Nvidia delays Kyber rack for Rubin Ultra chips to 2028

What happened?

Nvidia has delayed delivery of its Kyber rack, the server cabinet that was to house its Rubin Ultra chips starting in 2027, until 2028. The cause is a problem with a single circuit board, according to SemiAnalysis and as reported by CNBC. The Kyber rack is not a chip but a complete system integrating 144 Rubin Ultra GPUs and liquid cooling components, with a thermal design power (TDP) of up to 1.5 kW per chip. The delay affects the entire Rubin platform, which is now expected in 2028. This setback adds to a series of technical challenges Nvidia has faced in scaling its AI systems. The problematic board, referred to as a 'baseboard management controller' (BMC) according to industry sources, manages power and data distribution among the GPUs. Although Nvidia has not officially confirmed the delay, SemiAnalysis's information has been corroborated by multiple sources close to the company. This incident recalls the delay of the DGX GH200 rack in 2023, also due to cooling issues, but the scale of the Kyber is much larger: it integrates 144 GPUs in a single cabinet, with direct liquid cooling (DLC) and a next-generation NVLink interconnect architecture promising 1.8 TB/s bandwidth per GPU.

Why is this important?

Nvidia dominates the AI accelerator market with over 80% share, according to Mercury Research data from 2025. Any delay in its products has a ripple effect across the entire data center supply chain, from hyperscalers like Microsoft, Amazon, and Google to AI startups that depend on compute capacity. The Kyber rack delay means the next generation of massive AI model training infrastructure will not be available until at least a year later than planned. This could slow the development of larger and more complex models and give an opportunity to competitors like AMD and Intel, as well as custom solutions like Google's TPUs. For example, OpenAI's GPT-5, expected to require clusters of over 100,000 GPUs, could see training delays if Nvidia misses deadlines. Additionally, the delay affects Nvidia's roadmap: the company had planned to launch the Rubin architecture in 2026, followed by Rubin Ultra in 2027, but now everything shifts to 2028. This contrasts with AMD's strategy, which has launched its Instinct MI400 line with competitive performance and is gaining traction in the HPC market.

Consequences for businesses and users

For companies planning to upgrade their AI clusters in 2027, the delay means they will need to extend the life of their current Blackwell (B200)-based systems or turn to alternative suppliers. This could increase demand for Nvidia's H100/B200 GPUs, keeping their prices high. According to SemiAnalysis calculations, the cost of operating a cluster of 10,000 H100 GPUs is approximately $150 million annually in electricity and cooling, and the delay could increase these costs by preventing migration to more efficient systems. For end users, this could translate into less available compute capacity for AI applications, slowing innovation in areas like large language models, video generation, and scientific simulations. For example, startups like Anthropic and Cohere, which rely on Nvidia's capacity to train their models, could see delayed launches. Additionally, hyperscalers like Microsoft and Google may accelerate their custom chip efforts. Microsoft has already developed the Maia 100 chip, and Google has its TPU v5p, which directly compete with Nvidia. The Kyber delay could give them breathing room to improve their software and ecosystem.

Historical context

This is not the first time Nvidia has faced rack delays. The DGX GH200 rack, launched in 2023, also suffered delays due to cooling issues. However, the scale of the Kyber is much larger: it integrates 144 GPUs in a single cabinet, with direct liquid cooling (DLC) and a next-generation NVLink interconnect architecture. The delay underscores the difficulty of scaling AI systems beyond current limits of power density and bandwidth. In 2022, Nvidia delayed the launch of its H100 GPU due to packaging issues but resolved it within months. Now, the problem appears more complex, involving the integration of multiple third-party components. The BMC board is supplied by an external vendor, and redesign may require changes to the rack layout. Moreover, the Kyber delay comes at a time when Nvidia faces unprecedented demand: its revenue in the fourth quarter of 2025 reached $39.3 billion, up 78% year-over-year, according to its financial results. However, the company also faces regulatory pressures, such as the U.S. Federal Trade Commission's investigation into potential anticompetitive practices in the AI chip market.

What readers should know

• The delay is specific to the Kyber rack, not the Rubin Ultra chips themselves. Nvidia could launch the chips in other formats before 2028. For example, it could offer Rubin Ultra in HGX modules or as part of smaller DGX systems.
• The problem lies in a circuit board that manages power and data distribution among the GPUs. Precise technical details have not been disclosed, but sources indicate it is a BMC that controls telemetry and sequential power-on of the GPUs.
• The 2028 date is estimated; it could be adjusted if Nvidia finds a faster solution. Some analysts believe the delay could be only a few months if the redesign is minor.
• Competitors like AMD with its Instinct MI400 line and the open-source RISC-V AI project could seize this window. AMD has announced that its MI400 will have performance comparable to Rubin Ultra, and companies like Tenstorrent are developing RISC-V-based AI chips.

"Every delay in Nvidia's roadmap is an opportunity for the ecosystem to diversify its compute sources," says an analyst at SemiAnalysis. "The AI market cannot rely on a single supplier for its critical infrastructure."

Conclusion

The Kyber rack delay is a reminder that Moore's Law is no longer enough: AI innovation now depends on complete system engineering. While Nvidia resolves this hurdle, the market must adapt to a slower pace of innovation than expected. This incident also highlights the fragility of the supply chain for critical components like BMC boards, which can become bottlenecks. For investors, the news could generate volatility in Nvidia's stock, which has risen over 200% in the past year. However, the company has a track record of overcoming technical obstacles, and its market dominance will not be threatened in the short term. In the long run, the delay could accelerate the adoption of alternative architectures and foster competition, ultimately benefiting the AI industry as a whole.