NVIDIA's suppliers are accelerating the production of its flagship AI data center "racks" as the technical issue has delayed shipping as U.S. chipmakers exacerbate their global sales drive.
According to several familiar people, the semiconductor giant's partners, including Foxconn, Inventec, Dell and Worston, have made a series of breakthroughs that have enabled them to start shipping of Nvidia's highly anticipated "Blackwell" AI servers.
The latest fix is facilitated by CEO Jensen Huang, who announced Blackwell last year, promised to significantly improve the computing power needed to train and use large language models.
Technical issues that emerged late last year undermined their production, threatening the ambitious annual sales targets of U.S. companies.
The GB200 AI rack includes 36 "grace" central processing units and 72 Blackwell graphics processing units connected through NVIDIA's NVLINK communication system.
Nvidia's Taiwanese partners Foxconn and Wiston spoke at a Computex conference in Taipei last week, saying the shipment of GB200 shelves began at the end of the first quarter. They added that production capacity is now rapidly expanding.
"Our internal testing showed connectivity issues ... collaborating with NVIDIA, which happened two to three months ago," said one of NVIDIA's partner manufacturers.
The development is ahead of NVIDIA's quarterly earnings on Wednesday, where investors will watch signs of Blackwell shipments being carried out at speed after initial technical issues.
Saudi Arabia and the United Arab Emirates recently announced plans to acquire thousands of chips during Donald Trump's trip to the Gulf as NVIDIA looks beyond the large-scale tech "super standard" companies to diversify its states into its customer base.
NVIDIA's supply chain partners have spent months addressing several challenges on the GB200 rack, including overheating caused by its 72 high-performance GPUs, and leaks in the liquid cooling system.
Engineers also cite software errors and inter-chip connectivity issues, which are caused by the complexity of syncing such a large number of processors.
"This technology is really complicated. No company is trying to make these numerous AI processors work simultaneously in servers before and work in such a short time," said Chu Wei-Chia, a Taipei analyst at the consulting firm.
"NVIDIA does not allow supply chains to be fully prepared, so delays. As manufacturers increase rack production in the second half of the year, inventory risks will be mitigated."
To ensure smoother deployments for key customers such as Microsoft and Meta, suppliers have strengthened their testing protocols before shipping, running more checks to ensure rack-functional AI workloads.
NVIDIA is also preparing for the launch of its next-generation GB300 AI rack with enhanced memory capabilities designed to handle more complex inference models such as OpenAI's 01 and DeepSeek R1. Huang said last week that the GB300 will be launched in the third quarter.
To accelerate deployment, NVIDIA has been compromised in the design of GB300. It initially planned to introduce a new chipboard layout called "Cordelia" that allows for the replacement of a single GPU.
But in April, the company told partners that due to installation issues, the company would revert to the "Bianca" design used in the current GB200 rack.
This decision could help NVIDIA achieve its sales targets. The company said in February that the company targeted sales of about $43 billion in the quarter, with the record rising by about 65% in the digital year as of the end of April.
Analysts say Cordelia's board could have provided better profit margins and made it easier for customers to maintain.
According to three people familiar with the matter, NVIDIA has not abandoned Cordelia and has informed its suppliers that intend to redesign in its next-generation AI chips.
Additionally, NVIDIA is working to offset China's revenue losses after the U.S. government banned its H20 chip exports (a watered version of its AI processor). The company said it expects to incur $5.5 billion in fees on the ban due to inventory sales and purchase commitments.
Last week, Bank of America analyst Vivek Arya wrote that China's sales hit rate would lower Nvidia's gross margin in the quarter, from 71% to about 58% the company previously indicated.
But Blackwell's expectations were faster than expected as the company resumed Bianca's board of directors and could help offset Chinese revenue in the second half of the year, he wrote.