Suppliers including Dell, Foxconn, Inventec, and Wistron have ramped up production of Nvidia Grace Blackwell racks, having resolved technical issues that had previously delayed shipments of the product.

Per a report in the Financial Times, the companies have made a “series of breakthroughs,” allowing them to start shipping the GB200 servers to customers.

According to comments made by an engineer from one of Nvidia’s partner manufacturers at this month’s Computex event in Taiwan, internal tests showed “connectivity problems” with the AI servers, but the suppliers have been able to work with Nvidia, and the issue was ultimately solved two to three months ago.

This is not the first time problems with Nvidia’s Blackwell chips have been reported. In late 2024, it was claimed that Nvidia’s Blackwell GPUs were overheating when linked together in 72-chip data center racks, also known as NVL72, with the chip designer reportedly asking suppliers to make multiple changes to its custom rack design in a bid to solve the issue.

Software bugs, inter-chip connectivity problems, and leaking liquid cooling systems were also reported by engineers, the FT said. As a result, suppliers have increased the number of testing protocols the racks are subject to before shipping them to customers.

The technical fixes have been implemented at a time when Nvidia is preparing to deploy its GB300 AI rack. The Nvidia GB300 NVL72, a fully liquid-cooled rack-scale offering, featuring 72 Nvidia Blackwell Ultra GPUs and 36 Nvidia Grace CPUs, is expected to launch in Q3 2024.

However, the FT claimed that in order to accelerate deployment of the incoming GB300 servers, Nvidia has compromised on aspects of its design, including reverting to the chip board layout – dubbed Bianca – currently used in the GB200 rack, as opposed to moving forward with its new Cordelia layout, which would allow for the replacement on individual GPUs, making maintenance of the system easier.

Citing three people familiar with the matter, the FT reported that Nvidia has told suppliers it now intends to implement the Cordelia design in its next-generation AI chips.

Elsewhere this week, it was reported that the chipmaker was planning to launch new Blackwell-based chipset for Chinese market, following the recent decision from the US government to impose further restrictions on the company’s H20 chips – less sophisticated versions of Nvidia’s H100 processors that had also been designed specifically for the Chinese market in compliance with US export controls.

Nvidia, which said it would see $5.5 billion of charges associated with H20 products for inventory, purchase commitments, and related reserves as a result of the restrictions, will post its Q1 2026 results later today.