NVIDIA's Next AI Rack Delay Independently Confirmed

Research firm SemiAnalysis reported on July 5 that NVIDIA's Kyber NVL144 rack (a next-generation system that packs 144 GPUs into one massive computer) has slipped from 2027 into 2028. A related bigger system called NVL576 is also delayed. The problem is not the chips themselves. NVIDIA cannot reliably manufacture the giant circuit board that connects everything. An independent source in Taiwan with direct supply chain contacts confirmed the report. NVIDIA publicly denied it. The market split the difference: NVIDIA stock bounced 1% on the denial, but companies that make the circuit boards for these systems dropped by double digits because their customers know what is really happening.

The AI mania has been built on the assumption that NVIDIA can deliver whatever it promises. This delay proves that even NVIDIA hits real physical limits. When NVIDIA cannot ship the promised architecture, the entire chain of assumptions above it has to be recalculated by every hyperscaler placing orders. AMD is the biggest winner. Their competing product ships now and uses simpler technology that does not have the same manufacturing problems. Companies like Meta helped design AMD's alternative and cannot wait for NVIDIA to fix its problems. AMD gets roughly 18 months of open competitive window before their own next-generation product hits similar walls. The bigger picture matters more than the specific companies. If NVIDIA cannot deliver on schedule, memory demand growth curves flatten, hyperscaler capex assumptions get recalibrated, and the entire AI infrastructure story gets scrutinized. This is the first confirmed crack in the compute side of the AI trade.