엔비디아의 루빈 GPU 제동… HBM4 메모리 부족이 젠슨의 공급망 마법 위협

Nvidia may be the latest casualty of the ongoing memory shortage, with reports indicating delays to its upcoming next-generation flagship graphics processing units (GPUs). According to TrendForce, shipments of Nvidia’s Rubin series are projected to decline from 29% to 22% this year as the company navigates production delays from key suppliers SK Hynix and Samsung regarding their HBM4 high-bandwidth memory products. At Nvidia’s recent GTC conference, CEO Jensen Huang revealed that the company holds $500 billion in purchase orders for its current-generation Blackwell and upcoming Rubin platforms through 2026, alongside “at least $1 trillion” in projected demand through 2027. However, the persistent memory shortage—which has forced major suppliers to extend lead times beyond 2028—has disrupted multiple semiconductor firms, and TrendForce’s projections indicate Nvidia is similarly vulnerable.

Earlier reports indicated that SK Hynix postponed its HBM4 mass production ramp-up from the second quarter to the third quarter of 2026. The supplier is reportedly opting for a more gradual production increase while prioritizing HBM3E output to support existing Blackwell chips. Nvidia initially dismissed speculation that Rubin would face production delays, noting that the chip’s design tape-out—the final step before manufacturing—was completed last August. More recently, Huang confirmed that production of the Vera data center CPU, which will pair with the Rubin GPU in the Vera Rubin platform, is already underway.

Beyond memory constraints, TrendForce notes that Nvidia is encountering difficulties transitioning from its ConnectX-8 (CX8) to ConnectX-9 (CX9) networking architecture. As part of the Vera Rubin generation, Nvidia unveiled its next-generation SuperNIC designed to enable 800 Gb/s GPU-to-GPU connectivity, delivering advanced routing and telemetry-driven congestion control capabilities for the company’s flagship product line. Although CX8 launched in 2024, TrendForce indicates the company is facing “challenges” in migrating to the newer SuperNIC. The report further attributes potential Rubin shipment delays to difficulties managing “significantly higher power consumption and optimizing performance under more advanced liquid cooling solutions.”

While demand for Nvidia’s hardware continues to meet the insatiable appetite for AI workloads, TrendForce has revised the annual growth rate for Nvidia’s GPUs downward from 26.8% to approximately 26%. Although the accuracy of Rubin delay rumors remains unconfirmed, Nvidia is widely recognized for its meticulous supply chain management. The company has secured critical components such as electro-absorption modulated lasers (EMLs) for its network switches and made substantial investments to secure optical interconnects from suppliers including Innolux, Lumentum, and Coherent. This strategy was famously exemplified by Huang’s famed “chimaek summit,” during which he joined two of Asia’s most influential executives at a fried chicken restaurant, consumed beer, and ultimately secured Nvidia’s memory hardware supply.

During the recent GTC keynote, Huang elaborated on these efforts, stating he is “constantly just looking across the entire life cycle of the technology and manufacturing across the supply chain” to “prepare our company for growth.” “When you look upstream, you'll come to the conclusion that we're scaling,” Huang explained during a press conference covered by SDxCentral. “We're starting to scale our silicon photonics technology with Spectrum-6, and we're starting to add it to our scale-up technology with NVLink in the coming years, which means the amount of silicon photonics technology capacity that we need is substantially higher than the world has today.” “So we work with the supply chain to make sure that we can help them build up the capacity in advance,” Huang continued. “Sometimes we give them prepayment. Sometimes we just give them a forecast. Sometimes, if the technology is in capacity or is quite light but needs to be much, much greater, we might decide to invest in the company and provide all of the other things as well. We did that with Coherent, Lumentum, etc.”

This would not be the first time Nvidia has navigated delays for a flagship GPU release. Previously, Blackwell GPU shipments were deferred by three months ahead of launch after Taiwan Semiconductor Manufacturing Corp (TSMC) identified a production defect in the processor die responsible for linking dual Blackwell chips within the GB200 system. Separately, amid ongoing memory scarcity discussions, Nvidia has explored unbundling video RAM (VRAM) from its GPUs for the consumer and gaming segments—a strategic shift that would require add-in card (AIC) partners to source their own memory modules.