Nvidia Denies 12-Month Delay on Kyber Rack After SemiAnalysis Report, But Asian PCB Stocks Fall

Nvidia denied on Monday, July 6, a report by research firm SemiAnalysis that cited a delay of over twelve months for the Kyber NVL144, the next-generation rack designed to host the Rubin Ultra chips in 2027. According to SemiAnalysis, a manufacturing issue with the multilayer orthogonal PCB backplane would be the bottleneck, pushing the product to 2028. "Our roadmap remains unchanged," responded a Nvidia spokesperson hours later, without providing project-specific details.

The episode started in the early hours of Asian trading and derailed the stocks of the printed circuit board supply chain in South Korea, Taiwan, and Japan. The Kospi led the decline of the subsector. The drop comes ahead of SK Hynix's debut on Nasdaq, with a formal marketing phase starting on this Monday and a revised target of $28.21 billion, making the offering the largest ADR ever issued, surpassing Alibaba's $21.8 billion in 2014. Nvidia's response brought stocks back to positive territory in New York, but the PCB sector closed the day lower.

What’s at Stake with the Kyber Rack

The Kyber NVL144 is the rack that anchors the Rubin roadmap presented at GTC 2026 and is Nvidia's first project to integrate eighty GPU modules in a single densely cooled unit, a leap in architecture that requires eighteen-layer printed circuit boards and tighter tolerances than those used in the Blackwell GB200. SemiAnalysis argues that current suppliers have been unable to stabilize yields above a commercial level. Nvidia contests this interpretation and states that wafer allocation at TSMC and the qualification curve for HBM4 remain on track.

The importance of the signal lies in the hyperscalers' purchasing calendar. Microsoft, Amazon, Google, and Meta have signed multi-year agreements tied to the availability of Rubin, and a slip to 2028 would require infrastructure areas to extend the lifespan of Blackwell clusters beyond what was planned or diversify into AMD MI400 and Google’s internal ASICs TPU v6, Amazon's Trainium 3, and Microsoft's Maia 200. None of these alternatives are ready today to absorb the training demand for models like OpenAI’s GPT-5.6 Sol, still in restricted preview for U.S. government-approved partners, or Anthropic's Claude Fable 5, globally redistributed on July 1.

A Signal for the Buy-Side, Not for Operations

There are two plans that the controversy intertwines. On the industrial side, a twelve-month delay on the rack does not alter Nvidia's short-term economics, whose revenue from Blackwell over the next six quarters is already locked in. On the buy-side, the SemiAnalysis report reinforces the hypothesis posited by analysts from Morgan Stanley and JPMorgan that part of the current super-cycle reflects defensive stocking by hyperscalers and overly optimistic qualification timelines. The decline in Asian PCB stocks on this Monday shows that the market is pricing in this risk within the supply chain, even with Nvidia's formal denial.

From the hardware vendor's side, Jensen Huang's response to the same public thesis is well-known: Nvidia argues that any internal delay would be offset by subsequent demand absorption, meaning that the training market remains constrained by supply at least until the end of 2027. The company reiterated this message in a press briefing at the end of June and reaffirmed it on Monday. The more honest interpretation is that both sides may be right at the same time, and what changes is the timing of what is already certain.

Global Chain Impact

In the United States, hyperscalers are pushing for more flexible contractual terms for Rubin, including partial replacement clauses with their own ASICs. In Taiwan, TSMC is operating near the capacity ceiling of CoWoS-L reserved for Nvidia and is not expected to alter mix even in the event of a delay. In South Korea, Samsung and SK Hynix have already announced HBM4 capex for 2027, a decision that would become riskier if the Kyber moves to 2028. In Japan, the Ministry of Economy is monitoring the process due to the involvement of Japanese high-temperature resin suppliers and the recent approval of access to Anthropic's Claude Mythos and OpenAI’s GPT-5.5-series by MUFG, SMBC, and Mizuho for cybersecurity. The global silicon supply chain is so exposed to the Rubin roadmap that any sign of slippage triggers repricing, even if denied.

For CIOs, the operational effect is immediate. Capacity reservation contracts in Azure, AWS, and Google Cloud will return to the table in September for renegotiation of training windows starting in 2027, and any confirmed signs of delays will push prices and timelines up. A good stress test for the 2027 budget is simulating the impact of twelve additional months of GPU costs under the current architecture. The number is unsettling, and that is exactly why it matters today.