Nvidia disclosed a more than 12-month delay to its Kyber NVL144 rack-scale AI system on Monday, pushing the high-end infrastructure product to 2028 amid manufacturing challenges with a complex circuit board.
What Went Wrong With Kyber
The Kyber’s PCB midplane — an orthogonal circuit board with up to 78 layers — proved harder to mass-produce than Nvidia anticipated. The complexity of the design has created a manufacturing bottleneck that forced the company to push back its timeline. The system was originally slated to ship in 2027 alongside Nvidia’s Rubin Ultra processors.
Nvidia developed an interim workaround called NVL72x2, which bolted together two current-generation NVL72 racks to approximate Kyber’s compute capacity. But cloud providers and hyperscalers rejected the approach, saying it was operationally awkward and too expensive to justify the investment.
Rivals Smell an Opening
The delay cracks open a rare window for AMD, Google, and other chip rivals to chip away at Nvidia’s grip on high-end AI infrastructure. With Nvidia’s premium product off the table for another 18 months, cloud giants running trillion-dollar AI bets now have reason to look elsewhere for their next-generation systems.
Markets reflected the concern. Nvidia shares fell 1.4% when the manufacturing issue surfaced, wiping billions in value in a single session. But on July 7, Nvidia responded directly to the delay rumors, saying its roadmap remained unchanged and reaffirming confidence in the Kyber timeline.
The delay caps a bruising week for AI semiconductor bets. Meta announced it’s leasing excess data center capacity, and Samsung’s earnings beat expectations by only 6%. For cloud operators betting billions on next-gen gear, the Kyber timeline slip means they’re still years away from Nvidia’s newest flagship system.
