key Takeaways
- The Kyber NVL144 server cabinet packs 144 Rubin Ultra GPUs into a single unit using a high-density vertical mounting design.
- A 78-layer printed circuit board (PCB) midplane is the primary bottleneck, with current yields failing to meet production requirements.
- Nvidia scrapped a backup plan to link two current-generation racks after cloud providers rejected the proposal due to high operational costs and design complexity.
- The larger NVL576 system, which utilizes optical connections to link eight racks, is also facing likelihood of significant delays or volume restrictions.
Why It Matters
The delay marks a rare stumble in Nvidia's aggressive annual release cadence and creates a technical opening for rivals like AMD and Google. For streaming and AI-heavy platforms, this hardware bottleneck limits the density of next-generation training clusters, potentially slowing the transition to trillion-parameter models. As hyperscalers seek alternatives to Nvidia's premium-priced systems, the cancellation of Nvidia's transitional 'back-to-back' rack design reduces the immediate pressure on competitors to match Nvidia's architectural scale. Watch for whether cloud service providers (CSPs) like AWS and Azure increase allocations for AMD’s Helios platform or Google’s custom TPUs in 2027 to fill the capacity gap left by the Kyber slip.
Additional Context
The delay of the Kyber rack system follows a period of intense pressure on Nvidia's supply chain as it transitioned from the Hopper architecture to Blackwell. In early 2026, many hyperscalers faced backlogs for Blackwell B200 chips, which featured a performance improvement of up to 30x over the H100 but required complex liquid-cooling infrastructure that strained Taiwanese component suppliers, per WinZheng and Wccftech. This manufacturing complexity was exacerbated by geopolitical shifts; by late 2025, U.S. export restrictions on Blackwell-class silicon to China forced Nvidia to develop lower-specification variants to maintain international revenue streams, according to reporting from CapacityGlobal in November 2025. While Nvidia struggles with the 78-layer PCB midplane required for Kyber, competitors have positioned themselves to capture the resulting domestic demand. AMD, under CEO Lisa Su, reported record data center revenue of $5.8 billion in Q1 2026, a 57% year-over-year increase driven by the adoption of its MI300 and MI450 series GPUs, per Intellectia. Concurrent with Nvidia's Kyber struggle, AMD previewed its Helios AI platform at CES 2026, utilizing UALink-over-Ethernet interconnects as a direct challenge to Nvidia’s proprietary NVLink fabric. Additionally, Alphabet’s Google has scaled its in-house TPU production, booking capacity for over 3 million units in 2028 to reduce its infrastructure dependence on third-party silicon providers, according to industry analysts in July 2026.
Read full article at cnbc.com
Get this in your inbox → Subscribe
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source