Nvidia´s suppliers have surmounted key production challenges with the Blackwell-based GB200 Artificial Intelligence server racks, accelerating manufacturing and shipments after a delay caused by technical complications. According to reporting from the Financial Times, supply chain partners such as Foxconn, Inventec, Dell, and Wistron have made significant breakthroughs, addressing hurdles that initially impeded the deployment of GB200 racks. Issues resolved include overheating, liquid cooling system leaks, and complex inter-chip connectivity problems encountered during internal testing late last year.
The production setbacks led to shipment delays, but announcements from Nvidia´s Taiwanese partners at Computex 2025 confirm that shipments officially began at the end of the first quarter, and production capacity is now ramping up rapidly. Engineers cited in the report revealed that stabilizing connectivity between components required months of cooperation between suppliers and Nvidia. Furthermore, the partners spent considerable time mitigating software bugs and synchronizing the operation of a large number of processors, which is critical for reliable data center performance.
Despite the progress on the GB200, Nvidia appears to be compromising on some aspects of its upcoming GB300 design to meet timelines. The report notes that Nvidia has dropped the planned Cordelia chip board layout for the GB300 in favor of the established Bianca design used in the GB200, a move aimed at reducing installation issues but precluding replacement of individual GPUs. Additionally, the postponement of SOCAMM memory technology, originally intended for the Blackwell Ultra GB300, is attributed to the design switch. Inventory risk for the GB200 is expected to subside in the second half of 2025, while Nvidia plans to introduce the Cordelia design with its next-generation Rubin chips.
