Nvidia’s suppliers resolve AI ‘rack’ issues in boost to sales

Unlock the Editor’s Digest for free

Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.

Nvidia’s suppliers are accelerating production of its flagship AI data centre “racks” following a resolution of technical issues that had delayed shipments, as the US chipmaker intensifies its global sales push.

The semiconductor giant’s partners — including Foxconn, Inventec, Dell and Wistron — have made a series of breakthroughs that have allowed them to start shipments of Nvidia’s highly anticipated “Blackwell” AI servers, according to several people familiar with developments at the groups.

The recent fixes are a boost to chief executive Jensen Huang, who unveiled Blackwell last year promising it would massively increase the computing power needed to train and use large language models.

Technical problems that emerged at the end of last year had disrupted their production, threatening the US company’s ambitious annual sales targets.

The GB200 AI rack includes 36 “Grace” central processing units and 72 Blackwell graphics processing units, connected through Nvidia’s NVLink communication system.

Speaking at the Computex conference in Taipei last week, Nvidia’s Taiwanese partners Foxconn, Inventec and Wistron said shipments of the GB200 racks began at the end of the first quarter. Production capacity is now being rapidly scaled up, they added.

“Our internal tests showed connectivity problems . . . the supply chain collaborated with Nvidia to solve the issues, which happened two to three months ago,” said an engineer at one of Nvidia’s partner manufacturers.

The development comes ahead of Nvidia’s quarterly earnings on Wednesday, where investors will be watching for signs that Blackwell shipments are proceeding at pace following the initial technical problems.

Saudi Arabia and the United Arab Emirates recently announced plans to acquire thousands of Blackwell chips during President Donald Trump’s tour of the Gulf, as Nvidia looks beyond the Big Tech “hyperscaler” companies to nation states to diversify its customer base.

Nvidia’s supply chain partners have spent months tackling several challenges with the GB200 racks, including overheating caused by its 72 high-performance GPUs, and leaks in the liquid cooling systems.

Engineers also cited software bugs and inter-chip connectivity problems stemming from the complexity of synchronising such a large number of processors.

“This technology is really complicated. No company has tried to make this many AI processors work simultaneously in a server before, and in such a short timeframe,” said Chu Wei-Chia, a Taipei-based analyst at consultancy SemiAnalysis.

“Nvidia had not allowed the supply chain sufficient time to be fully ready, hence the delays. The inventory risk around GB200 will ease off as manufacturers increase rack output in the second half of the year,” Chu added.

To ensure a smoother deployment for major customers such as Microsoft and Meta, suppliers have beefed up testing protocols before shipping, running more checks to ensure the racks function for AI workloads.

Nvidia is also preparing for the rollout of its next-generation GB300 AI rack, which features enhanced memory capabilities and is designed to handle more complex reasoning models such as OpenAI’s 01 and DeepSeek R1. Huang said last week that GB300 will launch in the third quarter.

In a bid to accelerate deployment, Nvidia has compromised on aspects of the GB300’s design. It had initially planned to introduce a new chip board layout, known as “Cordelia,” allowing for the replacement of individual GPUs.

But in April, the company told partners it would revert to the earlier “Bianca” design — used in the the current GB200 rack — due to installation issues, according to two suppliers.

The decision could help Nvidia to achieve its sales targets. In February the company said it was aiming for around $43bn in sales for the quarter to the end of April, a record figure which would be up around 65 per cent year on year.

Analysts have said the Cordelia board would have offered the potential for better margins and made it easier for customers to do maintenance.

Nvidia has not abandoned Cordelia and has informed suppliers it intends to implement the redesign within its next-generation AI chips, according to three people familiar with the matter.

Separately, Nvidia is working to offset revenue losses in China, following a US government ban on exports of its H20 chip — a watered-down version of its AI processors. The company said it expects to incur $5.5bn in charges related to the ban, due to inventory write-offs and purchase commitments.

Last week Bank of America analyst Vivek Arya wrote that the China sales hit would drag down Nvidia’s gross margins for the quarter from the 71 per cent previously indicated by the company to around 58 per cent.

But he wrote that a faster than expected rollout of Blackwell due to the company reverting back to Bianca boards could help offset the China revenue hit in the second half of the year.