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 physical and manufacturing bottlenecks. In July 2026, the narrative around Nvidia's upcoming next-generation architectures—specifically the Vera Rubin and Rubin Ultra platforms planned for 2027—has split between detailed supply chain reports of major delays and firm management denials.
The Kyber NVL144 78-Layer PCB Manufacturing Bottleneck
On July 6, 2026, semiconductor research firm SemiAnalysis reported that Nvidia's next-generation Kyber NVL144 rack-scale AI system—designed to house 144 Rubin Ultra GPUs in a single unified compute domain—has been delayed by more than 12 months to 2028.
The source of the delay is a manufacturing failure at the absolute physical frontier of printed circuit board (PCB) fabrication. Unlike the current Vera Rubin NVL72 systems, which rely on approximately 5,000 copper cables totaling nearly two miles in length to link 72 GPUs, the Kyber architecture was designed to eliminate cables using an orthogonal backplane PCB. This board sits between vertically mounted compute trays and rear switch trays, routing all NVLink signals through rigid copper traces at SerDes speeds exceeding 448 gigabits per second per lane.
However, the physical specifications of this orthogonal backplane have proven too complex for current manufacturing capabilities:
- 78-Layer Stack: The board is roughly the size of a kitchen table and requires laminating three separate 26-layer circuit boards together under extreme heat and pressure.
- Specialty Materials: The stack uses a hybrid of M9-grade copper-clad laminate, quartz fabric, and PTFE.
- Extreme Precision: Copper traces must be held to widths and spacings of 25 micrometers or less, with impedance controlled within a ±5% tolerance across the entire board.
According to CITIC Securities, this combination of extra-large form factors and extremely high layer counts presents challenges "significantly greater than conventional products in areas such as yield control, impedance consistency, and thermal management." No advanced PCB fabricator has been able to produce this board at acceptable production yields.
Rejected Backup Plans and Rubin Ultra Spec Downgrades
To bridge the gap, Nvidia proposed a contingency design called the NVL72x2, which would place two existing Oberon NVL72 racks back-to-back and link them via copper NVLink cables. However, cloud service providers (CSPs) and hyperscalers heavily rejected the design due to its odd configuration and heavy operational and cabling burdens, forcing Nvidia to cancel the backup plan.
This delay follows another major setback reported on June 30, 2026: the quiet cancellation of the original quad-chiplet version of the Rubin Ultra GPU. Due to advanced packaging limits at TSMC, Nvidia had to abandon the quad-chiplet design (which would have carried 16 HBM4E memory stacks for 1 terabyte of memory per package). Instead, Nvidia will ship a dual-chiplet version carrying roughly 8 HBM4E stacks, effectively halving the planned memory capacity.
Nvidia's Response and Near-Term Financial Strengths
Nvidia has swiftly and urgently denied the delay claims. On July 15, 2026, CEO Jensen Huang dismissed reports of delays in the Vera Rubin rollout, stating that the next-generation systems are "in production and on track" and affirming "giant" production volumes. The company's official response remains: "Our roadmap is intact."
Analysts note that these roadmap issues do not affect Nvidia's immediate financial trajectory:
- NVL72 Production: The current Vera Rubin NVL72 systems—built on the existing Oberon rack architecture, which is entirely unaffected by the Kyber midplane issue—are in full production and scheduled for delivery this fall (2026) to eight confirmed cloud partners (AWS, Microsoft Azure, Google Cloud, Oracle Cloud, CoreWeave, Lambda, Nebius, and Nscale).
- Strong Financial Performance: In the quarter ending April 30, 2026, Nvidia reported a record $81.61 billion in revenue (+85.2% YoY) and $58.32 billion in net income, supported by a 74.1% gross margin and 63.0% profit margin (/markets/NVDA/2026/07/21).
- Valuation Support: Even with recent volatility, Nvidia's forward P/E multiple sits at a reasonable 21.7x, close to the S&P 500 average and well below its five-year average of 72x, while SemiAnalysis projects Nvidia's second-half fiscal 2027 data-center compute revenue will exceed Wall Street consensus estimates by 20%.
The high-stakes test of who is correct will arrive on August 26, 2026, when Nvidia reports Q2 FY2027 earnings and management is forced to address the Kyber timeline and PCB yields on the record.