Nvidia's AI Capex Sustainability: TSMC's Packaging Limits and the Kyber NVL144 78-Layer PCB Failure
The structural demand driving Nvidia’s artificial intelligence dominance remains robustly supported by physical supply chain signals, but its breakneck annual release cadence is hitting severe manufacturing bottlenecks at the physical frontier of materials science. While near-term financials remain insulated by the full production and upcoming fall shipment of its Vera Rubin NVL72 platform, Nvidia's longer-term scale-up roadmap has suffered its most significant technical setbacks yet.
The Kyber NVL144 Delay: Physics Defeats the Orthogonal Backplane
On July 6, 2026, semiconductor research firm SemiAnalysis reported that Nvidia's next-generation Kyber NVL144 rack-scale architecture—designed to pack 144 of its most powerful Rubin Ultra AI accelerators into a single unified compute domain—has been delayed by more than 12 months, pushing its commercial launch from 2027 into 2028.
The bottleneck is a manufacturing yield failure of its orthogonal backplane (PCB midplane), a massive circuit board roughly the size of a kitchen table that connects vertically mounted compute trays to rear-facing switch trays. To keep 448G-class SerDes signals intact without the massive cable bundles used in current-generation racks, the orthogonal backplane requires extreme specifications:
- Lamination of three separate 26-layer circuit boards into a single 78-layer stack.
- Copper traces no wider than 25 micrometers (one-quarter of a human hair).
- Specialty high-frequency materials including M9-grade copper-clad laminate, quartz fabric, and PTFE.
- Impedance controlled within a razor-thin ±5% tolerance across the entire board.
No advanced printed circuit board fabricator in the world, including Taiwan's most sophisticated PCB manufacturers, has been able to produce this board at viable production yields. Microscopic variations in laminate thickness or layer alignment at this level cause signal reflections and crosstalk that break high-speed links.
The Backup Plan and Customer Rejection
To bypass the midplane yield crisis, Nvidia developed a contingency design known as the NVL72x2, which placed two existing Oberon NVL72 racks back-to-back and linked them via copper NVLink. While this would have bridged the gap to deliver comparable GPU density, it introduced severe latency and heavy operational complexity.
Cloud Service Providers (CSPs) and hyperscalers heavily rejected the design. Faced with "heavy pushback... over its odd design and heavy operational burden," Nvidia canceled the NVL72x2, leaving the company with no proven solution to expand the scale-up world size for Rubin Ultra in 2027. Consequently, the larger NVL576 system (eight Kyber racks linked by co-packaged optics) has also been delayed or restricted to low-volume shipments.
Rubin Ultra Silicon Scaled Back
The Kyber delay follows another major setback on June 30, 2026, when SemiAnalysis reported that the original quad-chiplet version of the Rubin Ultra GPU had been quietly canceled. Due to TSMC advanced-packaging (CoWoS) constraints, Nvidia was forced to abandon the quad-die design, which would have carried 16 HBM4E memory stacks for 1 terabyte of memory per package. Instead, Nvidia will ship a dual-chiplet version with approximately 8 HBM4E stacks and half the memory capacity.
Together, these developments represent a substantial reduction in Nvidia's 2027 ambitions: a chip with half the originally planned memory capacity, housed in a rack architecture that is a full generation behind schedule.
Near-Term Financial Runway Remains Unmatched
Despite these longer-term roadmap snags, Nvidia’s current-generation financial trajectory remains exceptionally strong. According to its public-markets snapshot, Nvidia generated $81.61 billion in revenue and $58.32 billion in net income for the quarter ending April 30, 2026, reflecting year-over-year growth of 85.2% and 214.5% respectively, with a gross margin of 74.1% and a net profit margin of 63.0%.
The current Vera Rubin NVL72 systems—built on the existing Oberon rack architecture that is completely unaffected by the Kyber midplane issue—are in full production and scheduled for delivery this fall to eight confirmed cloud partners: Amazon Web Services, Microsoft Azure, Google Cloud, Oracle Cloud, CoreWeave, Lambda, Nebius, and Nscale. Even with the Kyber delay, SemiAnalysis projects Nvidia's data-center compute revenue will exceed Wall Street consensus estimates by 20% in the second half of fiscal 2027, highlighting that the physical demand for Nvidia's silicon remains insatiable in the near term.