Nvidia Kyber NVL144 delayed to 2028 on PCB defects

Key Takeaways

• Nvidia's Kyber NVL144 AI server rack is delayed to 2028 due to PCB midplane defects
• Asian suppliers lost up to 18% in stock value following the SemiAnalysis report
• AMD and Google gain a timing window as Nvidia also scraps the NVL72x2 rack and Rubin Ultra four-die chip

Nvidia's next AI server rack, the Kyber NVL144, has been pushed back more than twelve months to 2028. The cause: circuit board defects that have proven nearly impossible to eliminate at scale. The news, broken by analyst firm SemiAnalysis, sent shockwaves through Asian supply chains.

The report landed just three months after CEO Jensen Huang unveiled the Kyber NVL144 at Nvidia's GTC conference. For a company riding the biggest AI infrastructure boom in history, the delay is a rare stumble. It also hands competitors a window they didn't expect.

What caused the Kyber NVL144 delay?

The problem sits in the PCB midplane. This is the central circuit board that connects all individual components inside the rack. According to SemiAnalysis, manufacturers cannot produce it without defects at acceptable yields. When you're cramming 144 GPUs into a single system with extreme power and thermal demands, even minor PCB flaws become showstoppers.

The NVL144 was designed to double the GPU count of Nvidia's current flagship, the NVL72. That's not just twice the chips. It's exponentially more complexity in power delivery, cooling, and signal integrity. The midplane sits at the heart of all of it.

Asian suppliers took the hit

Markets reacted fast. Ibiden, a Japanese PCB maker that counts Nvidia as its largest customer, dropped as much as 10%. Kingboard Laminates fell 18% in Hong Kong. Elite Material lost 10% in Taiwan. Samsung Electro-Mechanics slid 11% in South Korea.

Context matters here. These stocks had been on a tear. Samsung Electro-Mechanics had gained more than 600% this year. Kingboard Laminates was up over 470%. Much of the selling looks like profit-taking from investors who had been waiting for any excuse.

Gary Tan of Allspring Global Investments told Bloomberg that a Kyber delay doesn't necessarily mean AI spending will shrink. The current weakness is mostly profit-taking, not a fundamental shift.

Nvidia scrapped other designs too

The Kyber delay isn't the only setback. SemiAnalysis reports that Nvidia has killed the NVL72x2, a planned design that would have placed two Oberon racks back to back. Cloud providers and data center operators pushed back. The form factor was unusual, and the operational overhead was too high.

The Rubin Ultra chip also got downsized. The four-die version with maximum compute has been canceled. Only the two-die version remains, delivering roughly half the real-world performance. For teams planning around Rubin Ultra's full specs, that's a material change.

Perhaps most significant: a key interconnect technology called CPO-NVSwitch won't arrive until Feynman, the generation after Rubin. CPO-NVSwitch was supposed to link many chips into a single large system. Without it, Nvidia lacks a proven way to scale Rubin Ultra to very large clusters.

AMD and Google get a timing advantage

This is where the delay gets strategic. AMD's MI500X and Google's TPUv8i Broadfly now have more runway. When the market leader stumbles on its next-generation system, competitors don't need to beat Nvidia's best hardware. They just need to ship.

Nvidia's response is to sell more Oberon-Rubin racks in the existing form factor. That buys time, but it doesn't close the gap on interconnect technology or rack density. Teams evaluating multi-year infrastructure investments now have to factor in the revised timeline.

Shawn Oh of NH Investment & Securities pointed to growing uncertainty around Nvidia's expansion plans. That uncertainty gives alternative AI platforms more room to compete.

What this means for AI infrastructure planning

For teams building AI products, the practical question is procurement timing. If you were planning around NVL144 availability in 2027, you're now looking at 2028. That changes capacity planning, budgeting, and potentially which hardware you choose.

The delay also highlights supply chain concentration risk. Nvidia's dominance means that when their roadmap slips, the entire industry feels it. PCB manufacturers, memory suppliers, and cooling system vendors all built capacity around Nvidia's timeline.

Frequently Asked Questions

When will Nvidia's Kyber NVL144 be available?

According to SemiAnalysis, the NVL144 has been pushed back to 2028, more than a year later than originally planned.

What caused the Nvidia NVL144 delay?

The PCB midplane, a central circuit board connecting all components, has proven extremely difficult to manufacture without defects.

Which Nvidia products were canceled alongside the Kyber delay?

Nvidia scrapped the NVL72x2 rack design and the four-die version of the Rubin Ultra chip. The CPO-NVSwitch interconnect is delayed until the Feynman generation.

How does the Nvidia delay affect AMD and Google?

Both companies gain a timing advantage. AMD's MI500X and Google's TPUv8i Broadfly face less pressure to match NVL144 specs in the near term.

Why did Asian tech stocks drop after the SemiAnalysis report?

PCB and component suppliers that had rallied on Nvidia demand saw sharp profit-taking. Kingboard Laminates fell 18%, Ibiden 10%, and Samsung Electro-Mechanics 11%.

Source: The Decoder / Maximilian Schreiner