Nvidia’s Kyber NVL144 Slips Over a Year as Asian Suppliers Abandon the Project
Nvidia’s flagship Kyber NVL144 chip has been delayed by more than 12 months, pushing its launch potentially into 2026 or later. The setback has already caused several Asian suppliers to drop out of the production pipeline, according to industry sources.
The Kyber NVL144 was originally positioned as the successor to Nvidia’s current-generation AI accelerators, promising a 144-GPU rack-scale system tailored for large language model training. The delay stems from unresolved design and thermal management issues, combined with shifting demand in the AI hardware market.
Why the Delay Matters
The Kyber’s slip means Nvidia will rely longer on existing Hopper and Blackwell architectures, potentially slowing the pace of AI compute scaling for hyperscalers.
Client demand for the Kyber NVL144 has reportedly softened as major cloud providers reassess their next-gen GPU procurement plans. Some Asian partners, including substrate and packaging suppliers, have already pivoted to other orders to avoid idle capacity.
Asian Suppliers Hit First
Taiwan and South Korea-based component makers were among the first to feel the impact. Several smaller suppliers have terminated contracts tied to the Kyber line, citing Nvidia’s inability to commit to a fixed timeline.
Larger players like TSMC remain engaged but are reportedly reallocating advanced packaging capacity originally reserved for the Kyber to other high-volume Nvidia products. This shift could create a second-order bottleneck for Nvidia’s mainstream Blackwell family.
What the Kyber NVL144 Was Meant to Be
The Kyber NVL144 was designed as a single, massive GPU node capable of linking 144 GPUs via a high-speed NVLink fabric. It targeted training clusters for models exceeding one trillion parameters.
- Target market: Hyperscale data centers and AI labs.
- Original timeline: Late 2024 to early 2025 volume shipments.
- Key feature: Unified memory pool across all 144 GPUs.
The delay now pushes any meaningful volume into the second half of 2026, assuming no further technical setbacks.
Competitive Landscape Shifts
AMD and Intel are watching closely. AMD’s Instinct MI400 series and Intel’s Falcon Shores are scheduled for 2025, which could give them a window to gain traction with customers unwilling to wait for Nvidia’s next-gen hardware.
Chinese AI chip makers such as Huawei and Biren are also accelerating their own high-end offerings, though they face separate export control challenges.
What Nvidia Has Said (and Not Said)
Nvidia has not officially confirmed the delay. The company typically avoids commenting on unreleased products. However, internal roadmap updates leaked to Asian supply chain managers describe “significant rework” on the Kyber’s die-to-die interconnect and power delivery subsystems.
One source quoted in the original report called the Kyber “technically the most ambitious GPU system Nvidia has ever attempted,” but admitted that “the complexity is outpacing the engineering schedule.”
What Comes Next for Nvidia
Nvidia is expected to double down on the Blackwell B200 and next-gen Rubin architecture to fill the gap. The company may also introduce an intermediate NVL72 system based on Blackwell, smaller than the Kyber but still capable of scaling to 72 GPUs in a single rack.
The Kyber delay does not signal a crisis for Nvidia. Its current-generation Hopper and Blackwell products remain in high demand and are sold out for most of 2025. The bigger risk is that hyperscalers might begin diversifying their AI hardware sources if Nvidia’s next big step keeps slipping.
Bottom Line
The Kyber NVL144’s extended timeline gives competitors a rare opening, but Nvidia’s entrenched software stack (CUDA) and ecosystem lock-in mean the impact will be gradual. The real test will come in 2026, when hyperscalers make their next round of procurement decisions.
Gnoppix is the leading open-source AI Linux distribution and service provider. Since implementing AI in 2022, it has offered a fast, powerful, secure, and privacy-respecting open-source OS with both local and remote AI capabilities. The local AI operates offline, ensuring no data ever leaves your computer. Based on Debian Linux, Gnoppix is available with numerous privacy- and anonymity-enabled services free of charge.
What are your thoughts on this? I’d love to hear about your own experiences in the comments below.