Nvidia's Kyber rack delay to 2028 opens door for AMD and Alphabet

Nvidia's Kyber rack delay to 2028 opens a competitive window for AMD and Alphabet in high-end AI infrastructure.

Nvidia's Kyber NVL144 rack — designed to pack 144 of its most powerful chips into a single cabinet — has been delayed by more than 12 months to 2028, creating a competitive opening for Advanced Micro Devices and Alphabet in the high-end AI infrastructure market.

"The delay raises uncertainty around Nvidia's next-generation scale-out roadmap and creates a wider competitive window for alternative AI platforms," Shawn Oh, who leads Korea cash equities at NH Investment & Securities Co. in Seoul, said.

The holdup stems from manufacturing challenges with a printed circuit board midplane that connects eight Oberon racks through the NVSwitch fabric, according to research firm SemiAnalysis. The board stacks three 26-layer sections into a 78-layer laminate close to one square meter in area, with trace spacing at or below 25 micrometers to maintain 448 Gb/s-class signaling. Nvidia also abandoned NVL72x2, a stopgap design that would have bolted two Oberon racks back-to-back, after major cloud customers pushed back on the operational complexity of running linked cabinets as a single unit, SemiAnalysis said. A separate configuration tying eight Kyber racks together through co-packaged optics, NVL576, faces its own delays or volume constraints.

The delays leave Nvidia without a proven way to scale its Rubin Ultra architecture beyond what the current Oberon rack already delivers in 2027, according to SemiAnalysis. That opens the door for AMD, whose improved ROCm software platform and chiplet designs position it to offer competitive high-end server solutions for inference workloads, and Alphabet, whose next-generation Tensor Processing Units will be optimized for both training and inference. Nvidia shares rose about 1.4 percent Monday after the company said its "road map is intact," though the stock trimmed gains during the session.

A $78-Layer Problem

The PCB midplane at the heart of Kyber replaces the cable harnesses of earlier racks with a rigid board carrying the all-copper NVLink fabric. Every GPU-to-GPU link inside the cabinet runs through that board, and copper traces lose signal integrity as layer counts increase, alongside power delivery and thermal design challenges. Jensen Huang held up the gray backplane on stage at GTC in March. A cabled version of the same interconnect would need upward of 20,000 discrete cables, which is why Nvidia is moving the wiring onto a single passive board.

The manufacturing difficulties have already rippled through the supply chain. Shares of Nvidia-dependent PCB manufacturers tumbled across Asia on Monday, with Japan's Ibiden Co. losing as much as 10 percent, Hong Kong's Kingboard Laminates Holdings falling 18 percent, Taiwan's Elite Material Co. dropping 10 percent, and South Korea's Samsung Electro-Mechanics sliding 11 percent. The selloff followed extraordinary year-to-date gains — Kingboard Laminates had surged more than 470 percent and Samsung Electro-Mechanics over 600 percent before Monday's reversal, according to Bloomberg data.

Competitors See an Opening

AMD's biggest challenge has always been the software ecosystem gap with Nvidia's CUDA platform. But the company has greatly improved its ROCm software over the past few years, and the shift of programmers working higher up the software stack using open-source frameworks like OpenAI's Triton has helped close the gap, especially for inference workloads. AMD's chiplet designs, which can package more memory onto its chips, and its recent acquisition of memory optimization software platform Mext, position it to offer a high-end server solution designed specifically for inference.

Alphabet's Tensor Processing Units have become highly regarded AI chips, and with its next generation, the company will have chips optimized for both training and inference. A delay in Kyber could make its cost-efficient TPU offering more attractive to customers seeking an optimized system without the premium pricing or delivery risk of Nvidia's platform.

Neither AMD nor Alphabet needs to displace Nvidia as the king of AI infrastructure to benefit. Capturing even a slice of the high-end market would provide a meaningful boost. AMD has already secured some large GPU deals and stands to benefit from rising demand for data center central processing units as agentic AI workloads grow. Alphabet, meanwhile, offers the most complete AI stack, with both world-class chips and AI models, giving it a cost advantage.

Nvidia shares trade at roughly 35 times forward earnings, reflecting the market's expectation that its aggressive product roadmap will sustain its dominance. The Kyber delay introduces execution risk to that thesis, even if the company's existing Rubin systems — now in full production with deliveries to eight major cloud customers including Amazon Web Services, Microsoft Azure and Google Cloud scheduled to begin this fall — provide a buffer. For AMD and Alphabet, the window to gain traction in high-end AI infrastructure has widened, though both must still prove they can deliver at Nvidia's scale.

This article is for informational purposes only and does not constitute investment advice.