NVIDIA's GB300 AI server is finally expected to begin shipping in volume in September as the supply chain adjusts to NVIDIA's design changes. NVIDIA's "Blackwell Ultra" series of AI servers did encounter some issues when they were released in the first half of 2025, as the Green team introduced some design changes that made it difficult for the supply chain to adapt. This has resulted in sluggish shipments, with higher-end NVL72 AI clusters only available to NVIDIA's exclusive partners such as Dell and Microsoft.
Now, according to a report in DigiTimes, NVIDIA plans to begin mass production of its GB300 AI cluster in September, making the company's high-end clusters available to a larger market.
The main reason is that NVIDIA uses the Cordelia motherboard design on the GB300, which integrates modular design features and integrates a new SOCAMM memory design that has never been used before. However, due to NVIDIA's short product update cycle and problems with SOCAMM memory, the company decided to switch to the Bianca architecture, which is also used in the GB200.
While the market viewed this as a move of weakness, it turned out to be a significant step.
It is said that because the adjustment range of GB300 is relatively low compared with GB200, which has the same fundamentals, the supply chain no longer faces pressure to increase customer supply. NVIDIA's suppliers are currently testing low-volume GB300 shipments using Bianca motherboards, and production is expected to follow in the next few quarters, which means the fourth quarter will be the time for "Blackwell Ultra" to enter and potentially disrupt the AI industry.
The GB300 AI server has started to receive significant orders, especially from NVIDIA's "Sovereign AI" initiative, so the demand is definitely there. Furthermore, considering how quickly NVIDIA is upgrading its architecture, no one can match it right now. Rubin will be available in late 2026 or early 2026, and NVIDIA's current product upgrade cycle is 6 to 8 months, one of the fastest ever.