Nvidia’s suppliers are accelerating production of their flagship AI data center, “Rack,” following the resolution of technical issues that delayed shipments as US chip makers strengthen their global sales drive.
Partners at the semiconductor giants, including Foxconn, Inventec, Dell and Wistron, have created a series of breakthroughs that will allow Nvidia to start shipping its highly anticipated “Blackwell” AI servers, according to several people familiar with the development of the group.
The recent revisions have driven the CEO of Jensen Fan, who announced Blackwell last year, pledging to significantly increase the computing power required to train and use large-scale language models.
The technical issues that emerged at the end of last year disrupted production and threatened the ambitious annual sales targets of US companies.
The GB200 AI rack includes 36 “Grace” central processing units and 72 Blackwell graphics processing units connected via NVIDIA’s NVLink communications system.
Nvidia’s GB200 Grace/Blackwell CPU/GPU unit is on display at Computex Conference© Annabelle Chih/Bloomberg in Taiwan
At the Computex Conference in Taipei last week, Nvidia’s Taiwanese partners Foxconn, Inventec and Wistron said shipments of the GB200 rack began at the end of the first quarter. Production capacity is currently expanding rapidly, they added.
“Our internal testing showed connectivity issues. The supply chain worked with NVIDIA to resolve issues that occurred two or three months ago,” said one engineer from Nvidia’s partner manufacturer.
The development is ahead of Nvidia’s quarterly revenue on Wednesday, with investors watching signs that Blackwell’s freight is ongoing at PACE following the initial technical issues.
Saudi Arabia and the United Arab Emirates recently announced plans to acquire thousands of Blackwell chips during President Donald Trump’s Gulf tour.
Nvidia’s supply chain partners have spent months tackling several challenges with their GB200 racks, including overheating and leaking liquid cooling systems due to 72 high-performance GPUs.
The engineers also cited software bugs and chip-to-chip connectivity issues that are caused by the complexity of synchronizing such a large number of processors.
“This technology is extremely complex. No company has tried to run these many AI processors simultaneously in such a short time frame.”
“Nvidia didn’t allow sufficient supply chain time to fully prepare, which resulted in delays. The inventory risk around the GB200 will be mitigated as manufacturers increase rack output later in the year,” Chu added.
To ensure smooth deployment for key customers such as Microsoft and Meta, suppliers will strengthen their testing protocols before shipping and perform further checks to ensure racking capabilities for AI workloads.
Nvidia is also preparing to deploy its next-generation GB300 AI racks. It has enhanced memory capabilities and is designed to handle more complex inference models such as Openai’s 01 and Deepseek R1. Huang said last week that the GB300 will be released in the third quarter.
Nvidia is preparing to deploy the next-generation GB300 AI system, photo © David Paul Morris/Bloomberg
To accelerate deployment, Nvidia has compromised on the design aspects of the GB300. Initially, they planned to introduce a new chipboard layout known as “Cordelia.”
However, in April, the company told partners that it would return to the previous “Bianca” designs that were used in current GB200 racks due to installation issues, according to the two suppliers.
This decision could help Nvidia achieve its sales goals. In February, the company said it was aiming for sales of approximately $43 million for the quarter ending April. This is a record number, with a 65% increase from the previous year.
Analysts say the Cordelia Committee has provided better margin possibilities and made maintenance easier for customers.
Nvidia has not abandoned Cordelia and has notified suppliers who are planning to implement the redesign within the next-generation AI chip, according to three people familiar with the issue.
Separately, NVIDIA is working to offset China’s revenue loss after the US government banned exports of H20 chips, a watering version of AI processors. The company said it expects that inventory amortization and purchase commitments will result in a $5.5 billion bill in claims related to the ban.
Last week, Bank of America analyst Vivek Arya wrote that if China’s sales hits, it would cut Nvidia’s gross profit margin for the quarter to around 58% from the previously shown 71%.
However, he writes that there is a faster rollout than Blackwell expected as the company that has returned to Bianca Boards could help offset China’s revenue later this year.
