Nvidia's Kyber NVL144 Rack System Delayed Over 12 Months to 2028 on PCB Manufacturing Snags

Kyber is the codename for Nvidia's rack-scale architecture that integrates 144 of its forthcoming Rubin Ultra GPUs into a single server cabinet. The design uses a dense all-copper NVLink interconnect enabled by the PCB midplane, allowing the chips to operate as a single massive compute unit — the kind of consolidated processing power that hyperscalers need to train and run frontier AI models ^[1][3].

The midplane board is the critical bottleneck. At 78 layers, it is among the most complex PCBs ever designed for a commercial computing product. SemiAnalysis identified challenges spanning signal integrity, power delivery, thermal design, and manufacturable board layer counts — issues that together have pushed production readiness beyond what Nvidia and its supply chain partners can resolve in the original timeframe ^[2][3].

Nvidia had developed an interim design called NVL72x2 that bolted together two of its current-generation NVL72 racks to approximate Kyber's compute density. That plan has been canceled after cloud providers and hyperscalers pushed back, deeming the configuration operationally awkward and prohibitively expensive ^[1][2].

The larger NVL576 system — which connects eight Oberon racks via co-packaged optics (CPO) — faces its own timeline pressure. SemiAnalysis indicated the system may also be delayed or limited to low-volume shipments given ongoing challenges with CPO technology. Meanwhile, the Rubin Ultra chip itself has been scaled back from a quad-chip to a dual-chip variant ^[3].

Nvidia continues shipping its current Oberon and Rubin rack systems to hyperscalers and enterprise customers, maintaining revenue fulfillment on existing product lines ^[3].

The delay opens a rare window for competitors at the high end of the AI infrastructure market. AMD and Google's custom TPU silicon stand to benefit if hyperscalers reallocate procurement budgets while waiting for Kyber ^[2][3].

Despite the product-level setback, SemiAnalysis struck a bullish note on Nvidia's near-term financials, projecting that the company's data-center compute revenue will exceed Wall Street consensus estimates by 20% in the second half of fiscal 2027. That projection suggests current-generation products remain in heavy demand even as the next-generation timeline slips ^[4].

The report sent ripples through the Asian supply chain. Bloomberg reported that PCB manufacturers in Taiwan saw their shares decline on Monday as investors recalibrated expectations for orders tied to the Kyber program ^[5].

Infrastructure procurement teams at major cloud providers now face capacity-planning uncertainty for the 2027–2028 cycle. The delay forces hyperscalers to either extend deployments of current-generation Nvidia hardware or evaluate alternative architectures from AMD, Google, and custom ASIC providers ^[2][3].

Nvidia has not publicly confirmed or denied the delay. The company's next scheduled earnings report and any updated product roadmap commentary from management will be closely watched for official guidance on the Kyber and Rubin Ultra timelines ^[1].