Nvidia's Kyber Rack Slips to 2028, Handing Rivals a Rare Opening at the AI Hardware Frontier

For three years, Nvidia has shipped a new generation of artificial intelligence hardware almost annually, a relentless drumbeat that rivals could not match and investors priced in as near-certainty. That rhythm has now collided with the physical limits of modern manufacturing. According to research firm SemiAnalysis, whose findings were subsequently reported by CNBC, Nvidia's next flagship rack-scale system — the Kyber NVL144 — has been delayed by more than 12 months, pushing its release date to 2028. The system, designed to house 144 of Nvidia's most powerful Rubin Ultra processors in a single cabinet and enabling them to function as one unified supercomputer, was originally scheduled to ship alongside the Vera Rubin Ultra platform in 2027.

The culprit is a single, unglamorous component: the printed circuit board midplane that connects the system's compute modules. SemiAnalysis reported that the PCB midplane at the centre of the Kyber design remains too difficult to produce reliably. According to analysis published by Proactive Investors, the board runs to approximately 78 layers, ranking it among the most complex ever attempted for a commercial product. Nvidia's alternative plan — pairing two existing Oberon racks back-to-back to replicate some of Kyber's performance — was also abandoned after cloud providers and hyperscalers pushed back against what SemiAnalysis described as an odd design and heavy operational burden. That cancellation, reported by CNBC, effectively removes any near-term stopgap and leaves Nvidia without a proven method to scale up its most powerful systems.

The knock-on effects are significant. SemiAnalysis also warned that the NVL576, a larger configuration intended to link eight Kyber racks through co-packaged optical connections, faces its own delays or production constraints. Separately, the Rubin Ultra chip itself has been scaled back from a four-chip to a two-chip design, roughly halving its peak performance even upon eventual delivery. Nvidia responded tersely to inquiries, with a spokesperson telling Tom's Hardware that the company's roadmap is intact, without elaborating on whether that statement referred to the original timeline or a revised one. Paul Triolo, a partner at consulting firm DGA-Albright Stonebridge Group, acknowledged the difficulty while noting that Nvidia has previously overcome comparable supplier challenges.

The timing sharpens the competitive stakes considerably. As Proactive Investors noted, the delay arrives barely three months after Nvidia chief executive Jensen Huang showcased Kyber on stage at the company's GTC conference. Meanwhile, the semiconductor sector is simultaneously digesting a broader market sell-off: as Forbes reported on July 8, chip stocks erased more than a trillion dollars in market value as Wall Street questioned the sustainability of record AI capital expenditure. Intel fell more than 20 percent, while memory chipmakers Sandisk and Micron dropped roughly 7 and 5 percent respectively after Samsung's record second-quarter results, reported by CNBC, nevertheless disappointed investors who had set an impossibly high AI bar. Against that turbulent backdrop, the Kyber slip deepens the sense of a sector hitting a moment of reckoning. On a more encouraging note for Nvidia, existing Rubin systems remain on schedule for autumn delivery to eight major cloud customers including Amazon Web Services, Microsoft Azure and Google Cloud, providing a commercial bridge even as the next generation stalls.

For competitors, the gap is an invitation. According to SemiAnalysis and analysis in The Next Web, the delay hands AMD and Google a rare opening at the premium end of the AI infrastructure market, where Nvidia's advantage was supposed to be widest. AMD's MI300 series has already been gaining traction among hyperscale customers seeking alternatives, and the Philadelphia Semiconductor Index's more than 47 percent year-to-date surge, reported by Intellectia.ai, reflects just how much investor capital has been staked on the assumption of uninterrupted Nvidia dominance. TSMC reports quarterly results on July 16, followed by Intel on July 23 — both sets of earnings will be closely watched as bellwethers for whether the sector can reframe the current turbulence as a mid-cycle reset, as several analysts have suggested, or whether the Kyber delay signals something more structurally troubling about the pace at which AI hardware ambitions are outrunning the engineering realities of production.