Nvidia's next-gen AI rack system delayed to 2028 on manufacturing snags, SemiAnalysis says

SemiAnalysis reports Nvidia’s Kyber rack-scale architecture—planned for Vera “Rubin Ultra” in 2027—is delayed by 12+ months to 2028 due to manufacturability issues with a key PCB midplane. The larger NVL576 system may be delayed or limited to small volumes, and a backup design combining two current-generation racks was canceled after cloud customers rejected it as awkward and costly to operate. While Nvidia’s current Rubin systems begin shipping this fall and SemiAnalysis projects data-center compute revenue could run ~20% above Wall Street consensus in 2H FY2027, the roadmap setbacks raise concerns about scaling at the high end versus rivals.

Analysis

The key market issue is not near-term revenue leakage; it is whether Nvidia’s premium multiple is anchored to a product cadence that customers can now see may not be operationally executable at the rack level. A 12-month slip in a future system matters because the next leg of AI capex is about scale-up efficiency, not just single-chip performance, so any manufacturability bottleneck shifts bargaining power back toward buyers and alternative architectures.

That creates a cleaner opening for second-source strategies. AMD is the obvious external beneficiary if hyperscalers decide they need a credible hedge against Nvidia roadmap risk, while GOOGL, AMZN, and MSFT gain leverage from their in-house silicon programs because the constraint moves from raw FLOPS to systems integration and deployment complexity. The more painful the rack design becomes, the more attractive “good-enough and shippable” looks versus “best-in-class but delayed.”

Near term, the stock reaction can stay muted because current-generation shipments are still flowing and AI spend is still expanding; this is more of a 1-3 quarter narrative risk than a 1-2 week earnings issue. Over 6-18 months, though, repeated slips could compress Nvidia’s multiple if investors start modeling a slower transition to next-gen density and lower ASP expansion. The thesis is falsified if Nvidia continues to beat data-center consensus by double digits and hyperscaler commentary shows no change in deployment or sourcing plans.

Key Decisions for Investors

• Initiate a 3-6 month long AMD / short NVDA relative-value pair on any post-news retracement: thesis is that hyperscaler diversification creates optionality for AMD faster than this delay can be absorbed by Nvidia; stop if AMD fails to win incremental design momentum over the next two earnings cycles.
• Use NVDA weakness to build a modest short via put spreads rather than outright short stock: the near-term revenue stream is still intact, so the cleaner expression is multiple compression if roadmap credibility erodes again; risk is a continued AI capex melt-up that offsets execution concerns.
• Add GOOGL as a structural beneficiary on weakness: Google’s TPU stack becomes more valuable when external scale-up hardware gets messy, and Cloud can use custom silicon as a negotiating lever against NVDA pricing; hold horizon 6-12 months, with the main risk being broader cloud spend volatility.
• Watch AMZN and MSFT upcoming capex commentary for signs of faster custom-silicon adoption or slower NVDA rack rollouts; if either signals a meaningful reallocation away from Nvidia in the next 1-2 quarters, add to the NVDA short leg.