NVIDIA is gearing up to begin large-scale shipments of its GB300 AI servers by September, following adjustments to its design changes. This transition alleviates previous supply chain constraints, allowing for broader accessibility to these high-performance clusters.
Initially, the “Blackwell Ultra” AI servers encountered challenges upon their release in early 2025, primarily due to significant design modifications that strained the supply chain. As a result, only a few exclusive partners, like Dell and Microsoft, received the high-end NVL72 AI clusters. However, reports now indicate that NVIDIA is set to start mass production of its GB300 AI clusters, opening up access to a wider market.
The production issues were largely due to the introduction of the Cordelia board design in the GB300, featuring novel modular elements and the SOCAMM memory design. These elements were not previously adopted, leading to short-term production cycle dilemmas. Consequently, NVIDIA pivoted to the Bianca architecture, previously used with the GB200. Although initially viewed as a risky move, this shift proved to be strategic.
Suppliers no longer feel pressured by the demand for GB300 units, as the transition to the Bianca board required fewer adjustments compared to past changes. Current low-volume testing with the Bianca design is underway, and production is expected to ramp up in the coming quarters. This positions Q4 as a critical period for the “Blackwell Ultra” series to make an impact on the AI industry.
Orders for the GB300 AI servers are already surging, particularly fueled by NVIDIA’s “Sovereign AI” initiative, indicating strong market demand. With rapid architectural advancements, NVIDIA maintains a competitive edge with its swift six to eight-month product cycle. The anticipated introduction of the Rubin model around late 2025 or early 2026 further illustrates their momentum in the tech world.
